Voices in AI – Episode 63: A Conversation with Hillery Hunter

[voices_in_ai_byline]

About this Episode

Episode 63 of Voices in AI features host Byron Reese and Hillery Hunter discuss AI, deep learning, power efficiency, and understanding the complexity of what AI does with the data it is fed. Hillery Hunter is an IBM Fellow and holds an MS and a PhD in electrical engineering from the University of Illinois Urbana-Champaign.
Visit www.VoicesinAI.com to listen to this one-hour podcast or read the full transcript.

Transcript Excerpt

Byron Reese: This is Voices in AI brought to you by GigaOm, I’m Byron Reese. Today, our guest is Hillery Hunter. She is an IBM Fellow, and she holds an MS and a PhD in electrical engineering from the University of Illinois Urbana-Champaign. Welcome to the show, Hillery.
Thank you it’s such a pleasure to be here today, looking forward to this discussion, Byron.
So, I always like to start off with my Rorschach test question, which is: what is artificial intelligence, and why is it artificial?
You know that’s a great question. My background is in hardware and in systems and in the actual compute substrate for AI. So one of the things I like to do is sort of demystify what AI is. There are certainly a lot of definitions out there, but I like to take people to the math that’s actually happening in the background. So when we talk about AI today, especially in the popular press and such and people talk about the things that AI is doing, be it understanding medical stands or labelling people’s pictures on a social media platform, or understanding speech or translating language, all those things that are considered core functions of AI today are actually deep learning, which means using many layered neural networks to solve a problem.
There’s also other parts of AI though, that are much less discussed in popular press, which include knowledge and reasoning and creativity and all these other aspects. And you know the reality is where we are today with AI, is we’re seeing a lot of productivity from the deep learning space and ultimately those are big math equations that are solved with lots of matrix math, and we’re basically creating a big equation that matches in its parameters to a set of data that it was fed.
So, would you say though that that it is actually intelligent, or that it is emulating intelligence, or would you say there’s no difference between those two things?
Yeah, so I’m really quite pragmatic as you just heard from me saying, “Okay, let’s go talk about what the math is that’s happening,” and right now where we’re at with AI is relatively narrow capabilities. AI is good at doing things like classification or answering yes and no kind of questions on data that it was fed and so in some sense, it’s mimicking intelligence in that it is taking in sort of human sensory data a computer can take in. What I mean by that is it can take in visual data or auditory data, people are even working on sensory data and things like that. But basically a computer can now take in things that we would consider sort of human process data, so visual things and auditory things, and make determinations as to what it thinks it is, but certainly far from something that’s actually thinking and reasoning and showing intelligence.
Well, staying squarely in the practical realm, that approach, which is basically, let’s look at the past and make guesses about the future, what is the limit of what that can do? I mean, for instance, is that approach going to master natural language for instance? Can you just feed a machine enough printed material and have it be able to converse? Like what are some things that model may not actually be able to do?
Yeah, you know it’s interesting because there’s a lot of debate. What are we doing today that’s different from analytics? We had the big data era, and we talked about doing analytics on the data. What’s new and what’s different and why are we calling it AI now? To refer to your question from that direction, one of the things that AI models do, be it anything from a deep learning model to something that’s more in the knowledge reasoning area, is that they’re much better interpolators, they’re much better able to predict on things that they’ve never seen before.
Classical rigid models that people programmed in computers, could answer “Oh, I’ve seen that thing before.” With deep learning and with more modern AI techniques, we are pushing forward into computers and models being able to guess on things that they haven’t exactly seen before. And so in that sense there’s a good amount of interpolation influx, whether or not and how AI pushes into forecasting on things well outside the bounds of what it’s never seen before and moving AI models to be effective at types of data that are very different from what they’ve seen before, is the type of advancement that people are really pushing for at this point.
Listen to this one-hour episode or read the full transcript at www.VoicesinAI.com
[voices_in_ai_link_back]
 
Byron explores issues around artificial intelligence and conscious computers in his new book The Fourth Age: Smart Robots, Conscious Computers, and the Future of Humanity.

Voices in AI – Episode 54: A Conversation with Ahmad Abdulkader

[voices_in_ai_byline]

About this Episode

Episode 54 of Voices in AI features host Byron Reese and Ahmad Abdulkader talking about the brain, learning, and education as well as privacy and AI policy. Ahmad Abdulkader is the CTO of Voicera. Before that he was the technical lead of Facebook’s DeepText, an AI text understanding engine. Prior to that he developed OCR engines, machine learning systems, and computer vision systems at Google.
Visit www.VoicesinAI.com to listen to this one-hour podcast or read the full transcript.

Transcript Excerpt

Byron Reese: This is Voices in AI brought to you by GigaOm. I am Byron Reese. Today our guest is Ahmad Abdulkader. He is the CTO of Voicera. Before that he was the lead architect for Facebook supplied AI efforts producing Deep Texts, which is a text understanding engine. Prior to that he worked at Google building OCR engines, machine learning systems, and computer vision systems. He holds a Bachelor of Science and Electrical Engineering degree from Cairo University and a Masters in Computer Science from the University of Washington. Welcome to the show.
Ahmad Abdulkader: Thank you, thanks Byron, thanks for having me.
I always like to start out by just asking people to define artificial intelligence because I have never had two people define it the same way before.
Yeah, I can imagine. I am not aware of a formal definition. So, to me AI is the ability of machines to do or perform cognitive tasks that humans can do or learn to do rather. And eventually learn to do it in a seamless way.
Is the calculator therefore artificial intelligence?
No, the calculator is not performing a cognitive task. A cognitive task I mean vision, speech understanding, understanding text, and such. Actually, in fact the brain is actually lousy at multiplying two six-digit numbers, which is what the calculator is good at. But the calculator is really bad at doing a cognitive test.
I see, well actually, that is a really interesting definition because you’re defining it not by some kind of an abstract notion of what it means to be intelligent, but you’ve got a really kind of narrow set of skills that once something can do those, it’s an AI. Do I understand you correctly?
Right, right, I have a sort of a yard stick, or I have a sort of a set of tasks a human can do in a seamless easy way without even knowing how to do it, and we want to actually have machines mimic that to some degree. And there will be some very specific set of tasks, some of them are more important than others and so far, we haven’t been able to build machines that actually get even close to the human beings around these tasks.
Help me understand how you are seeing the world that way, and I don’t want to get caught up on definitions, but this is really interesting.
Right.
So, if a computer couldn’t read, couldn’t recognize objects, and couldn’t do all those things you just said, but let’s say it was creative and it could write novels. Is that an AI?
First of all, this is hypothetical. I wouldn’t know, I wouldn’t call it AI, so it goes back to the definition of intelligence, and then there’s a natural intelligence that humans exhibit, and then there is artificial intelligence that machines will attempt to make and exhibit. So, the most important of these that we actually use sort of almost every second of the day are vision, speech understanding, or language understanding, and creativity is one of them. So if you were to do that I would say this machine performed a subset of AI, but haven’t exhibited the behavior to show that’s it good at the most important ones, being vision, speech and such.
When you say vision and speech are the most important ones, nobody’s ever really looked at the problem this way, so I really want to understand how you’re saying that, because it would seem to me those aren’t really the most important by a long shot. I mean, if I had an AI that could diagnose any disease, tell us how to generate unlimited energy, fix all the environmental woes, tell us how to do faster than light travel, all of those things, like, feed the hungry, and alleviate poverty and all of those things, but they couldn’t tell a tuna fish from a Land Rover. I would say that’s pretty important, I would take that hands down over what you’re calling to be more important stuff.
I think really important is an overloaded word. I think you’re talking about utility, right? So, you’re imagining a hypothetical situation where we’re able to build computers that will do the diagnosis or poverty and stuff like that. These would be way more useful for us, or that’s what we think, or that’s the hypothesis. But actually to do these tasks that you’re talking about, it probably implies, most probably that you have done or solved, to a great degree, solved vision. It’s hard to imagine that you would be doing diagnosis without actually solving vision. So, these are sort of the basic tasks that actually humans can do, and babies learn, and we see babies or children learn this as they grow up. So, perhaps the utility of what you talked about would be much more useful for us, but if you were to define importance as sort of the basic skills that you could build upon, I would say vision would be the most important one. Language understanding perhaps would be the second most important one. And I think doing well in these basic cognitive skills would enable us to solve the problems that you’re talking about.
Listen to this one-hour episode or read the full transcript at www.VoicesinAI.com 
[voices_in_ai_link_back]
 
Byron explores issues around artificial intelligence and conscious computers in his new book The Fourth Age: Smart Robots, Conscious Computers, and the Future of Humanity.

Voices in AI – Episode 47: A Conversation with Ira Cohen

[voices_in_ai_byline]
In this episode, Byron and Ira discuss transfer learning and AI ethics.
[podcast_player name=”Episode 47: A Conversation with Ira Cohen” artist=”Byron Reese” album=”Voices in AI” url=”https://voicesinai.s3.amazonaws.com/2018-06-05-(01-02-19)-ira-cohen.mp3″ cover_art_url=”https://voicesinai.com/wp-content/uploads/2018/06/voices-headshot-card.jpg”]
[voices_in_ai_byline]
Byron Reese: This is Voices in AI, brought to you by GigaOm, and I’m Byron Reese. Today our guest is Ira Cohen, he is the cofounder and chief data scientist at Anodot, which has created an AI-based anomaly detection system. Before that he was chief data scientist over at HP. He has a BS in electrical engineering and computer engineering, as well as an MS and a PhD in the same disciplines from The University of Illinois. Welcome to the show, Ira.
Ira Cohen: Thank you very much for having me.
So I’d love to start with the simple question, what is artificial intelligence?
Well there is the definition of artificial intelligence of machines being able to perform cognitive tasks, that we as humans can do very easily. What I like to think about in artificial intelligence, is machines taking on tasks for us that do require intelligence, but leave us time to do more thinking and more imagination, in the real world. So autonomous cars, I would love to have one, that requires artificial intelligence, and I hate driving, I hate the fact that I have to drive for 30 minutes to an hour every day, and waste a lot of time, my cognitive time, thinking about the road. So when I think about AI, I think how it improves my life to give me more time to think about even higher level things.
Well, let me ask the question a different way, what is intelligence?
That’s a very philosophical question, yes, so it has a lot of layers in it. So, when I think about intelligence for humans, it’s the ability to imagine something new, so imagine, have a problem and imagine a solution and think about how it will look like without actually having to build it yet, and then going in and implementing it. That’s what I think about [as] intelligence..
But a computer can’t do that, right?
That’s right, so when I think about artificial intelligence, personally at least, I don’t think that, at least in our lifetime, computers will be able to solve those kind of problems, but, there is a lower level of intelligence of understanding the context of where you are, and being able to take actions on it, and that’s where I think that machines can do a good task. So understanding a context of the environment and taking immediate actions based on that, that are not new, but are already… people know how to do them, and therefore we can code them into machines to do them.
I’m only going to ask you one more question along these lines and then we’ll move on, but you keep using the word “understand.” Can a computer understand anything?
So, yeah, the word understanding is another hard word to say. I think it can understand, well, at least it can recognize concepts. Understanding maybe requires a higher level of thinking, but understanding context and being able to take an action on it, is what I think understanding is. So if I see a kid going into the road while I’m driving, I understand that this is a kid, I understand that I need to hit the brake, and I think machines can do these types of understanding tasks.
Fair enough, so, if someone said what is the state of the art like, they said, where are we at with this, because it’s in the news all the time and people read about it all the time, so where are we at?
So, I think we’re at the point where machines can now recognize a lot of images and audio or various types of data, recognize with sensors, recognize that there are objects, recognize that there are words being spoken, and identify them. That’s really where we’re at today, we’re not… we’re getting to the point where they’re starting to also act on these recognition tasks, but most of the research, most of what AI is today, is the recognition tasks. That’s the first step.
And so let’s just talk about one of those. Give me something, some kind of recognition that you’ve worked on and have deep knowledge of, teaching a computer how to do…
All right, so, when I did my PhD, I worked on affective computing, so, part of the PHD was to have machines recognize emotions from facial expressions. So, it’s not really recognizing emotion, it’s recognizing a facial expression and what it may express. So there are 6 universal facial expressions that we as humans exhibit, so, smiling is associated with happiness, there is surprise, anger, disgust, and those are actually universal. So, the task that I worked on was to build classifiers, that given an image or a sequence of a video of a person, a person’s face, would recognize whether they’re happy or sad or disgusted or surprised or afraid…
So how do you do that? Like do you start with biology and you say “well how do people do it?” Or do you start by saying “it doesn’t really matter how people are doing it, I’m just going to brute force, show enough labeled data, that it can figure it out, that it just learns without ever having a deep understanding of it?”
All right so this was in the early 2000s, and we didn’t have deep learning yet, so we had neural networks, but we weren’t able to train them with huge amounts of data. There wasn’t a huge amount of data, so the brute force approach was not the way to go. What I actually worked on is based on research by a psychologist, that actually mapped facial movements to known expressions, and therefore to known emotions. So it started out in the 70s, by people in the psychology field, [such as] Charles Akemann, in San Francisco, who mapped out actual… he created a map of facial movements into facial expressions, and so that was the basis of what are the type of features I need to extract from video and then feed that to a classifier, and then you go through the regular process of machine learning of collecting a lot of data, but the data is transformed, so these videos were transformed into known features of facial movements, and then, you can feed that into a classifier that learns in a supervised way. So I think a lot of the tasks around intelligence are that way. It’s being changed a little bit by deep learning, which supposedly takes away the need to know the features are a priori, and do the feature engineering for the machinery task…
Why do you say “supposedly”?
Because it’s not completely true. You still have to do, even in speech, even in images, you still have to do some transformations of the raw data, it’s not just take it as is, and it will work magically and do everything for you. There is some… you do have to, for example in speech, you do have to do various transformations of the speech into all sorts of short term Fourier transform or other types of transformations, without which, the methods afterwards will not produce results.
So, if I look at a photo of a cat, that somebody’s posted online or a dog, that’s in surprise, you know, it’s kind of comical, the look of surprise, say, but a human can recognize that in something as simple as a stick figure… What are we doing there do you think? Is that a kind of transferred learning, or how is it that you can show me an alien and I would say, “Ah, he’s happy…”What do you think we’re doing there…?
Yeah, we’re doing transferred learning. Those are really examples of us taking one concept that we were trained on from the day we were born, with our visual cortex and also then in the brain, because our brain is designed to identify emotions, just out of the need to survive, and then when we see something else, we try to map it onto a concept that we already know, and then if something happens that is different from what we expected, then we start training to that new concept. So if we see an alien smiling, and all of a sudden when he smiles, he shoots at you, you would quickly understand that smiling for an alien, is not associated with happiness, but you will start offby thinking, “this could be happy”.
Yeah, I think that I remember reading that, hours after birth, children who haven’t even been trained on it, can recognize the difference between a happy and sad face. I think they got sticks and put drawings on them and try to see the baby’s reactions. It may even be even something deeper than something we learn, something that’s encoded in our DNA.
Yeah, and that may be true because we need to survive.
So why do you think we’re so good at it and machines aren’t, right, like, machines are terrible right now at transfer learning. We don’t really know how it works do we, because we can’t really code that abstraction that a human gets, so..
I think that from what I see first, it’s being changed. I see work coming out of Google AI labs that is starting to show how they are able to train single models, very large models, that are able to do some transfer learning on some tasks, and, so it is starting to change. So machines have a very different… they don’t have to survive –  they don’t have this notion of danger, and surviving, and I think until we are able to somehow encode that in them, we would always have to, ourselves, code the new concepts or understand how to code for them, how to learn new concepts using transfer learning…
You know the roboticist Rodney Brooks, talks about “the juice”, he talks about how, if you put an animal in a box, it feels trapped, it just tries and tries to get out and it clearly has a deep desire to get out, but you but in a robot to do it, the robot doesn’t have what he calls “the juice,” and he of course doesn’t think it’s anything spiritual or metaphysical or anything like that. But what do you think that is? What do you think is the juice? Because that’s what you just alluded to, machines don’t have to survive, so what do you think that is?
So I think he’s right, they don’t have the juice. Actually in my lab, during my PhD, we had some students working on teaching robots to move around, and actually, the way they did it was rewards and punishments. So they would get… they actually coded—just like you have in reinforcement learning—if you hit a wall, you get a negative reward. If the robot moved and did something he wasn’t supposed to, the PhD student would yell at them, and that would be encoded into a negative reward, and if he did something right, they had actions that gave them positive rewards. Now it was all kind of fun and games, but potentially if you do this for long enough, with enough feedback, the robot would learn what to do and what not to do, the main thing that’s different is that it still lives in the small world of where they were, in the lab or in the hallways of our labs. It didn’t have the intelligence to then take it and transfer it to somewhere else…
But the computer can never… I mean the inherent limit in that is that the computer can never be afraid, be ashamed, be motivated, be happy…
Yes. It doesn’t have the long term reward or the urge to survive, I guess.
You may be familiar with this, but I’d like to set it up anyway. There was a robot in Japan, it was released in a mall, and it was basically being taught how to get around and if it ran into a person, if it came up to a person, it would politely ask the person to move, and if the person didn’t, it would just zoom around them. And what happened was children would just kind of mess with it, maybe jump in front of it when it tried to go around them again and again and again, but the more kids there were, the more likely they were to get brutal. They would hit it with things, they would yell at it and all of that, and the programmers ended up having to program it, that if it had a bunch of short people around it, like children, it needed to find a tall person, an adult, and zip towards it, but the distressing thing about it is when they later asked those children who had done that, they said, “Did you cause the robot distress?” 75% of them said yes, and then they asked if it behaved human-like or machine-like, and only 15% said machine-like, and so they thought that they were actually causing distress and it was behavinglike a humanoid.What do you think that says? Does that concern you in any way?
Personally, it doesn’t, because I know that, as long as machines don’t have real affect in them, then, we might be transferring what we think stress is onto a machine that doesn’t really feel that stress… it’s really about codes…
I guess the concern is that if you get in the habit of treating something that you regard as being in distress, if you get into the habit of treating it callously, this is what Weizenbaum said, he thought that it would have a dampening effect on human empathy, which would not be good… Let me ask you this, what do you think about embodying artificial intelligence? Because you think about the different devices: Amazon has theirs, it’s right next to me, so I can’t say its name, but it’s a person’s name… Apple has Siri, Microsoft has Cortana… But Google just has the google system, it doesn’t have a name. Do you think there’s anything about that… why do you think it is? Why would we want to name it or not name it, why would we decide not to name it? Do you think we’re going to want to interact with these devices as if they’re other people? Or are we always going to want them to be obviously mechanistic?
My personal feeling is that we want them to be mechanistic, they’re there not to exist on their own accord, and reproduce and create a new world. They’re there to help us, that’s the way I think AI should be, to help us in our tasks. Therefore when you start humanizing it, then you’re going to either have the danger of mistreating it, treating it like basically slaves, or you’re going to give it other attributes that are not what they are, thinking that they are human, and then going the other route, and they’re there to help us, just like robots, or just like the industrial revolution brought machines that help humans manufacture things better… So they’re there to help us, I mean we’re creating them, not as beings, but rather as machines that help us improve humanity, and if we start humanizing them and then, either mistreating them, like you mentioned with the Japanese example, then it’s going to get muddled and strange things can happen…
But isn’t that really what is going to happen? Your PhD alone, which is how do you spot emotions? Presumably would be used in a robot, so it could spot your emotions, and then presumably it would be programmed to empathize with you, like “don’t be worried, it’s okay, don’t be worried,” and then to the degree it has empathy with you, you have emotional attachment to it, don’t you go down that path?
It might, but I think we can stop it. So the reason to identify the emotion is because it’s going to help me do something, so, for example, our research project was around creating assistance for kids to learn, so in order to help the kid learn better, we need to empathize with the state of mind of the child, so it can help them learn better. So that was the goal of the task, and I think as long as we encapsulate it in well-defined goals that help humans, then, we won’t have the danger of creating… the other way around.  Now, of course maybe in 20 years, what I’m saying now will be completely wrong and we will have a new world where we do have a world of robots that we have to think about how do we protect them from us. But I think we’re not there yet, I think it’s a bit science fiction, this one.
So I’m still referring back to your earlier “supposedly” comment about neural nets, what do you think are other misconceptions that you run across about artificial intelligence? What do you think are, like your own pet peeves, like “that’s not true, or that’s not how it works?” Does anything come to mind?
People think, because of the hype, that it does a lot more than it really does. We know that it’s really good at classification tasks, it’s not yet very good at anything that’s not classification, unsupervised tasks, it’s not being able to learn new concepts all by itself, you really have to code it, and it’s really hard. You need a lot of good people that know the art of applying neural nets to different problems. It doesn’t happen just magically, the way people think.
I mean you’re of course aware of high profile people: Elon Musk, Stephen Hawking, Bill Gates, and so forth who [have been] worried about what a general intelligence would do, they use terms like “existential threat” and all that, and they also, not to put words in their mouth, believe that it will happen sooner rather than later… Because you get Andrew Ng, who says, “worry about overpopulation of Mars,” maybe in a couple hundred years you have to give it some thought, but you don’t really right now…So where do you think their concern comes from?
So, I’m not really sure and I don’t want to put any words in their mouth either, but, I mean the way I see it, we’re still far off from it being an existential threat. The main concern is you might have people who will try to abuse AI, to actually fool other people, that I think is the biggest danger, I mean, I don’t know if you saw the South Park episode last week, they had their first episode where Cartman actually bought an Alexa and started talking to his Alexa, and I hope your Alexa doesn’t start working now…. So it basically activated a lot of Alexas around the country, so he was adding stuff to the shopping cart, really disgusting stuff, he was setting alarm clocks, he was doing all sorts of things, and I think the danger of the AI today is really getting abused by other people, for bad purposes, in this case it was just funny… But you can have cases where people will control autonomous cars, other people’s autonomous cars by putting pictures by the side of the road and causing them to swerve or stop, or do things they’re not supposed to, or building AI that will attack other types of AI machines. So I think the danger comes from the misuse of the technology, just like any other technology that came out into the world… And we have to… I think that’s where the worry comes from, and making sure that we put some sort of ethical code of how to do that…
What would that look like? I mean that’s a vexing problem…
Yes, I don’t know, I don’t have the answer to that…
So there are a number of countries, maybe as many as twenty, that are working on weaponizing, building AI-based weapons systems, that can make autonomous kill decisions. Does that worry you? Because that sounds like where you’re going with this… if they put a plastic deer on the side of the road and make the car swerve, that’s one thing, but if you literally make a killer robot that goes around killing people, that’s a whole different thing. Does that concern you, or would you call that a legitimate use of the technology…?
I mean this kind of use will happen, I think it will happen no matter what, it’s already happening with drones that are not completely autonomous, but they will be autonomous probably in the future. I think that I don’t know how it can be… this kind of progress can be stopped, the question is, I mean, the danger I think is, do these robots start having their own decision-making and intelligence that decides, just like in the movies, to attack all humankind, and not just the side they’re fighting on… Because technology in [the] military is something that… I don’t know how it can be stopped, because it’s driven by humans… Our need to wage war against each other… The real danger is, do they turn on us? And if there is real intelligence in the artificial intelligence, and real understanding and need to survive as a being, that’s where it becomes really scary…
So it sounds like you don’t necessarily think we’re anywhere near close to an AGI, and I’m going to ask you how far away you think we are… I want to set the question up as saying that, there are people who think we’re 5-10 years away from a general intelligence and then there are people who think we’re 500 years [away].Oren Etzioni was on the show, and he said he would give anyone 1000:1 odds that we wouldn’t have it in 5 years, so if you want to send him $10 he’ll put $10,000 against that. So why do you think there’s such a gap, and where are you in that continuum?
Well, because the methods we’re using are still so… as smart as they got, they’re still doing rudimentary tasks. They’re still recognizing images—the agents that are doing automated things for us, they’re still doing very rudimentary tasks. General intelligence requires a lot more than that, that requires a lot more understanding of context. I mean the example of Alexa last week, that’s a perfect example of not understanding context, for us as humans, we would never react to something on TV like that and add something to our shopping cart, just because Cartman said it, where even the very, very smart Alexa with amazing speech understanding, and taking actions based on that, it still doesn’t understand the context of the world, so I think prophecy is for fools, but I think it’s at least 20 years out…
You know, we often look at artificial intelligence and its progress based on games where it beats the best player, that goes back to [Garry] Kasparov in 97, you have of course Jeopardy, you have Alpha Go, you had… an AI beat some world rated poker players, what do you think…And those are all kind of… they create a stir, you want to reflect on it, what do you think is the next thing like that, that one day, snap your fingers and all of a sudden an AI just did… what?
Okay, I haven’t thought about that… All these games, what makes them unique is that they are a very closed world; the world of the game, is finite and the rules are very clear, even if there’s a lot of probability going on, the rules are very clear, and if you think in the real world—and this may be going back to the questions why it will take time—for artificial intelligence to really be general intelligence, the real world is almost infinite in possibilities and the way things can go, and even for us, it’s really hard.
Now trying to think of a game that machines would beat us next in. I wonder if we were able to build robots that can do lots of sports, I think they could beat us easily in a lot of games, because if you take any sports game like football or basketball, they require intelligence, they require a lot of thinking, very fast thinking and path finding by the players, and if we were able to build the body of the robot that can do the motions just like humans, I think they can easily beat us at all these games.
Do you, as a practitioner… I’m intrigued by it, on the topic of general intelligence, intrigued by the idea that, human DNA isn’t really that much code, and if you look at how much code that we are different than say a chimp, it’s very small, I mean it’s a few megabytes. That would be, how we are programmatically different, and yet, that little bit of code, makes us have a general intelligence and a chimp not. Does that persuade you or suggest to you that general intelligence is a simple thing, that we just haven’t discovered, or do you think that general intelligence is a hack of a hundred thousand different… like it’s going to be a long slog and then we finally get it together…?
So, I think [it’s] the latter, just because the way you see human progress, and it’s not just about one person’s intelligence. I think what makes us unique is the ability to combine intelligence of a lot of different people to solve tasks, and that’s another thing that makes us very different. So you do have some people that are geniuses that can solve really really hard tasks by themselves, but if you look at human progress, it’s always been around combined intelligence of getting one person’s contribution, then another person’s contribution, and thinking about how it comes together to solve that, and sometimes you have breakthroughs that come from an individual, but more often than not, it’s the combined intelligence that creates the drive forward, and that’s the part that I think is hard to put into a computer…
You know there are people that have, amazing savant-like abilities. I remember reading about a man named [George] Dantzig, and he was a graduate student in statistics, and his professor put two famous unsolvable/unsolved problems on the blackboard, and Dantzig arrived late that day. He saw them and just assumed that they were the homework, so he copied them down and went home, and later he said he thought they were a little harder than normal, but he solved them both and turned them in… and that like really happened. It’s not one of those urban legend kind of things, you have people who can read the left and right page of a book at the same exact time, you have… you just have people that are these extraordinarily edge cases of human ability,does that suggest that our intellects are actually far more robust than they are? Does that suggest anything to you as an artificial intelligence guy?
Right, so coming from the probability space, it just means that our intelligence has wide distribution, and there are always exceptions in the tails, right? And these kind of people are in the tails, and often when they are discovered, they can create monumental breakthroughs in our understanding of the world, and that’s what makes us so unique. You have a lot of people in the center of the distribution, that are still contributing a lot, and making advances to the world and to our understanding of it, and not just understanding, but actually creating new things. So I’m not a genius, most people are not geniuses, but we still create new things, and are able to advance things, and then, every once in a while you get these tails of a distribution intelligence, that could solve the really hard problems that nobody else can solve, and that’s a… so the combination of all that actually makes us push things forward in the world, and I think that kind of combined intelligence, I think that artificial intelligence is way, way off. It’s not anywhere near, because we don’t understand how it works, I think it would be hard for us to even code that into machines. That’s one of the reasons I think AI, the way people are afraid of it, it’s still way off…
But by that analysis, that sounds like, to circle that back, there will be somebody that comes along that has some big breakthrough in a general intelligence, and ta-da, it turns out all along it was, you know, bubble sort or….
I don’t think it’s that simple, that’s the thing, and solving a statistical problem that’s really, really tough, it’s not like… I don’t think it’s a well-defined enough problem, that some will take a genius just to understand.. “Oh, it’s that neuron going right to left,” and that’s it… so I don’t think it’s that simple… there might be breakthroughs in mathematics, that help you understand the computation better, maybe quantum computers that will help you do faster computation, so you can train much, much faster than machines so they can do the task much better, but, it’s not about understanding the concept of what makes a genius. I think that’s more complicated, but maybe it’s my limited way of thinking, maybe I’m not intelligent enough with it…
So to stay on that point for a minute… it’s interesting and I think perhaps, telling, that we don’t really understand how human intelligence works, like if you knew that.. like we don’t know how a thought is encoded in the brain… like if I said…Ira, what color was your first bicycle, can you answer that question?
I don’t remember… probably blue…
Let’s assume for a minute that you did remember. It makes my example bad, but there’s no bicycle location in your brain that stored the first “bicycle”… like an icon, or database lookup…like nobody knows how that happens… not only how it’s encoded, but how it’s retrieved… And then, you were talking earlier about synthesis and how we use it all together, we don’t know any of that… Does that suggest to you that, on the other end, maybe we can’t make a general intelligence… or at the very least, we cannot make a general intelligence until we understand how it is that people are intelligent…?
That may be, but yeah. First of all even if we made it, if we don’t understand it, then how would we know that we made it? Circling back to that… I think the way we… it’s just like the kids, they were thinking that they were causing stress to the robot, because they were giving it… they thought they understood stress and the affect of it, and they were transferring it onto the robot. So maybe when we create something very intelligent that looks to be like us, we would think we created intelligence, but we wouldn’t know that for sure until we know what is… general intelligence really is…
So do you believe that general intelligence is an evolutionary invention that will come along if, in 20 years, 50 years, 1,000 years… whatever it is, that it is something that will come along out of the techniques we use today from the early AI, like, are we building really, really, really primitive general intelligences, or do you have a feeling that a real AGI is going to be a whole different kind of approach in technology?
I think it’s going to be a whole different approach. I think what we’re building today are just machines that do tasks that we humans do, in a much, much better way, and just like we built machines in the industrial revolution that did what people did with their hands, but did it in a much faster way, and better way… that’s the way I see what we’re doing today… And maybe I’m wrong, maybe I’m totally wrong, and we’re giving them a lot more general intelligence than we’re thinking, but the way I see it, it’s driven by economic powers, it’s driven by the need of companies to advance, and take away tasks that cost too much money to do by humans, or are too slow to do by humans… And, revolutionizing that way, and I’m not sure that we’re really giving them general intelligence yet, still we’re giving them ways to solve specific tasks that we want them to solve, and not something very very general that can just live by itself, and create new things by itself.
Let’s take up this thread, that you just touched on, about, we build them to do jobs we don’t want to do, and you analogize it to the Industrial Revolution… so as you know, just to set the problem up, there are 3 different narratives about the effect this technology, combined with robotics, or we’ll call it automation, in general, are going to have on jobs. And the three scenarios are: one is that, it’s going to destroy an enormous number of quote, low-skill jobs, and that, they will, by definition, be fewer low skilled jobs, and more and more people competing for them and you will have this permanent class of unemployable… it’s like the Great Depression in the US, just forever. And then you have people who say, no, it’s different than that, what it really is, is, they’re going to be able to do everything we can do, they’re going to have escape… Once a machine can learn a new task faster than a person, they’ll take every job, even the creative ones, they’ll take everything. And the third one says no, for 250 years we’ve had 5-10% of unemployment, its never really gotten out of that range other than the anomalous depression, and in that time we had electricity, we had mechanization, we had steam power, we had the assembly line… we had all these things come along that sure looked like job eaters, but what people did is they used the new technology to increase their own productivity and drive their own wages higher, and that’s the story of progress, that we have experienced…So which of those three theories, or maybe a fourth one, do you think is the correct narrative?
I think the third theory is probably the more correct narrative. It just gives us more time to use our imagination and be more productive at doing more things, improve things, so, all of a sudden we’ll have time to think about going and conquering the stars, and living in the stars, or improving our lives here in various ways… The only thing that scares me is the speed of it, if it happens too quickly, too fast.. So, we’re humans, it takes, as a human race, some time to adapt. If the change happens so fast and people lose their jobs too quickly, before they’re able to retrain for the new economy, the new way of [work], the fact that some positions will not be available anymore, that’s the real danger and I think if it happens too fast around the world, then, there could be a backlash.
I think what will happen is that the progress will stop because some backlash will happen in the form of wars, or all sorts of uprisings, because, at the end, people need to live, people need to eat, and if they don’t have that, they don’t have anything to live for, they’re going to rise up, they’re not just going to disappear and die by themselves. So, that’s the real danger, if the change happens too rapidly, you can have a depression that will actually cause the progress to slow down, and I hope we don’t reach that because I would not want us, as a world, to reach that stage where we have to slow down, with all the weapons we have today, this could actually be catastrophic too…
What do you mean by that last sentence?
So I mean we have nuclear weapons…
Oh, I see, I see, I see.
We have actual weapons that can, not just… could actually annihilate us completely…
You know, I hear you  Like…what would “too fast” be? First of all, we had that when the Industrial Revolution came along… you had the Luddite movement, when Ludd broke two spinning wheels you had the thresher riots [or Swing riots] in England in the 1820s, when the automated threat, you had the… the first day the London Times was printed using steam power instead of people. They were going to go find the guy who invented that, and string him up, you had a deep-rooted fear of labor-changing technology, that’s a whole current that constantly runs, but what would too fast look like? The electrification of industry just happened lightning fast, we went from generating 5% of our power from steam to 85% in just22 years…Give me a “too fast” scenario. Are you thinking about the truck drivers, or… tell me how it could “be too fast,” because you seem to be very cautious, like, “man, these technologies are hard and they take a long time and there’s a lot of work and a lot of slog,” and then, so what would too fast look like to you?
If it’s less than a generation, let’s say in 5 years, really, all taxi drivers and truck drivers lose their job because everything becomes automated, that seems to be too fast. If it happens in 20 years, that’s probably enough time to adjust, and I think… the transition is starting, it will start in the next 5 years, but it will still take some time for it to really take hold, because if people lose those jobs today, and you have thousands or hundreds of thousands, or even millions of people doing that, what are they going to do?
Well, presumably, I mean, classic economics says that, if that happened, the cost of taking a cab goes way down, right? And if that happens, that frees up money that I no longer have to spend on an expensive cab, and therefore I spend that money elsewhere,  which generates demand for more jobs, but, is the 5-year scenario… it may be a technical possibility, like we may “technically” do it, if we don’t have a legislative hurdle.
I read this article in India, which said they’re not going to allow self-driving cars in India because that would put people out of work, then you have the retrofit problem, then every city’s going to want to regulate it and say well, you can have a self-driving car, but it needs to have a person behind the wheel just in case. I mean like you would say, look, we’ve been able to fly airplanes without a pilot for decades, yet no airline in the world would touch that, in this plane, we have no pilot… even though that’s probably a better way to do it…So, do you really think we can have all the taxi drivers gone in 5 years?
No, and exactly for that reason, even if our technology really allows it. First of all, I don’t think it will totally allow it, because for it to really take hold you have to have a majority of cars on the road to be autonomous. Just yesterday I was in San Francisco, and I heard a guy say he was driving behind one of those self-driving cars in San Francisco, and he got stuck behind it, because it wouldn’t take a left turn when it was green, and it just forever wouldn’t take a left turn that humans would… The reason why it wouldn’t take a left turn was there were other cars that are human-driven on the road, and it was coded to be very, very careful about it, and he was 15 minutes late to our meeting just because of that self-driving car…
Now, so I think there will be a long transition partly because legislation will regulate it, and slow it down a bit, which is a good thing. You don’t want to change too fast, too quickly without making sure that it really works well in the world, and as long as there is a mixture of humans driving and machines driving, the machines will be a little bit “lame,” because they will be coded to be a lot more careful than us, and we’re impatient, so, that will slow things down which is a good thing, I think making a change too fast can lead to all sorts of economic problems as well…
You know in Europe they had… I could be wrong on this, I think it was first passed in France, but I think it was being considered by the entire EU, and it’s the right to know why the AI decided what it did. If an AI made the decision to deny you a loan, or what have you, you have the right to know why it did that… I had a simple question which was, is that possible? Could Google ever say, I’m number four for this search and my competitor’s number three, why am I number four and they’re number three? Is Google big and complicated enough, and you don’t have to talk specifically about Google, but, are systems big and complicated enough that we don’t know… there are so many thousands of factors that go into this thing, that many people never even look at, it’s just a whole lot of training…
Right, so in principle, the methods could tell you why they made that decision. I mean, even if there are thousands of factors, you can go through all of them and have not just the output of their recognition, but also highlight what were the attributes that caused it to decide it’s one thing or another. So from the technology point of view, it’s possible, from the practical point of view, I think for a lot of problems, you don’t, you won’t really care. I mean, if it recognized that there’s a cat in the image, and you know it’s right, you won’t care why it’s recognized that cat. I guess for some problems where the system made a decision that you don’t necessarily know why it made the decision, or you have to take action based on that recognition, you would want to know. So if I predicted for you that your revenue is going to increase by 20% in the next week, you would probably want that system to tell you, why do you think that’s happened, because there isn’t a clear reason for it that you would imagine yourself, but, if the system told you there is a face in this image, and you just look at the image, and you can see that there’s a face in that image, then you won’t have a problem with it, so I think it really depends on the problem that you’re trying to solve…
We talked about games earlier and you pointed out that they were closed environments and that’s really a place with explicit rules, a place that an AI can excel, and I’ll add to that, there’s a clear cut idea of what winning looks like, and what a point is. I think somebody on the show said, “Who’s winning this conversation right now?” There’s no way to do that, so my question to you is,if you walk around an enterprise and you say “where can I apply artificial intelligence to my business?” would you look for things that looked like games? Like, okay, HR you have all these successful employees that get high performance ratings, and then you have all these people you had to fire because they didn’t, and then you get all these resumes in. Which ones more look like the good people as opposed to the bad people? Are there lots of things like that in life that look like games… or is the whole game thing really a distraction from solving real world problems, nothing really is a game in the real world…
Yeah, I think it’d be wrong to look at it as a game, because the rules… first there is no real clear notion of winning. What you want is progress, you have goals that you want to progress towards, you want, for example, in business, you want your company to grow. That could be your goal, or you want the profits to grow, you want your revenue to grow, so you make these goals, because that’s how you want things to progress and then you can look at all the factors that help it grow. The world of how to “make it grow” is very large, there are so many factors, so if I look at my employees, there might be a low-performing employee in one aspect of my business, but maybe that employee brings to the team, you know, a lot of humor that causes them to be productive, and I can’t measure that. Those kind of things are really, really hard to measure and, so looking at it from a very analytic point of view of just a “game,” would probably miss a lot of important factors.
So tell me about the company you co-founded, Anodot, because you make an anomaly detection system using AIs. So first of all, explain what that is and what that looks like, but how did you approach that problem? If it’s not a game, instead of… you looked at it this way…
So, what are anomalies? Anomalies are anything that’s unexpected, so our approach was: you’re a business and you’re collecting lots and lots and lots of data related to your business. At the end, you want to know what’s going on with the business, that’s the reason you collect a lot of data. Now, when today, people have a lot of different tools that help them kind of slice and dice the data, ask questions about what’s happening there, so you can make informed decisions about the future or react to things that are happening right now, that could affect your business.
The problem with that, is that basically… why isn’t it AI? It’s not AI because you’re basically asking a question and letting the computers compute something for you and giving you and answer; whereas anomalies, by nature, are things that happen that are unexpected, so you don’t necessarily know to ask the question in advance, and unexpected things could happen.  In businesses for example, you see a certain revenue for a product you’re selling going down in a certain city, why’s that happening? If you don’t look at it, and if you don’t ask the question in advance, you’re not even aware that that is happening… so, the great thing about AI, and machine learning algorithms, is they can process a lot of data, and if you can encode into a machine, an algorithm that identifies what are anomalies, you can find them in very, very large scale, and that helps the companies actually detect that things are going wrong, or detect the opportunities that they have, that they might miss otherwise. Where the endgame is very simple, to help you improve your business constantly and maintain it and avoid the risks of doing business, so, it’s not a “game,” it’s actually bringing immediate value to a company, highlighting, putting light on the data that they really need to look at with respect to their business, and the great thing about machine-learning algorithms, [is] they can process all of this data much better than we could, because what do humans do? We graph them, we visualize the data in various ways, you know, we create queries from database about questions that we think might be relevant, but we can’t really process all the data, all the time in an economical way. You would have to hire armies of people to do that, and machines are very good at that, so, that’s why we built Anodot…
Give me an example, like tell me a use case or a real world example of something that Anodot, well that you were able to spot that a person might not have been able to…?
So, we have various customers that are in the e-commerce business, and if you’re in e-commerce and you’re selling a lot of different products, various things could go wrong or opportunities might be missed. For example, if I’m selling coats, and I’m selling a thousand other products, I’m selling coats, and now in a certain area of the country, there is an anomalous weather condition that became cold, all of a sudden I’ll see, I won’t be able to see it because it’s hiding in my data, but people will start buying… in that state will start buying more coats. Now it’s not like if… if somebody actually looked at it, they would probably be able to spot it, but because there is so much data, so many things, so many moving parts, nobody actually notices it. Now our AI system finds… “Oh, there is an anomalous weather condition and there is an uptick in selling that coat, you better do something to seize that opportunity to sell more coats,” so either you have to send more inventory to that region to make sure that if somebody really wants a coat, you’re not out of stock. If you’re out of stock, you’re losing revenue, potential revenue, or you can even offer discounts for that region because you want to bring more people to your e-commerce site, rather than the competition, so, that’s one example…
And I assume it’s also used in security or fraud and what not, or are you really focused on an e-commerce-use case?
So we built a fairly generic platform that can handle a wide variety of use cases. We don’t focus on security as-is, but we do have customers that, in part of their data, we’re able to detect all sort of security-related breaches, like bot activity happening on a site or fraud rings—not the individual fraud of an individual person doing a transaction—but, it’s a lot of the time, frauds are not just one credit card, but somebody actually doing it over time, and then you can create or you can identify those fraud rings.
Most of our use cases have been around more the business-related data, either in ecommerce, ad tech companies, online services. And so online services, anybody that is really data-dependent to run their business, and very data-driven in running their business, and most businesses are transforming into that, even the old-fashioned businesses are transforming into that, because that data has competitive advantage, and being able to process that data to find all the anomalies, gives you an even larger competitive advantage.
So, last question: You made a comment earlier about freeing up people so we can focus on living in the stars. People who say that are generally science fiction fans I’ve noticed. If that is true, what view of the future, as expressed in science fiction, do you think is compelling or interesting or could happen?
That’s a great question. I think that that, what’s compelling to me about the future, really, is not whether we live in the stars or not in the stars, but really about having to free up our time to thinkabout stars, to thinkabout the next big things that progress humanity to the next levels, to be able to explore new dimensions and solve new problems, that…
Seek out new life and new civilizations…
Could be, and it could be in the stars, it could be on Earth, it could be just having more time, having more time on your hands, gives you more time to think about “What’s next?” When you’re busy surviving, then you don’t have any time to think about art, and think about music, and advancing it, or think about the stars, or think about the oceans, so, that’s the way I see AI and technology helping us—really freeing up our time to do more, and to use our collective intelligence and individual intelligence to imagine places that we haven’t thought about before… Or we don’t have time to think about before because we’re busy doing the mundane tasks. That’s really for me, what it’s all about…
Well that is a great place to end it, Ira. I want to thank you for taking the time and going on that journey with me of talking about all these different topics. It’s such an exciting time we live in and your reflections on them are fascinating, so thank you again..
Thank you very much, bye-bye.
Byron explores issues around artificial intelligence and conscious computers in his new book The Fourth Age: Smart Robots, Conscious Computers, and the Future of Humanity.
[voices_in_ai_link_back]

Voices in AI – Episode 26: A Conversation with Peter Lee

[voices_in_ai_byline]
In this episode, Byron and Peter talk about defining intelligence, Venn diagrams, transfer learning, image recognition, and Xiaoice.
[podcast_player name=”Episode 26: A Conversation with Peter Lee” artist=”Byron Reese” album=”Voices in AI” url=”https://voicesinai.s3.amazonaws.com/2017-12-04-(01-04-41)-peter-lee.mp3″ cover_art_url=”https://voicesinai.com/wp-content/uploads/2017/12/voices-headshot-card_preview-1-1.jpeg”]
[voices_in_ai_byline]
Byron Reese:  This is Voices in AI, brought to you by Gigaom. I’m Byron Reese. Today our guest is Peter Lee. He is a computer scientist and corporate Vice President at Microsoft Research. He leads Microsoft’s New Experiences and Technologies organization, or NExT, with the mission to create research powered technology and products and advance human knowledge through research. Prior to Microsoft, Dr. Lee held positions in both government and academia. At DARPA, he founded a division focused on R&D programs in computing and related areas. Welcome to the show, Peter. 
Peter Lee:  Thank you. It’s great to be here.
I always like to start with a seemingly simple question which turns out not to be quite so simple. What is artificial intelligence?
Wow. That is not a simple question at all. I guess the simple, one line answer is artificial intelligence is the science or the study of intelligent machines. And, I realize that definition is pretty circular, and I am guessing that you understand that that’s the fundamental difficulty, because it leaves open the question: what is intelligence? I think people have a lot of different ways to think about what is intelligence, but, in our world, intelligence is, “how do we compute how to set and achieve goals in the world.” And this is fundamentally what we’re all after, right now in AI.
That’s really fascinating because you’re right, there is no consensus definition on intelligence, or on life, or on death for that matter. So, I would ask that question: why do you think we have such a hard time defining what intelligence is?
I think we only have one model of intelligence, which is our own, and so when you think about trying to define intelligence it really comes down to a question of defining who we are. There’s fundamental discomfort with that. That fundamental circularity is difficult. If we were able to fly off in some starship to a far-off place, and find a different form of intelligence—or different species that we would recognize as intelligent—maybe we would have a chance to dispassionately study that, and come to some conclusions. But it’s a hard when you’re looking at something so introspective.
When you get into computer science research, at least here at Microsoft Research, you do have to find ways to focus on specific problems; so, we ended up focusing our research in AI—and our tech development in AI, roughly speaking—in four broad categories, and I think these categories are a little bit easier to grapple with. One is perception—that’s endowing machines with the ability to see and hear, much like we do. The second category is learning—how to get machines to get better with experience? The third is reasoning—how do you make inferences, logical inferences, commonsense inferences about the world? And then the fourth is language—how do we get machines to be intelligent in interacting with each other and with us through language? Those four buckets—perception, learning, reasoning and language—they don’t define what is intelligence, but they at least give us some kind of clear set of goals and directions to go after.
Well, I’m not going to spend too much time down in those weeds, but I think it’s really interesting. In what sense do you think it’s artificial? Because it’s either artificial in that it’s just mechanicalor that’s just a shorthand we use for thator it’s artificial in that it’s not really intelligence. You’re using words like “see,” “hear,” and reason.” Are you using those words euphemistically—can a computer really see or hear anything, or can it reason—or are you using them literally?
The question you’re asking really gets to the nub of things, because we really don’t know. If you were to draw the Venn diagram; you’d have a big circle and call that intelligence, and now you want to draw a circle for artificial intelligence—we don’t know if that circle is the same as the intelligence circle, whether it’s separate but overlapping, whether it’s a subset of intelligence… These are really basic questions that we debate, and people have different intuitions about, but we don’t really know. And then we get to what’s actually happening—what gets us excited and what is actually making it out into the real world, doing real things—and for the most part that has been a tiny subset of these big ideas; just focusing on machine learning, on learning from large amounts of data, models that are actually able to do some useful task, like recognize images.
Right. And I definitely want to go deep into that in just a minute, but I’m curious So, there’s a wide range of views about AI. Should we fear it? Should we love it? Will it take us into a new golden age? Will it do this? Will it cap out? Is an AGI possible? All of these questions. 
And, I mean, if you ask, “How will we get to Mars? Well, we don’t know exactly, but we kind of know. But if you ask, “What’s AI going to be like in fifty years?” it’s all over the map. And do you think that is because there isn’t agreement on the kinds of questions I’m askinglike people have different ideas on those questionsor are the questions I’m asking not really even germane to the day-to-day “get up and start building something”? 
I think there’s a lot of debate about this because the question is so important. Every technology is double-edged. Every technology has the ability to be used for both good purposes and for bad purposes, has good consequences and unintended consequences. And what’s interesting about computing technologies, generally, but especially with a powerful concept like artificial intelligence, is that in contrast to other powerful technologies—let’s say in the biological sciences, or in nuclear engineering, or in transportation and so on—AI has the potential to be highly democratized, to be codified into tools and technologies that literally every person on the planet can have access to. So, the question becomes really important: what kind of outcomes, what kinds of possibilities happen for this world when literally every person on the planet can have the power of intelligent machines at their fingertips? And because of that, all of the questions you’re asking become extremely large, and extremely important for us. People care about those futures, but ultimately, right now, our state of scientific knowledge is we don’t really know.
I sometimes talk in analogy about way, way back in the medieval times when Gutenberg invented mass-produced movable type, and the first printing press. And in a period of just fifty years, they went from thirty thousand books in all of Europe, to almost thirteen million books in all of Europe. It was sort of the first technological Moore’s Law. The spread of knowledge that that represented, did amazing things for humanity. It really democratized access to books, and therefore to a form of knowledge, but it was also incredibly disruptive in its time and has been since.
In a way, the potential we see with AI is very similar, and maybe even a bigger inflection point for humanity. So, while I can’t pretend to have any hard answers to the basic questions that you’re asking about the limits of AI and the nature of intelligence, it’s for sure important; and I think it’s a good thing that people are asking these questions and they’re thinking hard about it.
Well, I’m just going to ask you one more and then I want to get more down in the nitty-gritty. 
If the only intelligent thing we know of in the universe, the only general intelligence, is our brain, do you think it’s a settled question that that functionality can be reproduced mechanically? 
I think there is no evidence to the contrary. Every way that we look at what we do in our brains, we see mechanical systems. So, in principle, if we have enough understanding of how our own mechanical system of the brain works, then we should be able to, at a minimum, reproduce that. Now, of course, the way that technology develops, we tend to build things in different ways, and so I think it’s very likely that the kind of intelligent machines that we end up building will be different than our own intelligence. But there’s no evidence, at least so far, that would be contrary to the thesis that we can reproduce intelligence mechanically.
So, to say to take the opposite position for a moment. Somebody could say there’s absolutely no evidence to suggest that we can, for the following reasons. One, we don’t know how the brain works. We don’t know how thoughts are encoded. We don’t know how thoughts are retrieved. Aside from that, we don’t know how the mind works. We don’t know how it is that we have capabilities that seem to be beyond what a hunk of grey matter could dowe’re creative, we have a sense of humor and all these other things. We’re conscious, and we don’t even have a scientific language for understanding how consciousness could come about. We don’t even know how to ask that question or look for that answer, scientifically. So, somebody else might look at it and say, “There’s no reason whatsoever to believe we can reproduce it mechanically. 
I’m going to use a quote here from, of all people, a non-technologist Samuel Goldwyn, the old movie magnate. And I always reach to this when I get put in a corner like you’re doing to me right now, which is, “It’s absolutely impossible, but it has possibilities.”
All right.
Our current understanding is that brains are fundamentally closed systems, and so we’re learning more and more, and in fact what we learn is loosely inspiring some of the things we’re doing in AI systems, and making progress. How far that goes? It’s really, as you say, it’s unclear because there are so many mysteries, but it sure looks like there are a lot of possibilities.
Now to get kind of down to the nitty-gritty, let’s talk about difficulties and where we’re being successful and where we’re not. My first question is, why do you think AI is so hard? Because humans acquire their intelligence seemingly simply, right? You put a little kid in playschool and you show them some red, and you show them the number three, and then, all of a sudden, they understand what three red things are. I mean, we, kind of, become intelligent so naturally, and yet my frequent flyer program that I call in can’t tell, when I’m telling it my number if I said 8 or H. Why do you think it’s so hard?
What you said is true, although it took you many years to reach that point. And even a child that’s able to do the kinds of things that you just expressed has had years of life. The kinds of expectations that we have, at least today—especially in the commercial sphere for our intelligent machines—sometimes there’s a little bit less patience. But having said that, I think what you’re saying is right.
I mentioned before this Venn diagram; so, there’s this big circle which is intelligence, and let’s just assume that there is some large subset of that which is artificial intelligence. Then you zoom way, way in, and a tiny little bubble inside that AI bubble is machine learning—this is just simply machines that get better with experience. And then a tiny bubble inside that tiny bubble is machine learning from data—where the models that are extracted, that codify what has been learned, are all extracted from analyzing large amounts of data. That’s really where we’re at today—in this tiny bubble, inside this tiny bubble, inside this big bubble we call artificial intelligence.
What is remarkable is that, despite how narrow our understanding is—for the most part all of the exciting progress is just inside this little, tiny, narrow idea of machine learning from data, and there’s even a smaller bubble inside that that’s called a supervised manner—even from that we’re seeing tremendous power, a tremendous ability to create new computing systems that do some pretty impressive and valuable things. It is pretty crazy just how valuable that’s become to companies, like Microsoft. At the same time, it is such a narrow little slice of what we understand of intelligence.
The simple examples that you mentioned, for example, like one-shot learning, where you can show a small child a cartoon picture of a fire truck, and even if that child has never seen a fire truck before in her life, you can take her out on the street, and the first real fire truck that goes down the road the child will instantly recognize as a fire truck. That sort of one-shot idea, you’re right, our current systems aren’t good at.
While we are so excited about how much progress we’re making on learning from data, there are all the other things that are wrapped up in intelligence that are still pretty mysterious to us, and pretty limited. Sometimes, when that matters, our limits get in the way, and it creates this idea that AI is actually still really hard.
You’re talking about transfer learning. Would you say that the reason she can do that is because at another time she saw a drawing of a banana, and then a banana? And another time she saw a drawing of a cat, and then a cat. And so, it wasn’t really a one-shot deal. 
How do you think transfer learning works in humans? Because that seems to be what we’re super good at. We can take something that we learned in one place and transfer that knowledge to another contextYou know, “Find, in this picture, the Statue of Liberty covered in peanut butter,” and I can pick that out having never seen a Statue of Liberty in peanut butter, or anything like that. 
Do you think that’s a simple trick we don’t understand how to do yet? Is that what you want it to be, like an “a-ha” moment, where you discover the basic idea. Or do you think it’s a hundred tiny little hacks, and transfer learning in our minds is just, like, some spaghetti code written by some drunken programmer who was on a deadline, right? What do you think that is? Is it a simple thing, or is it a really convoluted, complicated thing? 
Transfer learning turns out to be incredibly interesting, scientifically, and also commercially for Microsoft, turns out to be something that we rely on in our business. What is kind of interesting is, when is transfer learning more generally applicable, versus being very brittle?
For example, in our speech processing systems, the actual commercial speech processing systems that Microsoft provides, we use transfer learning, routinely. When we train our speech systems to understand English speech, and then we train those same systems to understand Portuguese, or Mandarin, or Italian, we get a transfer learning effect, where the training for that second, and third, and fourth language requires less data and less computing power. And at the same time, each subsequent language that we add onto it improves the earlier languages. So, training that English-based system to understand Portuguese actually improves the performance of our speech systems in English, so there are transfer learning effects there.
In our image recognition tasks, there is something called the ImageNet competition that we participate in most years, and the last time that we competed was two years ago in 2015. There are five image processing categories. We trained our system to do well on Category 1—on the basic image classification—then we used transfer learning to not only win the first category, but to win all four other ImageNet competitions. And so, without any further kind of specialized training, there was a transfer learning effect.
Transfer learning actually does seem to happen. In our deep neural net, deep learning research activities, transfer learning effects—when we see them—are just really intoxicating. It makes you think about what you and I do as human beings.
At the same time, it seems to be this brittle thing. We don’t necessarily understand when and how this transfer learning effect is effective. The early evidence from studying these things is that there are different forms of learning, and that somehow the one-shot ideas that even small children are very good at, seem to be out of the purview of the deep neural net systems that we’re working on right now. Even this intuitive idea that you’ve expressed of transfer learning, the fact is we see it in some cases and it works so well and is even commercially-valuable to us, but then we also see simple transfer learning tasks where these systems just seem to fail. So, even those things are kind of mysterious to us right now.
It seemsand I don’t have any evidence to support this, but it seems, at a gut level to methat maybe what you’re describing isn’t pure transfer learning, but rather what you’re saying is, “We built a system that’s really good at translating languages, and it works on a lot of different languages.” 
It seems to me that the essence of transfer learning is when you take it to a different discipline, for example, “Because I learned a second language, I am now a better artist. Because I learned a second language, I’m now a better cook.” That, somehow, we take things that are in a discipline, and they add to this richness and depth and indimensionality of our knowledge in a way that they really impact our relationships. 
I was chatting with somebody the other day who said that learning a second language was the most valuable thing he’d ever done, and that his personality in that second language is different than his English personality. I hear what you’re saying, and I think those are hits that point us in the right direction. But I wonder if, at its core, it’s really multidimensional, what humans do, and that’s why we can seemingly do the one-shot things, because we’re taking things that are absolutely unrelated to cartoon drawings of something relating to real life. Do you have even any kind of a gut reaction to that?
One thing, at least in our current understanding of the research fields, is that there is a difference between learning and reasoning. The example I like to go to is, we’ve done quite a bit of work on language understanding, and specifically in something called machine reading—where you want to be able to read text and then answer questions about the text. And a classic place where you look to test your machine reading capabilities is parts of the verbal part of the SAT exam. The nice thing about the SAT exam is you can try to answer the questions and you can measure the progress just through the score that you get on the test. That’s steadily improving, and not just here at Microsoft Research, but at quite a few great university research areas and centers.
Now, subject those same systems to, say, the third-grade California Achievement Test, and the intelligence systems just fall apart. If you look at what third graders are expected to be able to do, there is a level of commonsense reasoning that seems to be beyond what we try to do in our machine reading system. So, for example, one kind of question you’ll get on that third-grade achievement test is, maybe, four cartoon drawings: a ball sitting on the grass, some raindrops, an umbrella, and a puppy dog—and you have to know which pairs of things go together. Third-graders are expected to be able to make the right logical inferences from having the right life experiences, the right commonsense reasoning inferences to put those two pairs together, but we don’t actually have the AI systems that, reliably, are able to do that. That commonsense reasoning is something that seems to be—at least today, with the state of today’s scientific and technological knowledge—outside of the realm of machine learning. It’s not something that we think machine learning will ultimately be effective at.
That distinction is important to us, even commercially. I’m looking at an e-mail today that someone here at Microsoft sent me to get ready to talk to you today. The e-mail says, it’s right in front of me here, “Here is the briefing doc for tomorrow morning’s podcast. If you want to review it tonight, I’ll print it for you tomorrow.” Right now, the system has underlined, “want to review tonight,” and the reason it’s underlined that is it’s somehow made the logical commonsense inference that I might want a reminder on my calendar to review the briefing documents. But it’s remarkable that it’s managed to do that, because there are references to tomorrow morning as well as tonight. So, making those sorts of commonsense inferences, doing that reasoning, is still just incredibly hard, and really still requires a lot of craftsmanship by a lot of smart researchers to make real.
It’s interesting because you say, you had just one line in there that solving the third-grade problem isn’t a machine learning task, so how would we solve that? Or put another way, I often ask these Turing Test systems, “What’s bigger, a nickel or the sun?” and none of them have ever been able to answer it. Because “sun is ambiguous, maybe, and nickel is ambiguous. 
In any case, if we don’t use machine learning for those, how do we get to the third grade? Or do we not even worry about the third grade? Because most of the problems we have in life aren’t third-grade problems, they’re 12th-grade problems that we really want the machines to be able to do. We want them to be able to translate documents, not match pictures of puppies. 
Well, for sure, if you just look at what companies like Microsoft, and the whole tech industry, are doing right now, we’re all seeing, I think, at least a decade, of incredible value to people in the world just with machine learning. There are just tremendous possibilities there, and so I think we are going to be very focused on machine learning and it’s going to matter a lot. It’s going to make people’s lives better, and it’s going to really provide a lot of commercial opportunities for companies like Microsoft. But that doesn’t mean that commonsense reasoning isn’t crucial, isn’t really important. Almost any kind of task that you might want help with—even simple things like making travel arrangements, shopping, or bigger issues like getting medical advice, advice about your own education—these things almost always involve some elements of what you would call commonsense reasoning, making inferences that somehow are not common, that are very particular and specific to you, and maybe haven’t been seen before in exactly that way.
Now, having said that, in the scientific community, in our research and amongst our researchers, there’s a lot of debate about how much of that kind of reasoning capability could be captured through machine learning, and how much of it could be captured simply by observing what people do for long enough and then just learning from it. But, for me at least, I see what is likely is that there’s a different kind of science that we’ll need to really develop much further if we want to capture that kind of commonsense reasoning.
Just to give you a sense of the debate, one thing that we’ve been doing—it’s been an experiment ongoing in China—is we have a new kind of chatbot technology in China that takes the form of a person named Xiaolce. Xiaolce is a persona that lives on social media in China, and actually has a large number of followers, tens of millions of followers.
Typically, when we think about chatbots and intelligent agents here in the US market—things like Cortana, or Siri, or Google Assistant, or Alexa—we put a lot of emphasis on semantic understanding; we really want the chatbot to understand what you’re saying at the semantic level. For Xiaolce, we ran a different experiment, and instead of trying to put in that level of semantic understanding, we instead looked at what people say on social media, and we used natural language processing to pick out statement response pairs, and templatize them, and put them in a large database. And so now, if you say something to Xiaolce in China, Xiaolce looks at what other people say in response to an utterance like that. Maybe it’ll come up with a hundred likely responses based on what other people have done, and then we use machine learning to rank order those likely responses, trying to optimize the enjoyment and engagement in the conversation, optimize the likelihood that the human being who is engaged in the conversation will stick with a conversation. Over time, Xiaolce has become extremely effective at doing that. In fact, for the top, say, twenty million people who interact with Xiaolce on a daily basis, the conversations are taking more than twenty-three turns.
What’s remarkable about that—and fuels the debate about what’s important in AI and what’s important in intelligence—is that at least the core of Xiaolce really doesn’t have any understanding at all about what you’re talking about. In a way, it’s just very intelligently mimicking what other people do in successful conversations. It raises the question, when we’re talking about machines and machines that at least appear to be intelligent, what’s really important? Is it really a purely mechanical, syntactic system, like we’re experimenting with Xiaolce, or is it something where we want to codify and encode our semantic understanding of the world and the way it works, the way we’re doing, say, with Cortana.
These are fundamental debates in AI. What’s sort of cool, at least in my day-to-day work here at Microsoft, is we are in a position where we’re able, and allowed, to do fundamental research in these things, but also build and deploy very large experiments just to see what happens and to try to learn from that. It’s pretty cool. At the same time, I can’t say that leaves me with clear answers yet. Not yet. It just leaves me with great experiences and we’re sharing what we’re learning with the world but it’s much, much harder to then say, definitively, what these things mean.
You know, it’s true. In 1950 Alan Turing said, “Can a machine think?” And that’s still a question that many can’t agree on because they don’t necessarily agree on the terms. But you’re right, that chatbot could pass the Turing Test, in theory. At twenty-three turns, if you didn’t tell somebody it was a chatbot, maybe it would pass it. 
But you’re right that that’s somehow unsatisfying that this is somehow this big milestone. Because if you saw it as a user in slow motionthat you ask a question, and then it did a query, and then it pulled back a hundred things and it rank ordered them, and looked for how many of those had successful follow-ups, and thumbs up, and smiley faces, and then it gave you one… It’s that whole thing about, once you know how the magic trick works, it isn’t nearly as interesting. 
It’s true. And with respect to achieving goals, or completing tasks in the world with the help of the Xiaolce chatbot, well, in some cases it’s pretty amazing how helpful Xiaolce is to people. If someone says, “I’m in the market for a new smartphone, I’m looking for a larger phablet, but I still want it to fit in my purse,” Xiaolce is amazingly effective at giving you a great answer to that question, because it’s something that a lot of people talk about when they’re shopping for a new phone.
At the same time, Xiaolce might not be so good at helping you decide which hotels to stay in, or helping you arrange your next vacation. It might provide some guidance, but maybe not exactly the right guidance that’s been well thought out. One more thing to say about this is, today—at least at the scale and practicality that we’re talking about—for the most part, we’re learning from data, and that data is essentially the digital exhaust from human thought and activity. There’s also another sense in which Xiaolce, while it passes the Turing Test, it’s also, in some ways, limited by human intelligence, because almost everything it’s able to do is observed and learned from what other people have done. We can’t discount the possibility of future systems which are less data dependent, and are able to just understand the structure of the world, and the problems, and learn from that.
Right. I guess Xiaolce wouldn’t know the difference, “What’s bigger, a nickel or the sun? 
That’s right, yes.
Unless the transcript of this very conversation were somehow part of the training set, but you notice, I’ve never answered it. I’ve never given the answer away, so, it still wouldn’t know
We should try the experiment at some point.
Why do you think we personify these AIs? You know about Weizenbaum and ELIZA and all of that, I assume. He got deeply disturbed when people were relating to a lie, knowing it was a chatbot. He got deeply concerned that people poured out their heart to it, and he said that when the machine says, “I understand,” it’s just a lie. That there’s no “I,” and there’s nothing that understands anything. Do you think that somehow confuses relationships with people and that there are unintended consequences to the personification of these technologies that we don’t necessarily know about yet? 
I’m always internally scolding myself for falling into this tendency to anthropomorphize our machine learning and AI systems, but I’m not alone. Even the most hardened, grounded researcher and scientist does this. I think this is something that is really at the heart of what it means to be human. The fundamental fascination that we have and drive to propagate our species is surfaced as a fascination with building autonomous intelligent beings. It’s not just AI, but it goes back to the Frankenstein kinds of stories that have just come up in different guises, and different forms throughout, really, all of human history.
I think we just have a tremendous drive to build machines, or other objects and beings, that somehow capture and codify, and therefore promulgate, what it means to be human. And nothing defines that more for us than some sort of codification of human intelligence, and especially human intelligence that is able to be autonomous, make its own decisions, make its own choices moving forward. It’s just something that is so primal in all of us. Even in AI research, where we really try to train ourselves and be disciplined about not making too many unfounded connections to biological systems, we fall into the language of biological intelligence all the time. Even the four categories I mentioned at the outset of our conversation—perception, learning, reasoning, language—these are pretty biologically inspired words. I just think it’s a very deep part of human nature.
That could well be the case. I have a book coming out on AI in April of 2018 that talks about these questions, and there’s a whole chapter about how long we’ve been doing this. And you’re right, it goes back to the Greeks, and the eagle that allegedly plucked out Prometheus’ liver every day, in some accounts, was a robot. There’s just tons of them. The difference of course, now, is that, up until a few years ago, it was all fiction, and so these were just stories. And we don’t necessarily want to build everything that we can imagine in fiction. I still wrestle with it, that, somehow, we are going to convolute humans and machines in a way which might be to the detriment of humans, and not to the ennobling of the machine, but time will tell. 
Every technology, as we discussed earlier, is double-edged. Just to strike an optimistic note here—to your last comment, which is, I think, very important—I do think that this is an area where people are really thinking hard about the kinds of issues you just raised. I think that’s in contrast to what was happening in computer science and the tech industry even just a decade ago, where there’s more or less an ethos of, “Technology is good and more technology is better.” I think now there’s much more enlightenment about this. I think we can’t impede the progress of science and technology development, but what is so good and so important is that, at least as a society, we’re really trying to be thoughtful about both the potential for good, as well as the potential for bad that comes out of all of this. I think that gives us a much better chance that we’ll get more of the good.
I would agree. I think the only other corollary to this, where there’s been so much philosophical discussion about the implications of the technology, is the harnessing of the atom. If you read the contemporary literature written at the time, people were like, “It could be energy too cheap to meter, or it could be weapons of colossal destruction, or it could be both. There was a precedent there for a long and thoughtful discussion about the implications of the technology. 
It’s funny you mentioned that because that reminds me of another favorite quote of mine which is from Albert Einstein, and I’m sure you’re familiar with it. “The difference between stupidity and genius is that genius has its limits.”
That’s good. 
And of course, he said that at the same time that a lot of this was developing. It was a pithy way to tell the scientific community, and the world, that we need to be thoughtful and careful. And I think we’re doing that today. I think that’s emerging very much so in the field of AI.
There’s a lot of practical concern about the effect of automation on employment, and these technologies on the planet. Do you have an opinion on how that’s all going to unfold? 
Well, for sure, I think it’s very likely that there’s going to be massive disruptions in how the world works. I mentioned the printing press, the Gutenberg press, movable type; there was incredible disruption there. When you have nine doublings in the spread of books and printing presses in the period of fifty years, that’s a real medieval Moore’s Law. And if you think about the disruptive effect of that, by the early 1500s, the whole notion of what it meant to educate your children suddenly involved making sure that they could read and write. That’s a skill that takes a lot of expense, and years of formal training and it has this sort of destructive impact. So, while the overall impact on the world and society was hugely positive—really the printing press laid the foundation for the Age of Enlightenment and the Renaissance—it had an absolutely disruptive effect on what it meant and what it took for people to succeed in the world.
AI, I’m pretty sure, is going to have the same kind of disruptive effect, because it has the same sort of democratizing force that the spread of books has had. And so, for us, we’ve been trying very hard to keep the focus on, “What can we do to put AI in the hands of people, that really empowers them, and augments what they’re able to do? What are the codifications of AI technologies that enable people to be more successful in whatever they’re pursuing in life?” And that focus, that intent by our research labs and by our company, I think, is incredibly important, because it takes a lot of the inventive and innovative genius that we have access to, and tries to point it in the right direction.
Talk to me about some of the interesting work you’re doing right now. Start with the healthcare stuff, what can you tell us about that?
Healthcare is just incredibly interesting. I think there are maybe three areas that just really get me excited. One is just fundamental life sciences, where we’re seeing some amazing opportunities and insights being unlocked through the use of machine learning, large-scale machine, and data analytics—the data that’s being produced increasingly cheaply through, say, gene sequencing, and through our ability to measure signals in the brain. What’s interesting about these things is that, over and over again, in other areas, if you put great innovative research minds and machine learning experts together with data and computing infrastructure, you get this burst of unplanned and unexpected innovations. Right now, in healthcare, we’re just getting to the point where we’re able to arrange the world in such a way that we’re able to get really interesting health data into the hands of these innovators, and genomics is one area that’s super interesting there.
Then, there is the basic question of, “What happens in the day-to-day lives of doctors and nurses?” Today, doctors are spending an average—there are several recent studies about this—of one hundred and eight minutes a day just entering health data into electronic health record systems. This is an incredible burden on those doctors, though it’s very important because it’s managed to digitize people’s health histories. But we’re now seeing an amazing ability for intelligent machines to just watch and listen to the conversation that goes on between the doctor and the patient, and to dramatically reduce the burden of all of that record keeping on doctors. So, doctors can stop being clerks and record keepers, and instead actually start to engage more personally with their patients.
And then the third area which I’m very excited about, but maybe is a little more geeky, is determining how we can create a system, how can we create a cloud, where more data is open to more innovators, where great researchers at universities, great innovators at startups who really want to make a difference in health, can provide a platform and a cloud where we can supply them with access to lots of valuable data, so they can innovate, they can create models that do amazing things.
Those three things just all really get me excited because the combination of these things I think can really make the lives of doctors, and nurses, and other clinicians better; can really lead to new diagnostics and therapeutic technologies, and unleash the potential of great minds and innovators. Stepping back for a minute, it really just amounts to creating systems that allow innovators, data, and computing infrastructure to all come together in one place, and then just having the faith that when you do that, great things will happen. Healthcare is just a huge opportunity area for doing this, that I’ve just become really passionate about.
I guess we will reach a point where you can have essentially the very best doctor in the world in your smartphone, and the very best psychologist, and the very best physical therapist, and the very best everything, right? All available at essentially no cost. I guess the internet always provided, at some abstract level, all of that information if you had an infinite amount of time and patience to find it. And the promise of AI, the kinds of things you’re doing, is that it was that difference, what did you say, between learning and reasoning, that it kind of bridges that gap. So, paint me a picture of what you think, just in the healthcare arena, the world of tomorrow will look like. What’s the thing that gets you excited? 
I don’t actually see healthcare ever getting away from being an essentially human-to-human activity. That’s something very important. In fact, I predict that healthcare will still be largely a local activity where it’s something that you will fundamentally access from another person in your locality. There are lots of reasons for this, but there’s something so personal about healthcare that it ends up being based in relationships. I see AI in the future relieving senseless and mundane burden from the heroes in healthcare—the doctors, and nurses, and administrators, and so on—that provide that personal service.
So, for example, we’ve been experimenting with a number of healthcare organizations with our chatbot technology. That chatbot technology is able to answer—on demand, through a conversation with a patient—routine and mundane questions about some health issue that comes up. It can do a, kind of, mundane textbook triage, and then, once all that is done, make an intelligent connection to a local healthcare provider, summarize very efficiently for the healthcare provider what’s going on, and then really allow the full creative potential and attention of the healthcare provider to be put to good use.
Another thing that we’ll be showing off to the world at a major radiology conference next week is the use of computer vision and machine learning to learn the habits and tricks of the trade for radiologists, that are doing radiation therapy planning. Right now, radiation therapy planning involves, kind of, a pixel by pixel clicking on radiological images that is extremely important; it has to be done precisely, but also has some artistry. Every good radiologist has his or her different kinds of approaches to this. So, one nice thing about the machine learning basic computer vision today, is that you can actually observe and learn what radiologists do, their practices, and then dramatically accelerate and relieve a lot of the mundane efforts, so that instead of two hours of work that is largely mundane with only maybe fifteen minutes of that being very creative, we can automate the noncreative aspects of this, and allow the radiologists to devote that full fifteen minutes, or even half an hour to really thinking through the creative aspects of radiology. So, it’s more of an empowerment model rather than replacing those healthcare workers. It still relies on human intuition; it still relies on human creativity, but hopefully allows more of that intuition, and more of that creativity to be harnessed by taking away some of the mundane, and time-consuming aspects of things.
These are approaches that I view as very human-focused, very humane ways to, not just make healthcare workers more productive, but to make them happier and more satisfied in what they do every day. Unlocking that with AI is just something that I feel is incredibly important. And it’s not just us here at Microsoft that are thinking this way, I’m seeing some really enlightened work going on, especially with some of our academic collaborators in this way. I find it truly inspiring to see what might be possible. Basically, I’m pushing back on the idea that we’ll be able to replace doctors, replace nurses. I don’t think that’s the world that we want, and I don’t even know that that’s the right idea. I don’t think that that necessarily leads to better healthcare.
To be clear, I’m talking about the great, immense parts of the world where there aren’t enough doctors for people, where there is this vast shortage of medical professionals, to somehow fill that gap, surely the technology can do that.
Yes. I think access is great. Even with some of the health chatbot pilot deployments that we’ve been experimenting with right now, you can just see that potential. If people are living in parts of the world where they have access issues, it’s an amazing and empowering thing to be able to just send a message to chatbot that’s always available and ready to listen, and answer questions. Those sorts of things, for sure, can make a big difference. At the same time, the real payoff is when technologies like that then enable healthcare workers—really great doctors, really great clinicians—to clear enough on their plate that their creative potential becomes available to more people; and so, you win on both ends. You win both on an instant access through automation, but you can also have a potential to win by expanding and enhancing the throughput and the number of patients that the clinics and clinicians can deal with. It’s a win-win situation in that respect.
Well said and I agree. It sounds like overall you are bullish on the future, you’re optimistic about the future and you think this technology overall is a force for great good, or am I just projecting that on to you? 
I’d say we think a lot about this. I would say, in my own career, I’ve had to confront both the good and bad outcomes, both the positive and unintended consequences of technology. I remember when I was back at DARPA—I arrived at DARPA in 2009—and in the summer of 2009, there was an election in Iran where the people in Iran felt that the results were not valid. This sparked what has been called the Iranian Twitter revolution. And what was interesting about the Iranian Twitter revolution is that people were using social media, Friendster and Twitter, in order to protest the results of this election and to organize protests.
This came to my attention at DARPA, through the State Department, because it became apparent that US-developed technologies to detect cyber intrusions and to help protect corporate networks were being used by the Iranian regime to hunt down and prosecute people who were using social media to organize these protests. The US took very quick steps to stop the sale of these technologies. But the thing that’s important is that these technologies, I’m pretty sure, were developed with only the best of intentions in mind—to help make computer networks safer. So, the idea that these technologies could be used to suppress free speech and freedom of assembly was, I’m sure never contemplated.
This really, kind of, highlights the double-edged nature of technology. So, for sure, we try to bring that thoughtfulness into every single research project we have across Microsoft Research, and that motivates our participation in things like the partnership on AI that involves a large number of industry and academic players, because we always want to have the technology, industry, and the research world be more and more thoughtful and enlightened on these ideas. So, yes, we’re optimistic. I’m optimistic certainly about the future, but that optimism, I think, is founded on a good dose of reality that if we don’t actually take proactive steps to be enlightened, on both the good and bad possibilities, good and bad outcomes, then the good things don’t just happen on their own automatically. So, it’s something that we work at, I guess, is the bottom line for what I’m trying to say. It’s earned optimism.
I like that. “Earned optimism,” I like that. It looks like we are out of time. I want to thank you for an hour of fascinating conversation about all of these topics. 
It was really fascinating, and you’ve asked some of the hardest question of the day. It was a challenge, and tons of fun to noodle on them with you.
Like, “What is bigger, the sun or a nickel? Turns out that’s a very hard question.
I’m going to ask Xiaolce that question and I’ll let you know what she says.
All right. Thank you again.
Thank you.
Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here.
[voices_in_ai_link_back]

Voices in AI – Episode 12: A Conversation with Scott Clark

[voices_in_ai_byline]
In this episode, Byron and Scott talk about algorithms, transfer learning, human intelligence, and pain and suffering.
[podcast_player name=”Episode 12: A Conversation with Scott Clark” artist=”Byron Reese” album=”Voices in AI” url=”https://voicesinai.s3.amazonaws.com/2017-10-16-(00-56-02)-scott-clark.mp3″ cover_art_url=”https://voicesinai.com/wp-content/uploads/2017/10/voices-headshot-card-4.jpg”]
[voices_in_ai_link_back]
Byron Reese: This is Voices in AI, brought to you by Gigaom. I’m Byron Reese. Today our guest is Scott Clark. He is the CEO and co-founder of SigOpt. They’re a SaaS startup for tuning complex systems and machine learning models. Before that, Scott worked on the ad targeting team at Yelp, leading the charge on academic research and outreach. He holds a PhD in Applied Mathematics and an MS in Computer Science from Cornell, and a BS in Mathematics, Physics, and Computational Physics from Oregon State University. He was chosen as one of Forbes 30 under 30 in 2016. Welcome to the show, Scott.
Scott Clark: Thanks for having me.
I’d like to start with the question, because I know two people never answer it the same: What is artificial intelligence?
I like to go back to an old quote… I don’t remember the attribution for it, but I think it actually fits the definition pretty well. Artificial intelligence is what machines can’t currently do. It’s the idea that there’s this moving goalpost for what artificial intelligence actually means. Ten years ago, artificial intelligence meant being able to classify images; like, can a machine look at a picture and tell you what’s in the picture?
Now we can do that pretty well. Maybe twenty, thirty years ago, if you told somebody that there would be a browser where you can type in words, and it would automatically correct your spelling and grammar and understand language, he would think that’s artificial intelligence. And I think there’s been a slight shift, somewhat recently, where people are calling deep learning artificial intelligence and things like that.
It’s got a little bit conflated with specific tools. So now people talk about artificial general intelligence as this impossible next thing. But I think a lot of people, in their minds, think of artificial intelligence as whatever it is that’s next that computers haven’t figured out how to do yet, that humans can do. But, as computers continually make progress on those fronts, the goalposts continually change.
I’d say today, people think of it as conversational systems, basic tasks that humans can do in five seconds or less, and then artificial general intelligence is everything after that. And things like spell check, or being able to do anomaly detection, are just taken for granted and that’s just machine learning now.
I’ll accept all of that, but that’s more of a sociological observation about how we think of it, and then actually… I’ll change the question. What is intelligence?
That’s a much more difficult question. Maybe the ability to reason about your environment and draw conclusions from it.
Do you think that what we’re building, our systems, are they artificial in the sense that we just built them, but they can do that? Or are they artificial in the sense that they can’t really do that, but they sure can think it well?
I think they’re artificial in the sense that they’re not biological systems. They seem to be able to perceive input in the same way that a human can perceive input, and draw conclusions based off of that input. Usually, the reward system in place in an artificial intelligence framework is designed to do a very specific thing, very well.
So is there a cat in this picture or not? As opposed to a human: It’s, “Try to live a fulfilling life.” The objective functions are slightly different, but they are interpreting outside stimuli via some input mechanism, and then trying to apply that towards a specific goal. The goals for artificial intelligence today are extremely short-term, but I think that they are performing them on the same level—or better sometimes—than a human presented with the exact same short-term goal.
The artificial component comes into the fact that they were constructed, non-biologically. But other than that, I think they meet the definition of observing stimuli, reasoning about an environment, and achieving some outcome.
You used the phrase ‘they draw conclusions’. Are you using that colloquially, or does the machine actually conclude? Or does it merely calculate?
It calculates, but then it comes to, I guess, a decision at the end of the day. If it’s a classification system, for example… going back to “Is there a cat in this picture?” It draws the conclusion that “Yes, there was a cat. No, that wasn’t a cat.” It can do that with various levels of certainty in the same way that, potentially, a human would solve the exact same problem. If I showed you a blurry Polaroid picture you might be able to say, “I’m pretty sure there’s a cat in there, but I’m not 100 percent certain.”
And if I show you a very crisp picture of a kitten, you could be like, “Yes, there’s a cat there.” And I think convolutional neural network is doing the exact same thing: taking in that outside stimuli. Not through an optical nerve, but through the raw encoding of pixels, and then coming to the exact same conclusion.
You make the really useful distinction between an AGI, which is a general intelligence—something as versatile as a human—and then the kinds of stuff we’re building now, which we call AI—which is doing this reasoning or drawing conclusions.
Is an AGI a linear development from what we have now? In other words, do we have all the pieces, and we just need faster computers, better algorithms, more data, a few nips and tucks, and we’re eventually going to get an AGI? Or is an AGI something very different, that is a whole different ball of wax?
I’m not convinced that, with the current tooling we have today, that it’s just like… if we add one more hidden layer to a neural network, all of a sudden it’ll be AGI. That being said, I think this is how science and computer science and progress in general works. Is that techniques are built upon each other, we make advancements.
It might be a completely new type of algorithm. It might not be a neural network. It might be reinforcement learning. It might not be reinforcement learning. It might be the next thing. It might not be on a CPU or a GPU. Maybe it’s on a quantum computer. If you think of scientific and technological process as this linear evolution of different techniques and ideas, then I definitely think we are marching towards that as an eventual outcome.
That being said, I don’t think that there’s some magic combinatorial setting of what we have today that will turn into this. I don’t think it’s one more hidden layer. I don’t think it’s a GPU that can do one more teraflop—or something like that—that’s going to push us over the edge. I think it’s going to be things built from the foundation that we have today, but it will continue to be new and novel techniques.
There was an interesting talk at the International Conference on Machine Learning in Sydney last week about AlphaGo, and how they got this massive speed-up when they put in deep learning. They were able to break through this plateau that they had found in terms of playing ability, where they could play at the amateur level.
And then once they started applying deep learning networks, that got them to the professional, and now best-in-the-world level. I think we’re going to continue to see plateaus for some of these current techniques, but then we’ll come up with some new strategy that will blast us through and get to the next plateau. But I think that’s an ever-stratifying process.
To continue on that vein… When in 1955, they convened in Dartmouth and said, “We can solve a big part of AI in the summer, with five people,” the assumption was that general intelligence, like all the other sciences, had a few simple laws.
You had Newton, Maxwell; you had electricity and magnetism, and all these things, and they were just a few simple laws. The idea was that all we need to do is figure out those for intelligence. And Pedro Domingos argues in The Master Algorithm, from a biological perspective that, in a sense, that may be true.  
That if you look at the DNA difference between us and an animal that isn’t generally intelligent… the amount of code is just a few megabytes that’s different, which teaches how to make my brain and your brain. It sounded like you were saying, “No, there’s not going to be some silver bullet, it’s going to be a bunch of silver buckshot and we’ll eventually get there.”
But do you hold any hope that maybe it is a simple and elegant thing?
Going back to my original statement about what is AI, I think when Marvin Minsky and everybody sat down in Dartmouth, the goalposts for AI were somewhat different. Because they were attacking it for the first time, some of the things were definitely overambitious. But certain things that they set out to do that summer, they actually accomplished reasonably well.
Things like the Lisp programming language, and things like that, came out of that and were extremely successful. But then, once these goals are accomplished, the next thing comes up. Obviously, in hindsight, it was overambitious to think that they could maybe match a human, but I think if you were to go back to Dartmouth and show them what we have today, and say: “Look, this computer can describe the scene in this picture completely accurately.”
I think that could be indistinguishable from the artificial intelligence that they were seeking, even if today what we want is someone we can have a conversation with. And then once we can have a conversation, the next thing is we want them to be able to plan our lives for us, or whatever it may be, solve world peace.
While I think there are some of the fundamental building blocks that will continue to be used—like, linear algebra and calculus, and things like that, will definitely be a core component of the algorithms that make up whatever does become AGI—I think there is a pretty big jump between that. Even if there’s only a few megabytes difference between us and a starfish or something like that, every piece of DNA is two bits.
If you have millions of differences, four-to-the-several million—like the state space for DNA—even though you can store it in a small amount of megabytes, there are so many different combinatorial combinations that it’s not like we’re just going to stumble upon it by editing something that we currently have.
It could be something very different in that configuration space. And I think those are the algorithmic advancements that will continue to push us to the next plateau, and the next plateau, until eventually we meet and/or surpass the human plateau.
You invoked quantum computers in passing, but putting that aside for a moment… Would you believe, just at a gut level—because nobody knows—that we have enough computing power to build an AGI, we just don’t know how?
Well, in the sense that if the human brain is general intelligence, the computing power in the human brain, while impressive… All of the computers in the world are probably better at performing some simple calculations than the biological gray matter mess that exists in all of our skulls. I think the raw amount of transistors and things like that might be there, if we had the right way to apply them, if they were all applied in the same direction.
That being said… Whether or not that’s enough to make it ubiquitous, or whether or not having all the computers in the world mimic a single human child will be considered artificial general intelligence, or if we’re going to need to apply it to many different situations before we claim victory, I think that’s up for semantic debate.
Do you think about how the brain works, even if [the context] is not biological? Is that how you start a problem: “Well, how do humans do this?” Does that even guide you? Does that even begin the conversation? And I know none of this is a map: Birds fly with wings, and airplanes, all of that. Is there anything to learn from human intelligence that you, in a practical, day-to-day sense, use?
Yeah, definitely. I think it often helps to try to approach a problem from fundamentally different ways. One way to approach that problem is from the purely mathematical, axiomatic way; where we’re trying to build up from first principles, and trying to get to something that has a nice proof or something associated with it.
Another way to try to attack the problem is from a more biological setting. If I had to solve this problem, and I couldn’t assume any of those axioms, then how would I begin to try to build heuristics around it? Sometimes you can go from that back to the proof, but there are many different ways to attack that problem. Obviously, there are a lot of things in computer science, and optimization in general, that are motivated by physical phenomena.
So a neural network, if you squint, looks kind of like a biological brain neural network. There’s things like simulated annealing, which is a global optimization strategy that mimics the way that like steel is annealed… where it tries to find some local lattice structure that has low energy, and then you pound the steel with the hammer, and that increases the energy to find a better global optima lattice structure that is harder steel.
But that’s also an extremely popular algorithm in the scientific literature. So it was come to from this auxiliary way, or a genetic algorithm where you’re slowly evolving a population to try to get to a good result. I think there is definitely room for a lot of these algorithms to be inspired by biological or physical phenomenon, whether or not they are required to be from that to be proficient. I would have trouble, off the top of my head, coming up with the biological equivalent for a support vector machine or something like that. So there’s two different ways to attack it, but both can produce really interesting results.
Let’s take a normal thing that a human does, which is: You show a human training data of the Maltese Falcon, the little statue from the movie, and then you show him a bunch of photos. And a human can instantly say, “There’s the falcon under water, and there it’s half-hidden by a tree, and there it’s upside down…” A human does that naturally. So it’s some kind of transferred learning. How do we do that?
Transfer learning is the way that that happens. You’ve seen trees before. You’ve seen water. You’ve seen how objects look inside and outside of water before. And then you’re able to apply that knowledge to this new context.
It might be difficult for a human who grew up in a sensory deprivation chamber to look at this object… and then you start to show them things that they’ve never seen before: “Here’s this object and a tree,” and they might not ‘see the forest for the trees’ as it were.
In addition to that, without any context whatsoever, you take someone who was raised in a sensory deprivation chamber, and you start showing them pictures and ask them to do classification type tasks. They may be completely unaware of what’s the reward function here. Who is this thing telling me to do things for the first time I’ve never seen before?
What does it mean to even classify things or describe an object? Because you’ve never seen an object before.
And when you start training these systems from scratch, with no previous knowledge, that’s how they work. They need to slowly learn what’s good, what’s bad. There’s a reward function associated with that.
But with no context, with no previous information, it’s actually very surprising how well they are able to perform these tasks; considering [that when] a child is born, four hours later it isn’t able to do this. A machine algorithm that’s trained from scratch over the course of four hours on a couple of GPUs is able to do this.
You mentioned the sensory deprivation chamber a couple of times. Do you have a sense that we’re going to need to embody these AIs to allow them to—and I use the word very loosely—‘experience’ the world? Are they locked in a sensory deprivation chamber right now, and that’s limiting them?
I think with transfer learning, and pre-training of data, and some reinforcement algorithm work, there’s definitely this idea of trying to make that better, and bootstrapping based off of previous knowledge in the same way that a human would attack this problem. I think it is a limitation. It would be very difficult to go from zero to artificial general intelligence without providing more of this context.
There’s been many papers recently, and OpenAI had this great blog post recently where, if you teach the machine language first, if you show it a bunch of contextual information—this idea of this unsupervised learning component of it, where it’s just absorbing information about the potential inputs it can get—that allows it to perform much better on a specific task, in the same way that a baby absorbs language for a long time before it actually starts to produce it itself.
And it could be in a very unstructured way, but it’s able to learn some of the actual language structure or sounds from the particular culture in which it was raised in this unstructured way.
Let’s talk a minute about human intelligence. Why do you think we understand so poorly how the brain works?
That’s a great question. It’s easier scientifically, with my background in math and physics—it seems like it’s easier to break down modular decomposable systems. Humanity has done a very good job at understanding, at least at a high level, how physical systems work, or things like chemistry.
Biology starts to get a little bit messier, because it’s less modular and less decomposable. And as you start to build larger and larger biological systems, it becomes a lot harder to understand all the different moving pieces. Then you go to the brain, and then you start to look at psychology and sociology, and all of the lines get much fuzzier.
It’s very difficult to build an axiomatic rule system. And humans aren’t even able to do that in some sort of grand unified way with physics, or understand quantum mechanics, or things like that; let alone being able to do it for these sometimes infinitely more complex systems.
Right. But the most successful animal on the planet is a nematode worm. Ten percent of all animals are nematode worms. They’re successful, they find food, and they reproduce and they move. Their brains have 302 neurons. We’ve spent twenty years trying to model that, a bunch of very smart people in the OpenWorm project…
 But twenty years trying to model 300 neurons to just reproduce this worm, make a digital version of it, and even to this day people in the project say it may not be possible.
I guess the argument is, 300 sounds like a small amount. One thing that’s very difficult for humans to internalize is the exponential function. So if intelligence grew linearly, then yeah. If we could understand one, then 300 might not be that much, whatever it is. But if the state space grows exponentially, or the complexity grows exponentially… if there’s ten different positions for every single one of those neurons, like 10300, that’s more than the number of atoms in the universe.
Right. But we aren’t starting by just rolling 300 dice and hoping for them all to be—we know how those neurons are arranged.
At a very high level we do.
I’m getting to a point, that we maybe don’t even understand how a neuron works. A neuron may be doing stuff down at the quantum level. It may be this gigantic supercomputer we don’t even have a hope of understanding, a single neuron.
From a chemical way, we can have an understanding of, “Okay, so we have neurotransmitters that carry a positive charge, that then cause a reaction based off of some threshold of charge, and there’s this catalyst that happens.” I think from a physics and chemical understanding, we can understand the base components of it, but as you start to build these complex systems that have this combinatorial set of states, it does become much more difficult.
And I think that’s that abstraction, where we can understand how simple chemical reactions work. But then it becomes much more difficult once you start adding more and more. Or even in physics… like if you have two bodies, and you’re trying to calculate the gravity, that’s relatively easy. Three? Harder. Four? Maybe impossible. It becomes much harder to solve these higher-order, higher-body problems. And even with 302 neurons, that starts to get pretty complex.
Oddly, two of them aren’t connected to anything, just like floating out there…
Do you think human intelligence is emergent?
In what respect?
I will clarify that. There are two sorts of emergence: one is weak, and one is strong. Weak emergence is where a system takes on characteristics which don’t appear at first glance to be derivable from them. So the intelligence displayed by an ant colony, or a beehive—the way that some bees can shimmer in unison to scare off predators. No bee is saying, “We need to do this.”  
The anthill behaves intelligently, even though… The queen isn’t, like, in charge; the queen is just another ant, but somehow it all adds intelligence. So that would be something where it takes on these attributes.
Can you really intuitively derive intelligence from neurons?
And then, to push that a step further, there are some who believe in something called ‘strong emergence’, where they literally are not derivable. You cannot look at a bunch of matter and explain how it can become conscious, for instance. It is what the minority of people believe about emergence, that there is some additional property of the universe we do not understand that makes these things happen.
The question I’m asking you is: Is reductionism the way to go to figure out intelligence? Is that how we’re going to kind of make advances towards an AGI? Just break it down into enough small pieces.
I think that is an approach, whether or not that’s ‘the’ ultimate approach that works is to be seen. As I was mentioning before, there are ways to take biological or physical systems, and then try to work them back into something that then can be used and applied in a different context. There’s other ways, where you start from the more theoretical or axiomatic way, and try to move forward into something that then can be applied to a specific problem.
I think there’s wide swaths of the universe that we don’t understand at many levels. Mathematics isn’t solved. Physics isn’t solved. Chemistry isn’t solved. All of these build on each other to get to these large, complex, biological systems. It may be a very long time, or we might need an AGI to help us solve some of these systems.
I don’t think it’s required to understand everything to be able to observe intelligence—like, proof by example. I can’t tell you why my brain thinks, but my brain is thinking, if you can assume that humans are thinking. So you don’t necessarily need to understand all of it to put it all together.
Let me ask you one more far-out question, and then we’ll go to a little more immediate future. Do you have an opinion on how consciousness comes about? And if you do or don’t, do you believe we’re going to build conscious machines?
Even to throw a little more into that one, do you think consciousness—that ability to change focus and all of that—is a requisite for general intelligence?
So, I would like to hear your definition of consciousness.
I would define it by example, to say that it’s subjective experience. It’s how you experience things. We’ve all had that experience when you’re driving, that you kind of space out, and then, all of a sudden, you kind of snap to. “Whoa! I don’t even remember getting here.”
And so that time when you were driving, your brain was elsewhere, you were clearly intelligent, because you were merging in and out of traffic. But in the sense I’m using the word, you were not ‘conscious’, you were not experiencing the world. If your foot caught on fire, you would feel it; but you weren’t experiencing the world. And then instantly, it all came on and you were an entity that experienced something.
Or, put another way… this is often illustrated with the problem of Mary by Frank Jackson:
He offers somebody named Mary, who knows everything about color, like, at a god-like level—knows every single thing about color. But the catch is, you might guess, she’s never seen it. She’s lived in a room, black-and-white, never seen it [color]. And one day, she opens the door, she looks outside and she sees red.  
The question becomes: Does she learn anything? Did she learn something new?  
In other words, is experiencing something different than knowing something? Those two things taken together, defining consciousness, is having an experience of the world…
I’ll give one final one. You can hook a sensor up to a computer, and you can program the computer to play an mp3 of somebody screaming if the sensor hits 500 degrees. But nobody would say, at this day and age, the computer feels the pain. Could a computer feel anything?
Okay. I think there’s a lot to unpack there. I think computers can perceive the environment. Your webcam is able to record the environment in the same way that your optical nerves are able to record the environment. When you’re driving a car, and daydreaming, and kind of going on autopilot, as it were, there still are processes running in the background.
If you were to close your eyes, you would be much worse at doing lane merging and things like that. And that’s because you’re still getting the sensory input, even if you’re not actively, consciously aware of the fact that you’re observing that input.
Maybe that’s where you’re getting at with consciousness here, is: Not only the actual task that’s being performed, which I think computers are very good at—and we have self-driving cars out on the street in the Bay Area every day—but that awareness of the fact that you are performing this task, is kind of meta-level of: “I’m assembling together all of these different subcomponents.”
Whether that’s driving a car, thinking about the meeting that I’m running late to, some fight that I had with my significant other the night before, or whatever it is. There’s all these individual processes running, and there could be this kind of global awareness of all of these different tasks.
I think today, where artificial intelligence sits is, performing each one of these individual tasks extremely well, toward some kind of objective function of, “I need to not crash this car. I need to figure out how to resolve this conflict,” or whatever it may be; or, “Play this game in an artificial intelligence setting.” But we don’t yet have that kind of governing overall strategy that’s aware of making these tradeoffs, and then making those tradeoffs in an intelligent way. But that overall strategy itself is just going to be going toward some specific reward function.
Probably when you’re out driving your car, and you’re spacing out, your overall reward function is, “I want to be happy and healthy. I want to live a meaningful life,” or something like that. It can be something nebulous, but you’re also just this collection of subroutines that are driving towards this specific end result.
But the direct question of what would it mean for a computer to feel pain? Will a computer feel pain? Now they can sense things, but nobody argues they have a self that experiences the pain. It matters, doesn’t it?
It depends on what you mean by pain. If you mean there’s a response of your nervous system to some outside stimuli that you perceive as pain, a negative response, and—
—It involves emotional distress. People know what pain is. It hurts. Can a computer ever hurt?
It’s a fundamentally negative response to what you’re trying to achieve. So pain and suffering is the opposite of happiness. And your objective function as a human is happiness, let’s say. So, by failing to achieve that objective, you feel something like pain. Evolutionarily, we might have evolved this in order to avoid specific things. Like, you get pain when you touch flame, so don’t touch flame.
And the reason behind that is biological systems degrade in high-temperature environments, and you’re not going to be able to reproduce or something like that.
You could argue that when a classification system fails to classify something, and it gets penalized in its reward function, that’s the equivalent of it finding something where, in its state of the world, it has failed to achieve its goal, and it’s getting the opposite of what its purpose is. And that’s similar to pain and suffering in some way.
But is it? Let’s be candid. You can’t take a person and torture them, because that’s a terrible thing to do… because they experience pain. [Whereas if] you write a program that has an infinite loop that causes your computer to crash, nobody’s going to suggest you should go to jail for that. Because people know that those are two very different things.
It is a negative neurological response based off of outside stimuli. A computer can have a negative response, and perform based off of outside stimuli poorly, relative to what it’s trying to achieve… Although I would definitely agree with you that that’s not a computer experiencing pain.
But from a pure chemical level, down to the algorithmic component of it, they’re not as fundamentally different… that because it’s a human, there’s something magic about it being a human. A dog can also experience pain.
These worms—I’m not as familiar with the literature on that, but [they] could potentially experience pain. And as you derive that further and further back, you might have to bend your definition of pain. Maybe they’re not feeling something in a central nervous system, like a human or a dog would, but they’re perceiving something that’s negative to what they’re trying to achieve with this utility function.
But we do draw a line. And I don’t know that I would use the word ‘magic’ the way you’re doing it. We draw this line by saying that dogs feel pain, so we outlaw animal cruelty. Bacteria don’t, so we don’t outlaw antibiotics. There is a material difference between those two things.
So if the difference is a central nervous system, and pain is being defined as a nervous response to some outside stimuli… then unless we explicitly design machines to have central nervous systems, then I don’t think they will ever experience pain.
Thanks for indulging me in all of that, because I think it matters… Because up until thirty years ago, veterinarians typically didn’t use anesthetic. They were told that animals couldn’t feel pain. Babies were operated on in the ‘90s—open heart surgery—under the theory they couldn’t feel pain.  
What really intrigues me is the idea of how would we know if a machine did? That’s what I’m trying to deconstruct. But enough of that. We’ll talk about jobs here in a minute, and those concerns…
There’s groups of people that are legitimately afraid of AI. You know all the names. You’ve got Elon Musk, you get Stephen Hawking. Bill Gates has thrown in his hat with that, Wozniak has. Nick Bostrom wrote a book that addressed existential threat and all of that. Then you have Mark Zuckerberg, who says no, no, no. You get Oren Etzioni over at the Allen Institute, just working on some very basic problem. You get Andrew Ng with his “overpopulation on Mars. This is not helpful to even have this conversation.”
What is different about those two groups in your mind? What is the difference in how they view the world that gives them these incredibly different viewpoints?
I think it goes down to a definition problem. As you mentioned at the beginning of this podcast, when you ask people, “What is artificial intelligence?” everybody gives you a different answer. I think each one of these experts would also give you a different answer.
If you define artificial intelligence as matrix multiplication and gradient descent in a deep learning system, trying to achieve a very specific classification output given some pixel input—or something like that—it’s very difficult to conceive that as some sort of existential threat for humanity.
But if you define artificial intelligence as this general intelligence, this kind of emergent singularity where the machines don’t hit the plateau, that they continue to advance well beyond humans… maybe to the point where they don’t need humans, or we become the ants in that system… that becomes very rapidly a very existential threat.
As I said before, I don’t think there’s an incremental improvement from algorithms—as they exist in the academic literature today—to that singularity, but I think it can be a slippery slope. And I think that’s what a lot of these experts are talking about… Where if it does become this dynamic system that feeds on itself, by the time we realize it’s happening, it’ll be too late.
Whether or not that’s because of the algorithms that we have today, or algorithms down the line, it does make sense to start having conversations about that, just because of the time scales over which governments and policies tend to work. But I don’t think someone is going to design a TensorFlow or MXNet algorithm tomorrow that’s going to take over the world.
There’s legislation in Europe to basically say, if an AI makes a decision about whether you should get an auto loan or something, you deserve to know why it turned you down. Is that a legitimate request, or is it like you go to somebody at Google and say, “Why is this site ranked number one and this site ranked number two?” There’s no way to know at this point.  
Or is that something that, with the auto loan thing, you’re like, “Nope, here are the big bullet points of what went into it.” And if that becomes the norm, does that slow down AI in any way?
I think it’s important to make sure, just from a societal standpoint, that we continue to strive towards not being discriminatory towards specific groups and people. It can be very difficult, when you have something that looks like a black box from the outside, to be able to say, “Okay, was this being fair?” based off of the fairness that we as a society have agreed upon.
The machine doesn’t have that context. The machine doesn’t have the policy, necessarily, inside to make sure that it’s being as fair as possible. We need to make sure that we do put these constraints on these systems, so that it meets what we’ve agreed upon as a society, in laws, etc., to adhere to. And that it should be held to the same standard as if there was a human making that same decision.
There is, of course, a lot of legitimate fear wrapped up about the effect of automation and artificial intelligence on employment. And just to set the problem up for the listeners, there’s broadly three camps, everybody intuitively knows this.
 There’s one group that says, “We’re going to advance our technology to the point that there will be a group of people who do not have the educational skills needed to compete with the machines, and we’ll have a permanent underclass of people who are unemployable.” It would be like the Great Depression never goes away.
And then there are people who say, “Oh, no, no, no. You don’t understand. Everything, every job, a machine is going to be able to do.” You’ll reach a point where the machine will learn it faster than the human, and that’s it.
And then you’ve got a third group that says, “No, that’s all ridiculous. We’ve had technology come along, as transformative as it is… We’ve had electricity, and machines replacing animals… and we’ve always maintained full employment.” Because people just learn how to use these tools to increase their own productivity, maintain full employment—and we have growing wages.
So, which of those, or a fourth one, do you identify with?
This might be an unsatisfying answer, but I think we’re going to go through all three phases. I think we’re in the third camp right now, where people are learning new systems, and it’s happening at a pace where people can go to a computer science boot camp and become an engineer, and try to retrain and learn some of these systems, and adapt to this changing scenario.
I think, very rapidly—especially at the exponential pace that technology tends to evolve—it does become very difficult. Fifty years ago, if you wanted to take apart your telephone and try to figure out how it works, repair it, that was something that a kid could do at a camp kind of thing, like an entry circuits camp. That’s impossible to do with an iPhone.
I think that’s going to continue to happen with some of these more advanced systems, and you’re going to need to spend your entire life understanding some subcomponent of it. And then, in the further future, as we move towards this direction of artificial general intelligence… Like, once a machine is a thousand times, ten thousand times, one hundred thousand times smarter—by whatever definition—than a human, and that increases at an exponential pace… We won’t need a lot of different things.
Whether or not that’s a fundamentally bad thing is up for debate. I think one thing that’s different about this than the Industrial Revolution, or the agricultural revolution, or things like that, that have happened throughout human history… is that instead of this happening over the course of generations or decades… Maybe if your father, and your grandfather, and your entire family tree did a specific job, but then that job doesn’t exist anymore, you train yourself to do something different.
Once it starts to happen over the course of a decade, or a year, or a month, it becomes much harder to completely retrain. That being said, there’s lots of thoughts about whether or not humans need to be working to be happy. And whether or not there could be some other fundamental thing that would increase the net happiness and fulfillment of people in the world, besides sitting at a desk for forty hours a week.
And maybe that’s actually a good thing, if we can set up the societal constructs to allow people to do that in a healthy and happy way.
Do you have any thoughts on computers displaying emotions, emulating emotions? Is that going to be a space where people are going to want authentic human experiences in those in the future? Or are we like, “No, look at how people talk to their dog,” or something? If it’s good enough to fool you, you just go along with the conceit?
The great thing about computers, and artificial intelligence systems, and things like that is if you point them towards a specific target, they’ll get pretty good at hitting that target. So if the goal is to mimic human emotion, I think that that’s something that’s achievable. Whether or not a human cares, or is even able to distinguish between that and actual human emotion, could be very difficult.
At Cornell, where I did my PhD, they had this psychology chatbot called ELIZA—I think this was back in the ‘70s. It went through a specific school of psychological behavioral therapy thought, replied with specific ways, and people found it incredibly helpful.
Even if they knew that it was just a machine responding to them, it was a way for them to get out their emotions and work through specific problems. As these machines get more sophisticated and able, as long as it’s providing utility to the end user, does it matter who’s behind the screen?
That’s a big question. Weizenbaum shut down ELIZA because he said that when a machine says, “I understand” that it’s a lie, there’s no ‘I’, and there’s nothing [there] that understands anything. He had real issues with that.
But then when they shut it down, some of the end users were upset, because they were still getting quite a bit of utility out of it. There’s this moral question of whether or not you can take away something from someone who is deriving benefit from it as well.
So I guess the concern is that maybe we reach a day where an AI best friend is better than a real one. An AI one doesn’t stand you up. And an AI spouse is better than a human spouse, because of all of those reasons. Is that a better world, or is it not?
I think it becomes a much more dangerous world, because as you said before, someone could decide to turn off the machine. When it’s someone taking away your psychologist, that could be very dangerous. When it’s someone deciding that you didn’t pay your monthly fee, so they’re going to turn off your spouse, that could be quite a bit worse as well.
As you mentioned before, people don’t necessarily associate the feelings or pain or anything like that with the machine, but as these get more and more life-like, and as they are designed with the reward function of becoming more and more human-like, I think that distinction is going to become quite a bit harder for us to understand.
And it not only affects the machine—which you can make the argument doesn’t have a voice—but it’ll start to affect the people as well.
One more question along these lines. You were a Forbes 30 Under 30. You’re fine with computer emotions, and you have this set of views. Do you notice any generational difference between researchers who have been in it longer than you, and people of your age and training? Do you look at it, as a whole, differently than another generation might have?
I think there are always going to be generational differences. People grow up in different times and contexts, societal norms shift… I would argue usually for the better, but not always. So I think that that context in which you were raised, that initial training data that you apply your transfer learning to for the rest of your life, has a huge effect on what you’re actually going to do, and how you perceive the world moving forward.
I spent a good amount of time today at SigOpt. Can you tell me what you’re trying to do there, and why you started or co-founded it, and what the mission is? Give me that whole story.
Yeah, definitely. SigOpt is an optimization-as-a-service company, or a software-as-a-service offering. What we do is help people configure these complex systems. So when you’re building a neural network—or maybe it’s a reinforcement learning system, or an algorithmic trading strategy—there’s often many different tunable configuration parameters.
These are the settings that you need to put in place before the system itself starts to do any sort of learning: things like the depth of the neural network, the learning rates, some of these stochastic gradient descent parameters, etc.
These are often kind of nuisance parameters that are brushed under the rug. They’re typically solved via relatively simplistic methods like brute forcing it or trying random configurations. What we do is we take an ensemble of the state-of-the-art research from academia, and Bayesian and global optimization, and we ensemble all of these algorithms behind a simple API.
So when you are downloading MxNet, or TensorFlow, or Caffe2, whatever it is, you don’t have to waste a bunch of time trying different things via trial-and-error. We can guide you to the best solution quite a bit faster.
Do you have any success stories that you like to talk about?
Yeah, definitely. One of our customers is Hotwire. They’re using us to do things like ranking systems. We work with a variety of different algorithmic trading firms to make their strategies more efficient. We also have this great academic program where SigOpt is free for any academic at any university or national lab anywhere in the world.
So we’re helping accelerate the flywheel of science by allowing people to spend less time doing trial-and-error. I wasted way too much of my PhD on this, to be completely honest—fine-tuning different configuration settings and bioinformatics algorithms.
So our goal is… If we can have humans do what they’re really good at, which is creativity—understanding the context in the domain of a problem—and then we can make the trial-and-error component as little as possible, hopefully, everything happens a little bit faster and a little bit better and more efficiently.
What are the big challenges you’re facing?
Where this system makes the biggest difference is in large complex systems, where it’s very difficult to manually tune, or brute force this problem. Humans tend to be pretty bad at doing 20-dimensional optimization in their head. But a surprising number of people still take that approach, because they’re unable to access some of this incredible research that’s been going on in academia for the last several decades.
Our goal is to make that as easy as possible. One of our challenges is finding people with these interesting complex problems. I think the recent surge of interest in deep learning and reinforcement learning, and the complexity that’s being imbued in a lot of these systems, is extremely good for us, and we’re able to ride that wave and help these people realize the potential of these systems quite a bit faster than they would otherwise.
But having the market come to us is something that we’re really excited about, but it’s not instant.
Do you find that people come to you and say, “Hey, we have this dataset, and we think somewhere in here we can figure out whatever”? Or do they just say, “We have this data, what can we do with it?” Or do they come to you and say, “We’ve heard about this AI thing, and want to know what we can do”?
There are companies that help solve that particular problem, where they’re given raw data and they help you build a model and apply it to some business context. Where SigOpt sits, which is slightly different than that, is when people come to us, they have something in place. They already have data scientists or machine learning engineers.
They’ve already applied their domain expertise to really understand their customers, the business problem they’re trying to solve, everything like that. And what they’re looking for is to get the most out of these systems that they’ve built. Or they want to build a more advanced system as rapidly as possible.
And so SigOpt bolts on top of these pre-existing systems, and gives them that boost by fine-tuning all of these different configuration parameters to get to their maximal performance. So, sometimes we do meet people like that, and we pass them on to some of our great partners. When someone has a problem and they just want to get the most out of it, that’s where we can come in and provide this black box optimization on top of it.
Final question-and-a-half. Do you speak a lot? Do you tweet? If people want to follow you and keep up with what you’re doing, what’s the best way to do that?
They can follow @SigOpt on Twitter. We have a blog where we post technical and high-level blog posts about optimization and some of the different advancements, and deep learning and reinforcement learning. We publish papers, but blog.sigopt.com and on Twitter @SigOpt is the best way to follow us along.
Alright. It has been an incredibly fascinating hour, and I want to thank you for taking the time.
Excellent. Thank you for having me. I’m really honored to be on the show.
Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here
[voices_in_ai_link_back]

Voices in AI – Episode 13: A Conversation with Bryan Catanzaro

[voices_in_ai_byline]
In this episode, Byron and Bryan talk about sentience, transfer learning, speech recognition, autonomous vehicles, and economic growth.
[podcast_player name=”Episode 13: A Conversation with Bryan Catanzaro” artist=”Byron Reese” album=”Voices in AI” url=”https://voicesinai.s3.amazonaws.com/2017-10-16-(00-54-18)-bryan-catanzaro.mp3″ cover_art_url=”https://voicesinai.com/wp-content/uploads/2017/10/voices-headshot-card-5.jpg”]
[voices_in_ai_link_back]
Byron Reese: This is “Voices in AI” brought to you by Gigaom. I’m Byron Reese. Today, our guest is Bryan Catanzaro. He is the head of Applied AI Research at NVIDIA. He has a BS in computer science and Russian from BYU, an MS in electrical engineering from BYU, and a PhD in both electrical engineering and computer science from UC Berkeley. Welcome to the show, Bryan.
Bryan Catanzaro: Thanks. It’s great to be here.
Let’s start off with my favorite opening question. What is artificial intelligence?
It’s such a great question. I like to think about artificial intelligence as making tools that can perform intellectual work. Hopefully, those are useful tools that can help people be more productive in the things that they need to do. There’s a lot of different ways of thinking about artificial intelligence, and maybe the way that I’m talking about it is a little bit more narrow, but I think it’s also a little bit more connected with why artificial intelligence is changing so many companies and so many things about the way that we do things in the world economy today is because it actually is a practical thing that helps people be more productive in their work. We’ve been able to create industrialized societies with a lot of mechanization that help people do physical work. Artificial intelligence is making tools that help people do intellectual work.
I ask you what artificial intelligence is, and you said it’s doing intellectual work. That’s sort of using the word to define it, isn’t it? What is that? What is intelligence?
Yeah, wow…I’m not a philosopher, so I actually don’t have like a…
Let me try a different tact. Is it artificial in the sense that it isn’t really intelligent and it’s just pretending to be, or is it really smart? Is it actually intelligent and we just call it artificial because we built it?
I really liked this idea from Yuval Harari that I read a while back where he said there’s the difference between intelligence and sentience, where intelligence is more about the capacity to do things and sentience is more about being self-aware and being able to reason in the way that human beings reason. My belief is that we’re building increasingly intelligent systems that can perform what I would call intellectual work. Things about understanding data, understanding the world around us that we can measure with sensors like video cameras or audio or that we can write down in text, or record in some form. The process of interpreting that data and making decisions about what it means, that’s intellectual work, and that’s something that we can create machines to be more and more intelligent at. I think the definitions of artificial intelligence that move more towards consciousness and sentience, I think we’re a lot farther away from that as a community. There are definitely people that are super excited about making generally intelligent machines, but I think that’s farther away and I don’t know how to define what general intelligence is well enough to start working on that problem myself. My work focuses mostly on practical things—helping computers understand data and make decisions about it.
Fair enough. I’ll only ask you one more question along those lines. I guess even down in narrow AI, though, if I had a sprinkler that comes on when my grass gets dry, it’s responding to its environment. Is that an AI?
I’d say it’s a very small form of AI. You could have a very smart sprinkler that was better than any person at figuring out when the grass needed to be watered. It could take into account all sorts of sensor data. It could take into account historical information. It might actually be more intelligent at figuring out how to irrigate than a human would be. And that’s a very narrow form of intelligence, but it’s a useful one. So yeah, I do think that could be considered a form of intelligence. Now it’s not philosophizing about the nature of irrigation and its harm on the planet or the history of human interventions on the world, or anything like that. So it’s very narrow, but it’s useful, and it is intelligent in its own way.
Fair enough. I do want to talk about AGI in a little while. I have some questions around…We’ll come to that in just a moment. Just in the narrow AI world, just in your world of using data and computers to solve problems, if somebody said, “Bryan, what is the state-of-the-art? Where are we at in AI? Is this the beginning and you ‘ain’t seen nothing yet’? Or are we really doing a lot of cool things, and we are well underway to mastering that world?”
I think we’re just at the beginning. We’ve seen so much progress over the past few years. It’s been really quite astonishing, the kind of progress we’ve seen in many different domains. It all started out with image recognition and speech recognition, but it’s gone a long way from there. A lot of the products that we interact with on a daily basis over the internet are using AI, and they are providing value to us. They provide our social media feeds, they provide recommendations and maps, they provide conversational interfaces like Siri or Android Assistant. All of those things are powered by AI and they are definitely providing value, but we’re still just at the beginning. There are so many things we don’t know yet how to do and so many underexplored problems to look at. So I believe we’ll continue to see applications of AI come up in new places for quite a while to come.
If I took a little statuette of a falcon, let’s say it’s a foot tall, and I showed it to you, and then I showed you some photographs, and said, “Spot the falcon.” And half the time it’s sticking halfway behind a tree, half the time it’s underwater; one time it’s got peanut butter smeared on it. A person can do that really well, but computers are far away from that. Is that an example of us being really good at transfer learning? We’re used to knowing what things with peanut butter on them look like? What is it that people are doing that computers are having a hard time to do there?
I believe that people have evolved, over a very long period of time, to operate on planet Earth with the sensors that we have. So we have a lot of built-in knowledge that tells us how to process the sensors that we have and models the world. A lot of it is instinctual, and some of it is learned. I have young children, like a year-old or so. They spend an awful lot of time just repetitively probing the world to see how it’s going to react when they do things, like pushing on a string, or a ball, and they do it over and over again because I think they’re trying to build up their models about the world. We have actually very sophisticated models of the world that maybe we take for granted sometimes because everyone seems to get them so easily. It’s not something that you have to learn in school. But these models are actually quite useful, and they’re more sophisticated than – and more general than – the models that we currently can build with today’s AI technology.
To your question about transfer learning, I feel like we’re really good at transfer learning within the domain of things that our eyes can see on planet Earth. There are probably a lot of situations where an AI would be better at transfer learning. Might actually have fewer assumptions baked in about how the world is structured, how objects look, what kind of composition of objects is actually permissible. I guess I’m just trying to say we shouldn’t forget that we come with a lot of context. That’s instinctual, and we use that, and it’s very sophisticated.
Do you take from that that we ought to learn how to embody an AI and just let it wander around the world, bumping into things and poking at them and all of that? Is that what you’re saying? How do we overcome that?
It’s an interesting question you note. I’m not personally working on trying to build artificial general intelligence, but it will be interesting for those people that are working on it to see what kind of childhood is necessary for an AI. I do think that childhood is a really important part of developing human intelligence, and plays a really important part of developing human intelligence because it helps us build and calibrate these models of how the world works, which then we apply to all sorts of things like your question of the falcon statue. Will computers need things like that? It’s possible. We’ll have to see. I think one of the things that’s different about computers is that they’re a lot better at transmitting information identically, so it may be the kind of thing that we can train once, and then just use repeatedly – as opposed to people, where the process of replicating a person is time-consuming and not exact.
But that transfer learning problem isn’t really an AGI problem at all, though. Right? We’ve taught a computer to recognize a cat, by giving it a gazillion images of a cat. But if we want to teach it how to recognize a bird, we have to start over, don’t we?
I don’t think we generally start over. I think most of the time if people wanted to create a new classifier, they would use transfer learning from an existing classifier that had been trained on a wide variety of different object types. It’s actually not very hard to do that, and people do that successfully all the time. So at least for image recognition, I think transfer learning works pretty well. For other kinds of domains, they can be a little bit more challenging. But at least for image recognition, we’ve been able to find a set of higher-level features that are very useful in discriminating between all sorts of different kinds of objects, even objects that we haven’t seen before.
What about audio? Because I’m talking to you now and I’m snapping my fingers. You don’t have any trouble continuing to hear me, but a computer trips over that. What do you think is going on in people’s minds? Why are we good at that, do you think? To get back to your point about we live on Earth, it’s one of those Earth things we do. But as a general rule, how do we teach that to a computer? Is that the same as teaching it to see something, as to teach it to hear something?
I think it’s similar. The best speech recognition accuracies come from systems that have been trained on huge amounts of data, and there does seem to be a relationship that the more data we can train a model on, the better the accuracy gets. We haven’t seen the end of that yet. I’m pretty excited about the prospects of being able to teach computers to continually understand audio, better and better. However, I wanted to point out, humans, this is kind of our superpower: conversation and communication. You watch birds flying in a flock, and the birds can all change direction instantaneously, and the whole flock just moves, and you’re like, “How do you do that and not run into each other?” They have a lot of built-in machinery that allows them to flock together. Humans have a lot of built-in machinery for conversation and for understanding spoken language. The pathways for speaking and the pathways for hearing evolve together, so they’re really well-matched.
With computers trying to understand audio, we haven’t gotten to that point yet. I remember some of the experiments that I’ve done in the past with speech recognition, that the recognition performance was very sensitive to compression artifacts that were actually not audible to humans. We could actually take a recording, like this one, and recompress it in a way that sounded identical to a person, and observe a measurable difference in the recognition accuracy of our model. That was a little disconcerting because we’re trying to train the model to be invariant to all the things that humans are invariant to, but it’s actually quite hard to do that. We certainly haven’t achieved that yet. Often, our models are still what we would call “overfitting”, where they’re paying attention to a lot of details that help it perform the tasks that we’re asking it to perform, but they’re not actually helpful to solving the fundamental tasks that we’re trying to perform. And we’re continually trying to improve our understanding of the tasks that we’re solving so that we can avoid this, but we’ve still got more work to do.
My standard question when I’m put in front of a chatbot or one of the devices that sits on everybody’s desktop, I can’t say them out loud because they’ll start talking to me right now, but the question I always ask is “What is bigger, a nickel or the sun?” To date, nothing has ever been able to answer that question. It doesn’t know how sun is spelled. “Whose son? The sun? Nickel? That’s actually a coin.” All of that. What all do we have to get good at, for the computer to answer that question? Run me down the litany of all the things we can’t do, or that we’re not doing well yet, because there’s no system I’ve ever tried that answered that correctly.
I think one of the things is that we’re typically not building chat systems to answer trivia questions just like that. I think if we were building a special-purpose trivia system for questions like that, we probably could answer it. IBM Watson did pretty well on Jeopardy, because it was trained to answer questions like that. I think we definitely have the databases, the knowledge bases, to answer questions like that. The problem is that kind of a question is really outside of the domain of most of the personal assistants that are being built as products today because honestly, trivia bots are fun, but they’re not as useful as a thing that can set a timer, or check the weather, or play a song. So those are mostly the things that those systems are focused on.
Fair enough, but I would differ. You can go to Wolfram Alpha and say, “What’s bigger, the Statue of Liberty or the Empire State Building?” and it’ll answer that. And you can ask Amazon’s product that same question, and it’ll answer it. Is that because those are legit questions and my question is not legit, or is it because we haven’t taught systems to disintermediate very well and so they don’t really know what I mean when I say “sun”?
I think that’s probably the issue. There’s a language modeling problem when you say, “What’s bigger, a nickel or the sun?” The sun can mean so many different things, like you were saying. Nickel, actually, can be spelled a couple of different ways and has a couple of different meanings. Dealing with ambiguities like that is a little bit hard. I think when you ask that question to me, I categorize this as a trivia question, and so I’m able to disambiguate all of those things, and look up the answer in my little knowledge base in my head, and answer your question. But I actually don’t think that particular question is impossible to solve. I just think it’s just not been a focus to try to solve stuff like that, and that’s why they’re not good.
AIs have done a really good job playing games: Deep Blue, Watson, AlphaGo, and all of that. I guess those are constrained environments with a fixed set of rules, and it’s easy to understand who wins, and what a point is, and all that. What is going to be the next thing, that’s a watershed event, that happens? Now they can outbluff people in poker. What’s something that’s going to be, in a year, or two years, five years down the road, that one day, it wasn’t like that in the universe, and the next day it was? And the next day, the best Go player in the world was a machine.
The thing that’s on my mind for that right now is autonomous vehicles. I think it’s going to change the world forever to unchain people from the driver’s seat. It’s going to give people hugely increased mobility. I have relatives that their doctors have asked them to stop driving cars because it’s no longer safe for them to be doing that, and it restricts their ability to get around the world, and that frustrates them. It’s going to change the way that we all live. It’s going to change the real estate markets, because we won’t have to park our cars in the same places that we’re going to. It’s going to change some things about the economy, because there’s going to be new delivery mechanisms that will become economically viable. I think intelligence that can help robots essentially drive around the roads, that’s the next thing that I’m most excited about, that I think is really going to change everything.
We’ll come to that in just a minute, but I’m actually asking…We have self-driving cars, and on an evolutionary basis, they’ll get a little better and a little better. You’ll see them more and more, and then someday there’ll be even more of them, and then they’ll be this and this and this. It’s not that surprise moment, though, of AlphaGo just beat Lee Sedol at Go. I’m wondering if there is something else like that—that it’s this binary milestone that we can all keep our eye open for?
I don’t know. As far as we have self-driving cars already, I don’t have a self-driving car that could say, for example, let me sit in it at nighttime, go to sleep and wake up, and it brought me to Disneyland. I would like that kind of self-driving car, but that car doesn’t exist yet. I think self-driving trucks that can go cross country carrying stuff, that’s going to radically change the way that we distribute things. I do think that we have, as you said, we’re on the evolutionary path to self-driving cars, but there’s going to be some discrete moments when people actually start using them to do new things that will feel pretty significant.
As far as games and stuff, and computers being better at games than people, it’s funny because I feel like Silicon Valley has, sometimes, a very linear idea of intelligence. That one person is smarter than another person maybe because of an SAT score, or an IQ test, or something. They use that sort of linearity of an intelligence to where some people feel threatened by artificial intelligence because they extrapolate that artificial intelligence is getting smarter and smarter along this linear scale, and that’s going to lead to all sorts of surprising things, like Lee Sedol losing to Go, but on a much bigger scale for all of us. I feel kind of the opposite. Intelligence is such a multidimensional thing. The fact that a computer is better at Go then I am doesn’t really change my life very much, because I’m not very good at Go. I don’t play Go. I don’t consider Go to be an important part of my intelligence. Same with chess. When Gary Kasparov lost to Deep Blue, that didn’t threaten my intelligence. I am sort of defining the way that I work and how I add value to the world, and what things make me happy on a lot of other axes besides “Can I play chess?” or “Can I play Go?” I think that speaks to the idea that intelligence really is very multifaceted. There’s a lot of different kinds – there’s probably thousands or millions of different kinds of intelligence – and it’s not very linearizable.
Because of that, I feel like, as we watch artificial intelligence develop, we’re going to see increasingly more intelligent machines, but they’re going to be increasingly more intelligent in some very narrow domains like “this is the better Go-playing robot than me”, or “this is the better car driver than me”. That’s going to be incredibly useful, but it’s not going to change the way that I think about myself, or about my work, or about what makes me happy. Because I feel like there are so many more dimensions of intelligence that are going to remain the province of humans. That’s going to take a very long time, if ever, for artificial intelligence to become better at all of them than us. Because, as I said, I don’t believe that intelligence is a linearizable thing.
And you said you weren’t a philosopher. I guess the thing that’s interesting to people, is there was a time when information couldn’t travel faster than a horse. And then the train came along, and information could travel. That’s why in the old Westerns – if they ever made it on the train, that was it, and they were out of range. Nothing traveled faster than the train. Then we had a telegraph and, all of a sudden, that was this amazing thing that information could travel at the speed of light. And then one time they ran these cables under the ocean, and somebody in England could talk to somebody in the United States instantly. Each one of them, and I think it’s just an opportunity to pause, and reflect, and to mark a milestone, and to think about what it all means. I think that’s why a computer just beat these awesome poker players. It learned to bluff. You just kind of want to think about it.
So let’s talk about jobs for a moment because you’ve been talking around that for just a second. Just to set the question up: Generally speaking, there are three views of what automation and artificial intelligence are going to do to jobs. One of them reflects kind of what you were saying is that there are going to be a certain group of workers who are considered low skilled, and there are going to be automation that takes these low-skilled jobs, and that there’s going to be a sizable part of the population that’s locked out of the labor market, and it’s kind of like the permanent Great Depression over and over and over forever. Then there’s another view that says, “No, you don’t understand. There’s going to be an inflection point where they can do every single thing. They’re going to be a better conductor and a better painter and a better novelist and a better everything than us. Don’t think that you’ve got something that a machine can’t do.” Clearly, that isn’t your viewpoint from what you said. Then there’s a third viewpoint that says, “No, in the past, even when we had these transformative technologies like electricity and mechanization, people take those technologies and they use them to increase their own productivity and, therefore, their own incomes. And you never have unemployment go up because of them, because people just take it and make a new job with it.” Of those three, or maybe a fourth one I didn’t cover; where do you find yourself?
I feel like I’m closer in spirit to number three. I’m optimistic. I believe that the primary way that we should expect economic growth in the future is by increased productivity. If you buy a house or buy some stock and you want to sell it 20 or 30 years from now, who’s going to buy it, and with what money, and why do you expect the price to go up? I think the answer to that question should be the people in the future should have more money than us because they’re more productive, and that’s why we should expect our world economy to continue growing. Because we find more productivity. I actually feel like this is actually necessary. World productivity growth has been slowing for the past several decades, and I feel like artificial intelligence is our way out of this trap where we have been unable to figure out how to grow our economy because our productivity hasn’t been improving. I actually feel like this is a necessary thing for all of us, is to figure out how to improve productivity, and I think AI is the way that we’re going to do that for the next several decades.
The one thing that I disagreed with in your third statement was this idea that unemployment would never go up. I think nothing is ever that simple. I actually am quite concerned about job displacement in the short-term. I think there will be people that suffer and in fact, I think, to a certain extent, this is already happening. The election of Donald Trump was an eye-opener to me that there really exists a lot of people that feel that they have been left behind by the economy, and they come to very different conclusions about the world than I might. I think that it’s possible that, as we continue to digitize our society, and AI becomes a lever that some people will become very good at using to increase their productivity, that we’re going to see increased inequality and that worries me.
The primary challenges that I’m worried about, for our society, with the rise of AI, have to do more with making sure that we give people purpose and meaning in their life that maybe doesn’t necessarily revolve around punching out a timecard, and showing up to work at 8 o’clock in the morning every day. I want to believe that that future exists. There are a lot of people right now that are brilliant people that have a lot that they could be contributing in many different ways – intellectually, artistically – that are currently not given that opportunity, because they maybe grew up in a place that didn’t have the right opportunities for them to get the right education so that they could apply their skills in that way, and many of them are doing jobs that I think don’t allow them to use their full potential.
So I’m hoping that, as we automate many of those jobs, that more people will be able to find work that provides meaning and purpose to them and allows them to actually use their talents and make the world a better place, but I acknowledge that it’s not going to be an easy transition. I do think that there’s going to be a lot of implications for how our government works and how our economy works, and I hope that we can figure out a way to help defray some of the pain that will happen during this transition.
You talked about two things. You mentioned income inequality as a thing, but then you also said, “I think we’re going to have unemployment from these technologies.” Separating those for a minute and just looking at the unemployment one for a minute, you say things are never that simple. But with the exception of the Great Depression, which nobody believes was caused by technology, unemployment has been between 5% and 10% in this country for 250 years and it only moves between 5% and 10% because of the business cycle, but there aren’t counterexamples. Just imagine if your job was you had animals that performed physical labor. They pulled, and pushed, and all of that. And somebody made the steam engine. That was disruptive. But even when we had that, we had electrification of industry. We adopted steam power. We went from 5% to 85% of our power being generated by steam in just 22 years. And even when you had that kind of disruption, you still didn’t have any increases in unemployment. I’m curious, what is the mechanism, in your mind, by which this time is different?
I think that’s a good point that you raise, and I actually haven’t studied all of those other transitions that our society has gone through. I’d like to believe that it’s not different. That would be a great story if we could all come to agreement, that we won’t see increased unemployment from AI. I think the reason why I’m a little bit worried is that I think this transition in some fields will happen quickly, maybe more quickly than some of the transitions in the past did. Just because, as I was saying, AI is easier to replicate than some other technologies, like electrification of a country. It takes a lot of time to build out physical infrastructure that can actually deliver that. Whereas I think for a lot of AI applications, that infrastructure will be cheaper and quicker to build, so the velocity of the change might be faster and that could lead to a little bit more shock. But it’s an interesting point you raise, and I certainly hope that we can find a way through this transition that is less painful than I’m worried it could be.
Do you worry about misuse of AI? I’m an optimist on all of this. And I know that every time we have some new technology come along, people are always looking at the bad cases. You take something like the internet, and the internet has overwhelmingly been a force for good. It connects people in a profound way. There’s a million things. And yeah, some people abuse it. But on net, all technology, I believe, almost all technology on net is used for good because I think, on net, people, on average, are more inclined to build than to destroy. That being said, do you worry about nefarious uses of AI, specifically in warfare?
Yeah. I think that there definitely are going to be some scary killer robots that armies make. Armies love to build machinery that kills things and AI will help them do that, and that will be scary. I think it’s interesting, like, where is the real threat going to come from? Sometimes, I feel like the threat of malevolent AI being deployed against people is going to be more subtle than that. It’s going to be more about things that you can do after compromising fiber systems of some adversary, and things that you can do to manipulate them using AI. There’s been a lot of discussion about Russian involvement in the 2016 election in the US, and that wasn’t about sending evil killer robots. It was more about changing people’s opinions, or attempting to change their opinions, and AI will give entities tools to do that on a scale that maybe we haven’t seen before. I think there may be nefarious uses of AI that are more subtle and harder to see than a full-frontal assault from a movie with evil killer robots. I do worry about all of those things, but I also share your optimism. I think we humans, we make lots of mistakes and we shouldn’t give ourselves too easy of a time here. We should learn from those mistakes, but we also do a lot of things well. And we have used technologies in the past to make the world better, and I hope AI will do so as well.
Pedro Domingo wrote a book called The Master Algorithm where he says there are all of these different tools and techniques that we use in artificial intelligence. And he surmises that there is probably a grandparent algorithm, the master algorithm, that can solve any problem, any range of problems. Does that seem possible to you or likely, or do you have any thoughts on that?
I think it’s a little bit far away, at least from AI as it’s practiced today. Right now, the practical, on-the-ground experience of researchers trying to use AI to do something new is filled with a lot of pain, suffering, blood, sweat, tears, and perseverance if they are to succeed, and I see that in my lab every day. Most of the researchers – and I have brilliant researchers in my lab that are working very hard, and they’re doing amazing work. And most of the things they try fail. And they have to keep trying. I think that’s generally the case right now across all the people that are working on AI. The thing that’s different is we’ve actually started to see some big successes, along with all of those more frustrating everyday occurrences. So I do think that we’re making the progress, but I think having a master algorithm that’s pushbutton that can solve any problem you pose to it that’s something that’s hard for me to conceive of with today’s state of artificial intelligence.
AI, of course, it’s doubtful we’ll have another AI winter because, like you said, it’s kind of delivering the goods, and there have been three things that have happened that made that possible. One of them is better hardware, and obviously you’re part of that world. The second thing is better algorithms. We’ve learned to do things a lot smarter. And the third thing is we have more data, because we are able to collect it, and store it, and whatnot. Assuming you think the hardware is the biggest of the driving factors, what would you think has been the bigger advance? Is it that we have so much more data, or so much better algorithms?
I think the most important thing is more data. I think the algorithms that we’re using in AI right now are, more or less, clever variations of algorithms that have been around for decades, and used to not work. When I was a PhD student and I was studying AI, all the smart people told me, “Don’t work with deep learning, because it doesn’t work. Use this other algorithm called support vector machines.” Which, at the time, that was the hope that that was going to be the master algorithm. So I stayed away from deep learning back then because, at the time, it didn’t work. I think now we have so much more data, and deep learning models have been so successful at taking advantage of that data, that we’ve been able to make a lot of progress. I wouldn’t characterize deep learning as a master algorithm, though, because deep learning is like a fuzzy cloud of things that have some relationships to each other, but actually finding a space inside that fuzzy cloud to solve a particular problem requires a lot of human ingenuity.
Is there a phrase – it’s such a jargon-loaded industry now – are there any of the words that you just find rub you the wrong way? Because they don’t mean anything and people use them as if they do? Do you have anything like that?
Everybody has pet peeves. I would say that my biggest pet peeve right now is the word neuromorphic. I have almost an allergic reaction every time I hear that word, mostly because I don’t think we know what neurons are or what they do, and I think modeling neurons in a way that actually could lead to brain simulations that actually worked is a very long project that we’re decades away from solving. I could be wrong on that. I’m always waiting for somebody to prove me wrong. Strong opinions, weakly held. But so far, neuromorphic is a word that I just have an allergic reaction to, every time.
Tell me about what you do. You are the head of Applied AI Research at NVIDIA, so what does your day look like? What does your team work on? What’s your biggest challenge right now, and all of that?
NVIDIA sells GPUs which have powered most of the deep learning revolution, so pretty much all of the work that’s going on with deep learning across the entire world right now, runs on NVIDIA GPUs. And that’s been very exciting for NVIDIA, and exciting for me to be involved in building that. The next step, I think, for NVIDIA is to figure out how to use AI to change the way that it does its own work. NVIDIA is incentivized to do this because we see the value that AI is bringing to our customers. Our GPU sales have been going up quite a bit because we’re providing a lot of value to everyone else who’s trying to use AI for their own problems. So the next step is to figure out how to use AI for NVIDIA’s problems directly. Andrew Ng, who I used to work with, has this great quote that “AI is the new electricity,” and I believe that. I think that we’re going to see AI applied in many different ways to many different kinds of problems, and my job at NVIDIA is to figure out how to do that here. So that’s what my team focuses on.
We have projects going on in quite a few different domains, ranging from graphics to audio, and text, and others. We’re trying to change the way that everything at NVIDIA happens: from chip design, to video games, and everything in between. As far as my day-to-day work goes, I lead this team, so that means I spend a lot of time talking with people on the team about the work that they’re doing, and trying to make sure they have the right resources, data, the right hardware, the right ideas, the right connections, so that they can make progress on problems that they’re trying to solve. Then when we have prototypes that we’ve built showing how to apply AI to a particular problem, then I work with people around the company to show them the promise of AI applied to problems that they care about.
I think one of the things that’s really exciting to me about this mission is that we’re really trying to change NVIDIA’s work at the core of the company. So rather than working on applied AI, that could maybe help some peripheral part of the company that maybe could be nice if we did that, we’re actually trying to solve very fundamental problems that the company faces with AI, and hopefully we’ll be able to change the way that the company does business, and transform NVIDIA into an AI company, and not just a company that makes hardware for AI.
You are the head of the Applied AI Research. Is there a Pure AI Research group, as well?
Yes, there is.
So everything you do, you have an internal customer for already?
That’s the idea. To me, the difference between fundamental research and applied research is more a question of emphasis on what’s the fundamental goal of your work. If the goal is academic novelty, that would be fundamental research. Our goal is, we think about applications all the time, and we don’t work on problems unless we have a clear application that we’re trying to build that could use a solution.
In most cases, do other groups come to you and say, “We have this problem we really want to solve. Can you help us?” Or is the science nascent enough that you go and say, “Did you know that we can actually solve this problem for you?”
It kind of works all of those ways. We have a list of projects that people around the company have proposed to us, and we also have a list of projects that we ourselves think are interesting to look at. There’s also a few projects that my management tells me, “I really want you to look at this problem. I think it’s really important.” We get input from all directions, and then prioritize, and go after the ones we think are most feasible, and most important.
And do you find a talent shortage? You’re NVIDIA on the one hand, but on the other hand, you know: it’s AI.
I think the entire field, no matter what company you work at, the entire field has a shortage of qualified scientists that can do AI research, and that’s despite the fact that the amount of people jumping into AI is increasing every year. If you go to any of the academic AI conferences, you’ll see how much energy and how much excitement, and how many people that are there that didn’t used to be there. That’s really wonderful to see. But even with all of that growth and change, it is a big problem for the industry. So, to all of your listeners that are trying to figure out what to do next, come work on AI. We have lots of fun problems to work on, and not nearly enough people doing it.
I know a lot of your projects I’m sure you can’t talk about, but tell me something you have done, that you can talk about, and what the goal was, and what you were able to achieve. Give us a success story.
I’ll give you one that’s relevant to the last question that you asked, which is about how to find talent for AI. We’ve actually built a system that can match candidates to job openings at NVIDIA. Basically, it can predict how well we think a particular candidate is a fit for a particular job. That system is actually performing pretty well. So we’re trialing it with hiring managers around the company to figure out if it can help them be more efficient in their work as they search for people to come join NVIDIA.
That looks like a game, isn’t it? I assume you have a pool of resumes or LinkedIn profiles or whatever, and then you have a pool of successful employees, and you have a pool of job descriptions and you’re trying to say, “How can I pull from that big pool, based on these job descriptions, and actually pick the people that did well in the end?”
That’s right.
That’s like a game, right? You have points.
That’s right.
Would you ever productize anything, or is everything that you’re doing just for your own use?
We focus primarily on building prototypes, not products, in my team. I think that’s what the research is about. Once we build a prototype that shows promise for a particular problem, then we work with other people in the company to get that actually deployed, and they would be the people that think about business strategy about whether something should be productized, or not.
But you, in theory, might turn “NVIDIA Resume Pro” into something people could use?
Possibly. NVIDIA also works with a lot of other companies. As we enable companies in many different parts of the economy to apply AI to their problems, we work with them to help them do that. So it might make more sense for us, for example, to deliver this prototype to some of our partners that are in a position to deliver products like this more directly, and then they can figure out how to enlarge its capabilities, and make it more general to try to solve bigger problems that address their whole market and not just one company’s needs. Partnering with other companies is good for NVIDIA because it helps us grow AI which is something we want to do because, as AI grows, we grow. Personally, I think some of the things that we’re working on; it just doesn’t really make sense. It’s not really in NVIDIA’s DNA to productize them directly because it’s just not the business model that the company has.
I’m sure you’re familiar with the “right to know” legislation in Europe: the idea that if an AI makes a decision about you, you have a right to know why it made that decision. AI researchers are like, “It’s not necessarily that easy to do that.” So in your case, your AI would actually be subject to that. It would say, “Why did you pick that person over this person for that job?” Is that an answerable question?
First of all, I don’t think that this system – or I can’t imagine – using it to actually make hiring decisions. I think that would be irresponsible. This system makes mistakes. What we’re trying to do is improve productivity. If instead of having to sort through 200 resumes to find 3 that I want to talk to—if I can look at 10 instead—then that’s a pretty good improvement in my productivity, but I’m still going to be involved, as a hiring manager, to figure out who is the right fit for my jobs.
But an AI excluded 190 people from that position.
It didn’t exclude them. It sorted them, and then the person decided how to allocate their time in a search.
Let’s look at the problem more abstractly. What do you think, just in general, about the idea that every decision an AI makes, should be, and can be, explained?
I think it’s a little bit utopian. Certainly, I don’t have the ability to explain all of the decisions that I make, and people, generally, are not very good at explaining their decisions, which is why there are significant legal battles going on about factual things, that people see in different ways, and remember in different ways. So asking a person to explain their intent is actually a very complicated thing, and we’re not actually very good at it. So I don’t actually think that we’re going to be able to enforce that AI is able to explain all of its decisions in a way that makes sense to humans. I do think that there are things that we can do to make the results of these systems more interpretable. For example, on the resume job description matching system that I mentioned earlier, we’ve built a prototype that can highlight parts of the resume that were most interesting to the model, both in a positive, and in a negative sense. That’s a baby step towards interpretability so that if you were to pull up that job description and a particular person and you could see how they matched, that might explain to you what the model was paying attention to as it made a ranking.
It’s funny because when you hear reasons why people exclude a resume, I remember one person said, “I’m not going to hire him. He has the same first name as somebody else on the team. That’d just be too confusing.” And somebody else I remember said that the applicant was a vegan and the place they like to order pizza from didn’t have a vegan alternative that the team liked to order from. Those are anecdotal of course, but people use all kinds of other things when they’re thinking about it.
Yeah. That’s actually one of the reasons why I’m excited about this particular system is that I feel like we should be able to construct it in a way that actually has fewer biases than people do, because we know that people harbor all sorts of biases. We have employment laws that guide us to stay away from making decisions based on protected classes. I don’t know if veganism is a protected class, but it’s verging on that. If you’re making hiring decisions based on people’s personal lifestyle choices, that’s suspect. You could get in trouble for that. Our models, we should be able to train them to be more dispassionate than any human could be.
We’re running out of time. Let’s close up by: do you consume science fiction? Do you ever watch movies or read books or any of that? And if so, is there any of it that you look at, especially any that portrays artificial intelligence, like Ex Machina, or Her, or Westworld or any of that stuff, that you look at and you’re like, “Wow, that’s really interesting,” or “That could happen,” or “That’s fascinating,” or anything like that?
I do consume science fiction. I love science fiction. I don’t actually feel like current science fiction matches my understanding of AI very well. Ex Machina, for example, that was a fun movie. I enjoyed watching that movie, but I felt, from a scientific point of view, it just wasn’t very interesting. I was talking about our built-in models of the world. One of the things that humans, over thousands of years, have drilled into our heads is that there’s somebody out to get you. We have a large part of our brain that’s worrying all the time, like, “Who’s going to come kill me tonight? Who’s going to take away my job? Who’s going to take my food? Who’s going to burn down my house?” There’s all these things that we worry about. So a lot of the depictions of AI in science fiction inflame that part of the brain that is worrying about the future, rather than actually speak to the technology and its potential.
I think probably the part of science fiction that has had the most impact on my thoughts about AI is Isaac Asimov’s Three Laws. Those, I think, are pretty classic, and I hope that some of them can be adapted to the kinds of problems that we’re trying to solve with AI, to make AI safe, and make it possible for people to feel confident that they’re interacting with AI, and not worry about it. But I feel like most of science fiction is, especially movies – maybe books can be a little bit more intellectual and maybe a little bit more interesting – but especially movies, it just sells more movies to make people afraid, than it does to show people a mundane existence where AI is helping people live better lives. It’s just not nearly as compelling of a movie, so I don’t actually feel like popular culture treatment of AI is very realistic.
All right. Well, on that note, I say, we wrap up. I want to thank you for a great hour. We covered a lot of ground, and I appreciate you traveling all that way with me.
It was fun.
Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here
Byron Reese: This is “Voices in AI” brought to you by Gigaom. I’m Byron Reese. Today, our guest is Bryan Catanzaro. He is the head of Applied AI Research at NVIDIA. He has a BS in computer science and Russian from BYU, an MS in electrical engineering from BYU, and a PhD in both electrical engineering and computer science from UC Berkeley. Welcome to the show, Bryan.
Bryan Catanzaro: Thanks. It’s great to be here.
Let’s start off with my favorite opening question. What is artificial intelligence?
It’s such a great question. I like to think about artificial intelligence as making tools that can perform intellectual work. Hopefully, those are useful tools that can help people be more productive in the things that they need to do. There’s a lot of different ways of thinking about artificial intelligence, and maybe the way that I’m talking about it is a little bit more narrow, but I think it’s also a little bit more connected with why artificial intelligence is changing so many companies and so many things about the way that we do things in the world economy today is because it actually is a practical thing that helps people be more productive in their work. We’ve been able to create industrialized societies with a lot of mechanization that help people do physical work. Artificial intelligence is making tools that help people do intellectual work.
I ask you what artificial intelligence is, and you said it’s doing intellectual work. That’s sort of using the word to define it, isn’t it? What is that? What is intelligence?
Yeah, wow…I’m not a philosopher, so I actually don’t have like a…
Let me try a different tact. Is it artificial in the sense that it isn’t really intelligent and it’s just pretending to be, or is it really smart? Is it actually intelligent and we just call it artificial because we built it?
I really liked this idea from Yuval Harari that I read a while back where he said there’s the difference between intelligence and sentience, where intelligence is more about the capacity to do things and sentience is more about being self-aware and being able to reason in the way that human beings reason. My belief is that we’re building increasingly intelligent systems that can perform what I would call intellectual work. Things about understanding data, understanding the world around us that we can measure with sensors like video cameras or audio or that we can write down in text, or record in some form. The process of interpreting that data and making decisions about what it means, that’s intellectual work, and that’s something that we can create machines to be more and more intelligent at. I think the definitions of artificial intelligence that move more towards consciousness and sentience, I think we’re a lot farther away from that as a community. There are definitely people that are super excited about making generally intelligent machines, but I think that’s farther away and I don’t know how to define what general intelligence is well enough to start working on that problem myself. My work focuses mostly on practical things—helping computers understand data and make decisions about it.
Fair enough. I’ll only ask you one more question along those lines. I guess even down in narrow AI, though, if I had a sprinkler that comes on when my grass gets dry, it’s responding to its environment. Is that an AI?
I’d say it’s a very small form of AI. You could have a very smart sprinkler that was better than any person at figuring out when the grass needed to be watered. It could take into account all sorts of sensor data. It could take into account historical information. It might actually be more intelligent at figuring out how to irrigate than a human would be. And that’s a very narrow form of intelligence, but it’s a useful one. So yeah, I do think that could be considered a form of intelligence. Now it’s not philosophizing about the nature of irrigation and its harm on the planet or the history of human interventions on the world, or anything like that. So it’s very narrow, but it’s useful, and it is intelligent in its own way.
Fair enough. I do want to talk about AGI in a little while. I have some questions around…We’ll come to that in just a moment. Just in the narrow AI world, just in your world of using data and computers to solve problems, if somebody said, “Bryan, what is the state-of-the-art? Where are we at in AI? Is this the beginning and you ‘ain’t seen nothing yet’? Or are we really doing a lot of cool things, and we are well underway to mastering that world?”
I think we’re just at the beginning. We’ve seen so much progress over the past few years. It’s been really quite astonishing, the kind of progress we’ve seen in many different domains. It all started out with image recognition and speech recognition, but it’s gone a long way from there. A lot of the products that we interact with on a daily basis over the internet are using AI, and they are providing value to us. They provide our social media feeds, they provide recommendations and maps, they provide conversational interfaces like Siri or Android Assistant. All of those things are powered by AI and they are definitely providing value, but we’re still just at the beginning. There are so many things we don’t know yet how to do and so many underexplored problems to look at. So I believe we’ll continue to see applications of AI come up in new places for quite a while to come.
If I took a little statuette of a falcon, let’s say it’s a foot tall, and I showed it to you, and then I showed you some photographs, and said, “Spot the falcon.” And half the time it’s sticking halfway behind a tree, half the time it’s underwater; one time it’s got peanut butter smeared on it. A person can do that really well, but computers are far away from that. Is that an example of us being really good at transfer learning? We’re used to knowing what things with peanut butter on them look like? What is it that people are doing that computers are having a hard time to do there?
I believe that people have evolved, over a very long period of time, to operate on planet Earth with the sensors that we have. So we have a lot of built-in knowledge that tells us how to process the sensors that we have and models the world. A lot of it is instinctual, and some of it is learned. I have young children, like a year-old or so. They spend an awful lot of time just repetitively probing the world to see how it’s going to react when they do things, like pushing on a string, or a ball, and they do it over and over again because I think they’re trying to build up their models about the world. We have actually very sophisticated models of the world that maybe we take for granted sometimes because everyone seems to get them so easily. It’s not something that you have to learn in school. But these models are actually quite useful, and they’re more sophisticated than – and more general than – the models that we currently can build with today’s AI technology.
To your question about transfer learning, I feel like we’re really good at transfer learning within the domain of things that our eyes can see on planet Earth. There are probably a lot of situations where an AI would be better at transfer learning. Might actually have fewer assumptions baked in about how the world is structured, how objects look, what kind of composition of objects is actually permissible. I guess I’m just trying to say we shouldn’t forget that we come with a lot of context. That’s instinctual, and we use that, and it’s very sophisticated.
Do you take from that that we ought to learn how to embody an AI and just let it wander around the world, bumping into things and poking at them and all of that? Is that what you’re saying? How do we overcome that?
It’s an interesting question you note. I’m not personally working on trying to build artificial general intelligence, but it will be interesting for those people that are working on it to see what kind of childhood is necessary for an AI. I do think that childhood is a really important part of developing human intelligence, and plays a really important part of developing human intelligence because it helps us build and calibrate these models of how the world works, which then we apply to all sorts of things like your question of the falcon statue. Will computers need things like that? It’s possible. We’ll have to see. I think one of the things that’s different about computers is that they’re a lot better at transmitting information identically, so it may be the kind of thing that we can train once, and then just use repeatedly – as opposed to people, where the process of replicating a person is time-consuming and not exact.
But that transfer learning problem isn’t really an AGI problem at all, though. Right? We’ve taught a computer to recognize a cat, by giving it a gazillion images of a cat. But if we want to teach it how to recognize a bird, we have to start over, don’t we?
I don’t think we generally start over. I think most of the time if people wanted to create a new classifier, they would use transfer learning from an existing classifier that had been trained on a wide variety of different object types. It’s actually not very hard to do that, and people do that successfully all the time. So at least for image recognition, I think transfer learning works pretty well. For other kinds of domains, they can be a little bit more challenging. But at least for image recognition, we’ve been able to find a set of higher-level features that are very useful in discriminating between all sorts of different kinds of objects, even objects that we haven’t seen before.
What about audio? Because I’m talking to you now and I’m snapping my fingers. You don’t have any trouble continuing to hear me, but a computer trips over that. What do you think is going on in people’s minds? Why are we good at that, do you think? To get back to your point about we live on Earth, it’s one of those Earth things we do. But as a general rule, how do we teach that to a computer? Is that the same as teaching it to see something, as to teach it to hear something?
I think it’s similar. The best speech recognition accuracies come from systems that have been trained on huge amounts of data, and there does seem to be a relationship that the more data we can train a model on, the better the accuracy gets. We haven’t seen the end of that yet. I’m pretty excited about the prospects of being able to teach computers to continually understand audio, better and better. However, I wanted to point out, humans, this is kind of our superpower: conversation and communication. You watch birds flying in a flock, and the birds can all change direction instantaneously, and the whole flock just moves, and you’re like, “How do you do that and not run into each other?” They have a lot of built-in machinery that allows them to flock together. Humans have a lot of built-in machinery for conversation and for understanding spoken language. The pathways for speaking and the pathways for hearing evolve together, so they’re really well-matched.
With computers trying to understand audio, we haven’t gotten to that point yet. I remember some of the experiments that I’ve done in the past with speech recognition, that the recognition performance was very sensitive to compression artifacts that were actually not audible to humans. We could actually take a recording, like this one, and recompress it in a way that sounded identical to a person, and observe a measurable difference in the recognition accuracy of our model. That was a little disconcerting because we’re trying to train the model to be invariant to all the things that humans are invariant to, but it’s actually quite hard to do that. We certainly haven’t achieved that yet. Often, our models are still what we would call “overfitting”, where they’re paying attention to a lot of details that help it perform the tasks that we’re asking it to perform, but they’re not actually helpful to solving the fundamental tasks that we’re trying to perform. And we’re continually trying to improve our understanding of the tasks that we’re solving so that we can avoid this, but we’ve still got more work to do.
My standard question when I’m put in front of a chatbot or one of the devices that sits on everybody’s desktop, I can’t say them out loud because they’ll start talking to me right now, but the question I always ask is “What is bigger, a nickel or the sun?” To date, nothing has ever been able to answer that question. It doesn’t know how sun is spelled. “Whose son? The sun? Nickel? That’s actually a coin.” All of that. What all do we have to get good at, for the computer to answer that question? Run me down the litany of all the things we can’t do, or that we’re not doing well yet, because there’s no system I’ve ever tried that answered that correctly.
I think one of the things is that we’re typically not building chat systems to answer trivia questions just like that. I think if we were building a special-purpose trivia system for questions like that, we probably could answer it. IBM Watson did pretty well on Jeopardy, because it was trained to answer questions like that. I think we definitely have the databases, the knowledge bases, to answer questions like that. The problem is that kind of a question is really outside of the domain of most of the personal assistants that are being built as products today because honestly, trivia bots are fun, but they’re not as useful as a thing that can set a timer, or check the weather, or play a song. So those are mostly the things that those systems are focused on.
Fair enough, but I would differ. You can go to Wolfram Alpha and say, “What’s bigger, the Statue of Liberty or the Empire State Building?” and it’ll answer that. And you can ask Amazon’s product that same question, and it’ll answer it. Is that because those are legit questions and my question is not legit, or is it because we haven’t taught systems to disintermediate very well and so they don’t really know what I mean when I say “sun”?
I think that’s probably the issue. There’s a language modeling problem when you say, “What’s bigger, a nickel or the sun?” The sun can mean so many different things, like you were saying. Nickel, actually, can be spelled a couple of different ways and has a couple of different meanings. Dealing with ambiguities like that is a little bit hard. I think when you ask that question to me, I categorize this as a trivia question, and so I’m able to disambiguate all of those things, and look up the answer in my little knowledge base in my head, and answer your question. But I actually don’t think that particular question is impossible to solve. I just think it’s just not been a focus to try to solve stuff like that, and that’s why they’re not good.
AIs have done a really good job playing games: Deep Blue, Watson, AlphaGo, and all of that. I guess those are constrained environments with a fixed set of rules, and it’s easy to understand who wins, and what a point is, and all that. What is going to be the next thing, that’s a watershed event, that happens? Now they can outbluff people in poker. What’s something that’s going to be, in a year, or two years, five years down the road, that one day, it wasn’t like that in the universe, and the next day it was? And the next day, the best Go player in the world was a machine.
The thing that’s on my mind for that right now is autonomous vehicles. I think it’s going to change the world forever to unchain people from the driver’s seat. It’s going to give people hugely increased mobility. I have relatives that their doctors have asked them to stop driving cars because it’s no longer safe for them to be doing that, and it restricts their ability to get around the world, and that frustrates them. It’s going to change the way that we all live. It’s going to change the real estate markets, because we won’t have to park our cars in the same places that we’re going to. It’s going to change some things about the economy, because there’s going to be new delivery mechanisms that will become economically viable. I think intelligence that can help robots essentially drive around the roads, that’s the next thing that I’m most excited about, that I think is really going to change everything.
We’ll come to that in just a minute, but I’m actually asking…We have self-driving cars, and on an evolutionary basis, they’ll get a little better and a little better. You’ll see them more and more, and then someday there’ll be even more of them, and then they’ll be this and this and this. It’s not that surprise moment, though, of AlphaGo just beat Lee Sedol at Go. I’m wondering if there is something else like that—that it’s this binary milestone that we can all keep our eye open for?
I don’t know. As far as we have self-driving cars already, I don’t have a self-driving car that could say, for example, let me sit in it at nighttime, go to sleep and wake up, and it brought me to Disneyland. I would like that kind of self-driving car, but that car doesn’t exist yet. I think self-driving trucks that can go cross country carrying stuff, that’s going to radically change the way that we distribute things. I do think that we have, as you said, we’re on the evolutionary path to self-driving cars, but there’s going to be some discrete moments when people actually start using them to do new things that will feel pretty significant.
As far as games and stuff, and computers being better at games than people, it’s funny because I feel like Silicon Valley has, sometimes, a very linear idea of intelligence. That one person is smarter than another person maybe because of an SAT score, or an IQ test, or something. They use that sort of linearity of an intelligence to where some people feel threatened by artificial intelligence because they extrapolate that artificial intelligence is getting smarter and smarter along this linear scale, and that’s going to lead to all sorts of surprising things, like Lee Sedol losing to Go, but on a much bigger scale for all of us. I feel kind of the opposite. Intelligence is such a multidimensional thing. The fact that a computer is better at Go then I am doesn’t really change my life very much, because I’m not very good at Go. I don’t play Go. I don’t consider Go to be an important part of my intelligence. Same with chess. When Gary Kasparov lost to Deep Blue, that didn’t threaten my intelligence. I am sort of defining the way that I work and how I add value to the world, and what things make me happy on a lot of other axes besides “Can I play chess?” or “Can I play Go?” I think that speaks to the idea that intelligence really is very multifaceted. There’s a lot of different kinds – there’s probably thousands or millions of different kinds of intelligence – and it’s not very linearizable.
Because of that, I feel like, as we watch artificial intelligence develop, we’re going to see increasingly more intelligent machines, but they’re going to be increasingly more intelligent in some very narrow domains like “this is the better Go-playing robot than me”, or “this is the better car driver than me”. That’s going to be incredibly useful, but it’s not going to change the way that I think about myself, or about my work, or about what makes me happy. Because I feel like there are so many more dimensions of intelligence that are going to remain the province of humans. That’s going to take a very long time, if ever, for artificial intelligence to become better at all of them than us. Because, as I said, I don’t believe that intelligence is a linearizable thing.
And you said you weren’t a philosopher. I guess the thing that’s interesting to people, is there was a time when information couldn’t travel faster than a horse. And then the train came along, and information could travel. That’s why in the old Westerns – if they ever made it on the train, that was it, and they were out of range. Nothing traveled faster than the train. Then we had a telegraph and, all of a sudden, that was this amazing thing that information could travel at the speed of light. And then one time they ran these cables under the ocean, and somebody in England could talk to somebody in the United States instantly. Each one of them, and I think it’s just an opportunity to pause, and reflect, and to mark a milestone, and to think about what it all means. I think that’s why a computer just beat these awesome poker players. It learned to bluff. You just kind of want to think about it.
So let’s talk about jobs for a moment because you’ve been talking around that for just a second. Just to set the question up: Generally speaking, there are three views of what automation and artificial intelligence are going to do to jobs. One of them reflects kind of what you were saying is that there are going to be a certain group of workers who are considered low skilled, and there are going to be automation that takes these low-skilled jobs, and that there’s going to be a sizable part of the population that’s locked out of the labor market, and it’s kind of like the permanent Great Depression over and over and over forever. Then there’s another view that says, “No, you don’t understand. There’s going to be an inflection point where they can do every single thing. They’re going to be a better conductor and a better painter and a better novelist and a better everything than us. Don’t think that you’ve got something that a machine can’t do.” Clearly, that isn’t your viewpoint from what you said. Then there’s a third viewpoint that says, “No, in the past, even when we had these transformative technologies like electricity and mechanization, people take those technologies and they use them to increase their own productivity and, therefore, their own incomes. And you never have unemployment go up because of them, because people just take it and make a new job with it.” Of those three, or maybe a fourth one I didn’t cover; where do you find yourself?
I feel like I’m closer in spirit to number three. I’m optimistic. I believe that the primary way that we should expect economic growth in the future is by increased productivity. If you buy a house or buy some stock and you want to sell it 20 or 30 years from now, who’s going to buy it, and with what money, and why do you expect the price to go up? I think the answer to that question should be the people in the future should have more money than us because they’re more productive, and that’s why we should expect our world economy to continue growing. Because we find more productivity. I actually feel like this is actually necessary. World productivity growth has been slowing for the past several decades, and I feel like artificial intelligence is our way out of this trap where we have been unable to figure out how to grow our economy because our productivity hasn’t been improving. I actually feel like this is a necessary thing for all of us, is to figure out how to improve productivity, and I think AI is the way that we’re going to do that for the next several decades.
The one thing that I disagreed with in your third statement was this idea that unemployment would never go up. I think nothing is ever that simple. I actually am quite concerned about job displacement in the short-term. I think there will be people that suffer and in fact, I think, to a certain extent, this is already happening. The election of Donald Trump was an eye-opener to me that there really exists a lot of people that feel that they have been left behind by the economy, and they come to very different conclusions about the world than I might. I think that it’s possible that, as we continue to digitize our society, and AI becomes a lever that some people will become very good at using to increase their productivity, that we’re going to see increased inequality and that worries me.
The primary challenges that I’m worried about, for our society, with the rise of AI, have to do more with making sure that we give people purpose and meaning in their life that maybe doesn’t necessarily revolve around punching out a timecard, and showing up to work at 8 o’clock in the morning every day. I want to believe that that future exists. There are a lot of people right now that are brilliant people that have a lot that they could be contributing in many different ways – intellectually, artistically – that are currently not given that opportunity, because they maybe grew up in a place that didn’t have the right opportunities for them to get the right education so that they could apply their skills in that way, and many of them are doing jobs that I think don’t allow them to use their full potential.
So I’m hoping that, as we automate many of those jobs, that more people will be able to find work that provides meaning and purpose to them and allows them to actually use their talents and make the world a better place, but I acknowledge that it’s not going to be an easy transition. I do think that there’s going to be a lot of implications for how our government works and how our economy works, and I hope that we can figure out a way to help defray some of the pain that will happen during this transition.
You talked about two things. You mentioned income inequality as a thing, but then you also said, “I think we’re going to have unemployment from these technologies.” Separating those for a minute and just looking at the unemployment one for a minute, you say things are never that simple. But with the exception of the Great Depression, which nobody believes was caused by technology, unemployment has been between 5% and 10% in this country for 250 years and it only moves between 5% and 10% because of the business cycle, but there aren’t counterexamples. Just imagine if your job was you had animals that performed physical labor. They pulled, and pushed, and all of that. And somebody made the steam engine. That was disruptive. But even when we had that, we had electrification of industry. We adopted steam power. We went from 5% to 85% of our power being generated by steam in just 22 years. And even when you had that kind of disruption, you still didn’t have any increases in unemployment. I’m curious, what is the mechanism, in your mind, by which this time is different?
I think that’s a good point that you raise, and I actually haven’t studied all of those other transitions that our society has gone through. I’d like to believe that it’s not different. That would be a great story if we could all come to agreement, that we won’t see increased unemployment from AI. I think the reason why I’m a little bit worried is that I think this transition in some fields will happen quickly, maybe more quickly than some of the transitions in the past did. Just because, as I was saying, AI is easier to replicate than some other technologies, like electrification of a country. It takes a lot of time to build out physical infrastructure that can actually deliver that. Whereas I think for a lot of AI applications, that infrastructure will be cheaper and quicker to build, so the velocity of the change might be faster and that could lead to a little bit more shock. But it’s an interesting point you raise, and I certainly hope that we can find a way through this transition that is less painful than I’m worried it could be.
Do you worry about misuse of AI? I’m an optimist on all of this. And I know that every time we have some new technology come along, people are always looking at the bad cases. You take something like the internet, and the internet has overwhelmingly been a force for good. It connects people in a profound way. There’s a million things. And yeah, some people abuse it. But on net, all technology, I believe, almost all technology on net is used for good because I think, on net, people, on average, are more inclined to build than to destroy. That being said, do you worry about nefarious uses of AI, specifically in warfare?
Yeah. I think that there definitely are going to be some scary killer robots that armies make. Armies love to build machinery that kills things and AI will help them do that, and that will be scary. I think it’s interesting, like, where is the real threat going to come from? Sometimes, I feel like the threat of malevolent AI being deployed against people is going to be more subtle than that. It’s going to be more about things that you can do after compromising fiber systems of some adversary, and things that you can do to manipulate them using AI. There’s been a lot of discussion about Russian involvement in the 2016 election in the US, and that wasn’t about sending evil killer robots. It was more about changing people’s opinions, or attempting to change their opinions, and AI will give entities tools to do that on a scale that maybe we haven’t seen before. I think there may be nefarious uses of AI that are more subtle and harder to see than a full-frontal assault from a movie with evil killer robots. I do worry about all of those things, but I also share your optimism. I think we humans, we make lots of mistakes and we shouldn’t give ourselves too easy of a time here. We should learn from those mistakes, but we also do a lot of things well. And we have used technologies in the past to make the world better, and I hope AI will do so as well.
Pedro Domingo wrote a book called The Master Algorithm where he says there are all of these different tools and techniques that we use in artificial intelligence. And he surmises that there is probably a grandparent algorithm, the master algorithm, that can solve any problem, any range of problems. Does that seem possible to you or likely, or do you have any thoughts on that?
I think it’s a little bit far away, at least from AI as it’s practiced today. Right now, the practical, on-the-ground experience of researchers trying to use AI to do something new is filled with a lot of pain, suffering, blood, sweat, tears, and perseverance if they are to succeed, and I see that in my lab every day. Most of the researchers – and I have brilliant researchers in my lab that are working very hard, and they’re doing amazing work. And most of the things they try fail. And they have to keep trying. I think that’s generally the case right now across all the people that are working on AI. The thing that’s different is we’ve actually started to see some big successes, along with all of those more frustrating everyday occurrences. So I do think that we’re making the progress, but I think having a master algorithm that’s pushbutton that can solve any problem you pose to it that’s something that’s hard for me to conceive of with today’s state of artificial intelligence.
AI, of course, it’s doubtful we’ll have another AI winter because, like you said, it’s kind of delivering the goods, and there have been three things that have happened that made that possible. One of them is better hardware, and obviously you’re part of that world. The second thing is better algorithms. We’ve learned to do things a lot smarter. And the third thing is we have more data, because we are able to collect it, and store it, and whatnot. Assuming you think the hardware is the biggest of the driving factors, what would you think has been the bigger advance? Is it that we have so much more data, or so much better algorithms?
I think the most important thing is more data. I think the algorithms that we’re using in AI right now are, more or less, clever variations of algorithms that have been around for decades, and used to not work. When I was a PhD student and I was studying AI, all the smart people told me, “Don’t work with deep learning, because it doesn’t work. Use this other algorithm called support vector machines.” Which, at the time, that was the hope that that was going to be the master algorithm. So I stayed away from deep learning back then because, at the time, it didn’t work. I think now we have so much more data, and deep learning models have been so successful at taking advantage of that data, that we’ve been able to make a lot of progress. I wouldn’t characterize deep learning as a master algorithm, though, because deep learning is like a fuzzy cloud of things that have some relationships to each other, but actually finding a space inside that fuzzy cloud to solve a particular problem requires a lot of human ingenuity.
Is there a phrase – it’s such a jargon-loaded industry now – are there any of the words that you just find rub you the wrong way? Because they don’t mean anything and people use them as if they do? Do you have anything like that?
Everybody has pet peeves. I would say that my biggest pet peeve right now is the word neuromorphic. I have almost an allergic reaction every time I hear that word, mostly because I don’t think we know what neurons are or what they do, and I think modeling neurons in a way that actually could lead to brain simulations that actually worked is a very long project that we’re decades away from solving. I could be wrong on that. I’m always waiting for somebody to prove me wrong. Strong opinions, weakly held. But so far, neuromorphic is a word that I just have an allergic reaction to, every time.
Tell me about what you do. You are the head of Applied AI Research at NVIDIA, so what does your day look like? What does your team work on? What’s your biggest challenge right now, and all of that?
NVIDIA sells GPUs which have powered most of the deep learning revolution, so pretty much all of the work that’s going on with deep learning across the entire world right now, runs on NVIDIA GPUs. And that’s been very exciting for NVIDIA, and exciting for me to be involved in building that. The next step, I think, for NVIDIA is to figure out how to use AI to change the way that it does its own work. NVIDIA is incentivized to do this because we see the value that AI is bringing to our customers. Our GPU sales have been going up quite a bit because we’re providing a lot of value to everyone else who’s trying to use AI for their own problems. So the next step is to figure out how to use AI for NVIDIA’s problems directly. Andrew Ng, who I used to work with, has this great quote that “AI is the new electricity,” and I believe that. I think that we’re going to see AI applied in many different ways to many different kinds of problems, and my job at NVIDIA is to figure out how to do that here. So that’s what my team focuses on.
We have projects going on in quite a few different domains, ranging from graphics to audio, and text, and others. We’re trying to change the way that everything at NVIDIA happens: from chip design, to video games, and everything in between. As far as my day-to-day work goes, I lead this team, so that means I spend a lot of time talking with people on the team about the work that they’re doing, and trying to make sure they have the right resources, data, the right hardware, the right ideas, the right connections, so that they can make progress on problems that they’re trying to solve. Then when we have prototypes that we’ve built showing how to apply AI to a particular problem, then I work with people around the company to show them the promise of AI applied to problems that they care about.
I think one of the things that’s really exciting to me about this mission is that we’re really trying to change NVIDIA’s work at the core of the company. So rather than working on applied AI, that could maybe help some peripheral part of the company that maybe could be nice if we did that, we’re actually trying to solve very fundamental problems that the company faces with AI, and hopefully we’ll be able to change the way that the company does business, and transform NVIDIA into an AI company, and not just a company that makes hardware for AI.
You are the head of the Applied AI Research. Is there a Pure AI Research group, as well?
Yes, there is.
So everything you do, you have an internal customer for already?
That’s the idea. To me, the difference between fundamental research and applied research is more a question of emphasis on what’s the fundamental goal of your work. If the goal is academic novelty, that would be fundamental research. Our goal is, we think about applications all the time, and we don’t work on problems unless we have a clear application that we’re trying to build that could use a solution.
In most cases, do other groups come to you and say, “We have this problem we really want to solve. Can you help us?” Or is the science nascent enough that you go and say, “Did you know that we can actually solve this problem for you?”
It kind of works all of those ways. We have a list of projects that people around the company have proposed to us, and we also have a list of projects that we ourselves think are interesting to look at. There’s also a few projects that my management tells me, “I really want you to look at this problem. I think it’s really important.” We get input from all directions, and then prioritize, and go after the ones we think are most feasible, and most important.
And do you find a talent shortage? You’re NVIDIA on the one hand, but on the other hand, you know: it’s AI.
I think the entire field, no matter what company you work at, the entire field has a shortage of qualified scientists that can do AI research, and that’s despite the fact that the amount of people jumping into AI is increasing every year. If you go to any of the academic AI conferences, you’ll see how much energy and how much excitement, and how many people that are there that didn’t used to be there. That’s really wonderful to see. But even with all of that growth and change, it is a big problem for the industry. So, to all of your listeners that are trying to figure out what to do next, come work on AI. We have lots of fun problems to work on, and not nearly enough people doing it.
I know a lot of your projects I’m sure you can’t talk about, but tell me something you have done, that you can talk about, and what the goal was, and what you were able to achieve. Give us a success story.
I’ll give you one that’s relevant to the last question that you asked, which is about how to find talent for AI. We’ve actually built a system that can match candidates to job openings at NVIDIA. Basically, it can predict how well we think a particular candidate is a fit for a particular job. That system is actually performing pretty well. So we’re trialing it with hiring managers around the company to figure out if it can help them be more efficient in their work as they search for people to come join NVIDIA.
That looks like a game, isn’t it? I assume you have a pool of resumes or LinkedIn profiles or whatever, and then you have a pool of successful employees, and you have a pool of job descriptions and you’re trying to say, “How can I pull from that big pool, based on these job descriptions, and actually pick the people that did well in the end?”
That’s right.
That’s like a game, right? You have points.
That’s right.
Would you ever productize anything, or is everything that you’re doing just for your own use?
We focus primarily on building prototypes, not products, in my team. I think that’s what the research is about. Once we build a prototype that shows promise for a particular problem, then we work with other people in the company to get that actually deployed, and they would be the people that think about business strategy about whether something should be productized, or not.
But you, in theory, might turn “NVIDIA Resume Pro” into something people could use?
Possibly. NVIDIA also works with a lot of other companies. As we enable companies in many different parts of the economy to apply AI to their problems, we work with them to help them do that. So it might make more sense for us, for example, to deliver this prototype to some of our partners that are in a position to deliver products like this more directly, and then they can figure out how to enlarge its capabilities, and make it more general to try to solve bigger problems that address their whole market and not just one company’s needs. Partnering with other companies is good for NVIDIA because it helps us grow AI which is something we want to do because, as AI grows, we grow. Personally, I think some of the things that we’re working on; it just doesn’t really make sense. It’s not really in NVIDIA’s DNA to productize them directly because it’s just not the business model that the company has.
I’m sure you’re familiar with the “right to know” legislation in Europe: the idea that if an AI makes a decision about you, you have a right to know why it made that decision. AI researchers are like, “It’s not necessarily that easy to do that.” So in your case, your AI would actually be subject to that. It would say, “Why did you pick that person over this person for that job?” Is that an answerable question?
First of all, I don’t think that this system – or I can’t imagine – using it to actually make hiring decisions. I think that would be irresponsible. This system makes mistakes. What we’re trying to do is improve productivity. If instead of having to sort through 200 resumes to find 3 that I want to talk to—if I can look at 10 instead—then that’s a pretty good improvement in my productivity, but I’m still going to be involved, as a hiring manager, to figure out who is the right fit for my jobs.
But an AI excluded 190 people from that position.
It didn’t exclude them. It sorted them, and then the person decided how to allocate their time in a search.
Let’s look at the problem more abstractly. What do you think, just in general, about the idea that every decision an AI makes, should be, and can be, explained?
I think it’s a little bit utopian. Certainly, I don’t have the ability to explain all of the decisions that I make, and people, generally, are not very good at explaining their decisions, which is why there are significant legal battles going on about factual things, that people see in different ways, and remember in different ways. So asking a person to explain their intent is actually a very complicated thing, and we’re not actually very good at it. So I don’t actually think that we’re going to be able to enforce that AI is able to explain all of its decisions in a way that makes sense to humans. I do think that there are things that we can do to make the results of these systems more interpretable. For example, on the resume job description matching system that I mentioned earlier, we’ve built a prototype that can highlight parts of the resume that were most interesting to the model, both in a positive, and in a negative sense. That’s a baby step towards interpretability so that if you were to pull up that job description and a particular person and you could see how they matched, that might explain to you what the model was paying attention to as it made a ranking.
It’s funny because when you hear reasons why people exclude a resume, I remember one person said, “I’m not going to hire him. He has the same first name as somebody else on the team. That’d just be too confusing.” And somebody else I remember said that the applicant was a vegan and the place they like to order pizza from didn’t have a vegan alternative that the team liked to order from. Those are anecdotal of course, but people use all kinds of other things when they’re thinking about it.
Yeah. That’s actually one of the reasons why I’m excited about this particular system is that I feel like we should be able to construct it in a way that actually has fewer biases than people do, because we know that people harbor all sorts of biases. We have employment laws that guide us to stay away from making decisions based on protected classes. I don’t know if veganism is a protected class, but it’s verging on that. If you’re making hiring decisions based on people’s personal lifestyle choices, that’s suspect. You could get in trouble for that. Our models, we should be able to train them to be more dispassionate than any human could be.
We’re running out of time. Let’s close up by: do you consume science fiction? Do you ever watch movies or read books or any of that? And if so, is there any of it that you look at, especially any that portrays artificial intelligence, like Ex Machina, or Her, or Westworld or any of that stuff, that you look at and you’re like, “Wow, that’s really interesting,” or “That could happen,” or “That’s fascinating,” or anything like that?
I do consume science fiction. I love science fiction. I don’t actually feel like current science fiction matches my understanding of AI very well. Ex Machina, for example, that was a fun movie. I enjoyed watching that movie, but I felt, from a scientific point of view, it just wasn’t very interesting. I was talking about our built-in models of the world. One of the things that humans, over thousands of years, have drilled into our heads is that there’s somebody out to get you. We have a large part of our brain that’s worrying all the time, like, “Who’s going to come kill me tonight? Who’s going to take away my job? Who’s going to take my food? Who’s going to burn down my house?” There’s all these things that we worry about. So a lot of the depictions of AI in science fiction inflame that part of the brain that is worrying about the future, rather than actually speak to the technology and its potential.
I think probably the part of science fiction that has had the most impact on my thoughts about AI is Isaac Asimov’s Three Laws. Those, I think, are pretty classic, and I hope that some of them can be adapted to the kinds of problems that we’re trying to solve with AI, to make AI safe, and make it possible for people to feel confident that they’re interacting with AI, and not worry about it. But I feel like most of science fiction is, especially movies – maybe books can be a little bit more intellectual and maybe a little bit more interesting – but especially movies, it just sells more movies to make people afraid, than it does to show people a mundane existence where AI is helping people live better lives. It’s just not nearly as compelling of a movie, so I don’t actually feel like popular culture treatment of AI is very realistic.
All right. Well, on that note, I say, we wrap up. I want to thank you for a great hour. We covered a lot of ground, and I appreciate you traveling all that way with me.
It was fun.
Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here
[voices_in_ai_link_back]

Voices in AI – Episode 11: A Conversation with Gregory Piatetsky-Shapiro

[voices_in_ai_byline]
In this episode, Byron and Gregory talk about consciousness, jobs, data science, transfer learning.
[podcast_player name=”Episode 11: A Conversation with Gregory Piatetsky-Shapiro” artist=”Byron Reese” album=”Voices in AI” url=”https://voicesinai.s3.amazonaws.com/2017-10-16-(00-43-05)-gregory-piatestsky.mp3″ cover_art_url=”https://voicesinai.com/wp-content/uploads/2017/10/voices-headshot-card-3.jpg”]
[voices_in_ai_link_back]
Byron Reese: This is “Voices in AI”, brought to you by Gigaom. I’m Byron Reese. Today our guest is Gregory Piatetsky. He’s a leading voice in Business Analytics, Data Mining, and Data Science. Twenty years ago, he founded and continues to operate a site called KDnuggets about knowledge discovery. It’s dedicated to the various topics he’s interested in. Many people think it’s a must-read resource. It has over 400,000 regular monthly readers. He holds an MS and a PhD in computer science from NYU. 
Welcome to the show.
Gregory Piatetsky: Thank you, Byron. Glad to be with you.
I always like to start off with definitions, because in a way we’re in such a nascent field in the grand scheme of things that people don’t necessarily start off agreeing on what terms mean. How do you define artificial intelligence?
Artificial intelligence is really machines doing things that people think require intelligence, and by that definition the goalposts of artificial intelligence are constantly moving. It was considered very intelligent to play checkers back in the 1950s, then there was a program. The next boundary was playing chess, and then computers mastered it. Then people thought playing Go would be incredibly difficult, or driving cars. General artificial intelligence is the field that tries to develop intelligent machines. And what is intelligence? I’m sure we will discuss, but it’s usually in the eye of the beholder.
Well, you’re right. I think a lot of the problem with the term artificial intelligence is that there is no consensus definition of what intelligence is. So, are you saying if we’re constantly moving the goalposts, it sounds like you’re saying we don’t have systems today that are intelligent.
No, no. On the contrary, we have lots of systems today that would have been considered amazingly intelligent 20 or even 10 years ago. And the progress is such that I think it’s very likely that those systems will exceed our intelligence in many areas, you know maybe not everywhere, but in many narrow, defined areas they’ve already exceeded our intelligence. We have many systems that are somewhat useful. We don’t have any systems that are fully intelligent, possessing what is a new term now, AGI, Artificial General Intelligence. Those systems remain still ahead in the future.
Well, let’s talk about that. Let’s talk about an AGI. We have a set of techniques that we use to build the weak or narrow AI we use today. Do you think that achieving an AGI is just continuing to apply to evolve those faster chips, better algorithms, bigger datasets, and all of that? Or do you think that an AGI really is qualitatively a different thing?
I think AGI is qualitatively a different thing, but I think that it is not only achievable but also inevitable. Humans also can be considered as biological machines, so unless there is something magical that we possess that we cannot transfer to machines, I think it’s quite possible that the smartest people can develop some of the smartest algorithms, and machines can eventually achieve AGI. And I’m sure it will require additional breakthroughs. Just like deep learning was a major breakthrough that contributed to significant advances in state of the art, I think we will see several such great breakthroughs before AGI is achieved.
So if you read the press about it and you look at people’s predictions on when we might get an AGI, they range, in my experience, from 5 to 500 years, which is a pretty telling fact alone that it’s that kind of range. Do you care to even throw in a dart in that general area? Like do you think you’ll live to see it or not?
Well, my specialty as a data scientist is making predictions, and I know when we don’t have enough information. I think nobody really knows. And I have no basis on which to make a prediction. I hope it’s not 5 years and I think our experience as a society shows that we have no idea how to make predictions for 100 years from now. It’s very instructive to find so-called futurology articles, things that were written 50 years ago about what will happen in 50 years, and see how naive were those people 50 years ago. I don’t think we will be very successful in predicting in 50 years. I have no idea how long it will take, but I think it will be more than 5 years.
So some people think that what makes us intelligent, or an indispensable part of our intelligence, is our consciousness. Do you think a machine would need to achieve consciousness in order to be an AGI?
We don’t know what is consciousness. I think machine intelligence would be very different from human intelligence, just like airplane flight is very different from a bird, you know. Both airplanes and birds fly, the flight is governed by the same laws of aerodynamics and physics, but they use very different principles. The airplane flight does not copy bird flight, it is inspired by it. I think in the same way, we’re likely to see that machine intelligence doesn’t copy human intelligence, or human consciousness. “What exactly is consciousness?” is more a question for philosophers, but probably it involves some form of self-awareness. And we can certainly see that machines and robots can develop self-awareness. And you know, self-driving cars already need to do some of that. They need to know exactly where they’re located. They need to predict what will happen. If they do something, what will other cars do? They have a form that is called model of the mind, mirror intelligence. One interesting anecdote on this topic is that when Google’s self-driving car was originally started their experiments, it couldn’t cross the intersection because it was always yielding to other cars. It was following the rules as they were written, but not the rules as people actually execute them. And so it was stuck at that intersection supposedly for an hour or so. Then the engineers adjusted the algorithm so it would better predict what people will do and what it will do, and it’s now able to negotiate the intersections. It has some form of self-awareness. I think other robots and machine intelligence will develop some form of self-awareness, and whether it will be called consciousness or not will be to our descendants to discuss.
Well, I think that there is an agreed upon definition of consciousness. I mean, you’re right that nobody knows how it comes about, but it’s qualia, it’s experiencing things. It’s, if you’ve ever had that sensation when you’re driving and you kind of space, and all of a sudden two miles later you kind of snap to and think, “Oh my gosh, I’ve got no recollection of how I got here.” That time you were driving, that’s intelligence without consciousness. And then when you kind of snap to, and all of the sudden you’re aware, you’re experiencing the world again. Do you think a computer can actually experience something? Because wouldn’t it need to experience the world in order to really be intelligent?
Well computers, if they have sensors, actually they already experience the world. The self-driving car is experiencing the world through its radar and LIDAR and various other sensors and so on, so they do experience and they do have sensors. I think it’s not useful to debate computer consciousness, because it’s like a question of, you know, how many angels can fit on the pin of a needle. I think what we can discuss is what they can or cannot do. How they experience it is more a question for philosophers.
So a lot of people are worried – you know all of this, of course – there’s two big buckets of worry about artificial intelligence. The first one is that it’s going to take human jobs and they’re going to have mass unemployment, and any number of dystopian movies play that scenario out. And then other people say, no, every technology that’s come along, even disruptive ones like electricity, and mechanical power replacing animal power and all of that, were merely then turned around and used by humans to increase their productivity, and that’s how you get increases in standard of living. On that question, where do you come down?
I’m much more worried than I am optimistic. I’m optimistic that technology will progress. What I’m concerned with is it will lead to increasing inequality and increasingly unequal distribution of wealth and benefits. In Massachusetts, there used to be many toll collectors. And toll collector is not a very sophisticated job, but recently they were eliminated. And the machines that eliminated them didn’t require full intelligence, basically just an RFID sensor. So we already see many jobs being eliminated by a simpler form of automation. And what society will do about it is not clear. I think the previous disruptions had much longer timespans. But now when people like these toll collectors are being laid off, they don’t have enough time to retrain themselves to become, let’s say computer programmers or doctors. What I’d like to do about it, I’m not sure. But I like a proposal by Andrew Ng, who was from Stanford Coursera. Andrew, he proposed the modified version of basic income, that people who are unemployed and cannot find jobs get some form of basic income. Not just to sit around, but they would be required to learn new skills and learn something new and useful. So maybe that would be a possible solution.
So do you really think that when you look back across time – you know, the United States, I can only speak to that, went from generating 5% of its energy with steam to 80% in just 22 years. Electrification happened electrifyingly fast. The minute we had engines there was wholesale replacement of the animals, they were just so much more efficient. Isn’t it actually the case that when these destructive technologies come along, they are so empowering that they are actually adopted incredibly quickly? And again, just talking about the US, unemployment for 230 years has been between 5% and 9%, other than the Great Depression, but in all the other time, it never bumped. When these highly disruptive technologies came along, it didn’t cause unemployment generally to go up, and they happened quickly, and they eliminated an enormous number of positions. Why do you think this one is different?
The main reason why I think it is different is because it is qualitatively different. Previously, the machines that came, like the steam and electricity-driven, it would eliminate some of the manual work and people could climb up on the pyramid of skills to do more sophisticated work. But nowadays, artificial general intelligence sort of captures this pyramid of skills, and it now competes with people on the cognitive skills. And it can eventually climb to the top of the pyramid, so there will be nowhere to climb to exceed it. And once you generate one general intelligence, it’s very easy to copy it. So you would have a very large number, let’s say, of intelligent robots that will do a very large number of things. They will compete with people to do other things. It’s just very hard to retrain, let’s say, a coal miner to become, let’s say, producer of YouTube videos.
Well that isn’t really how it ever happens, is it? I mean, that’s kind of a rigged set-up, isn’t it? What matters is, can everybody do a job a little bit harder than they have? Because the maker of YouTube videos is a film student. And then somebody else goes to film school, and then the junior college professor decides to… I mean, everybody just goes up a little bit. You never take one group of people and train them to do an incredibly radically different thing, do you?
Well, I don’t know about that exactly, but to return to your analogy, you mentioned that the United States for 200 years the pattern was such. But, you know, the United States is not the only country in the world, and 200 years is a very small part of our history. We look at several thousand years, and look with what happened in the north, we see they’re very complex things. Unemployment rate in the Middle Ages was much higher than 5% or 10%.
Well, I think the important thing, and the reason why I used 200 years is because that’s the period of industrialization that we’ve seen, and automation. And so the argument is Artificial Intelligence is going to automate jobs, so you really only need to look over the period you’ve had other things automating jobs to say, “What happens when you automate a lot of jobs?” I mean, by your analogy, wouldn’t the invention of the calculator have put mathematicians out of business? I mean like with ATM machines, an ATM machine in theory replaces a bank teller. And yet we have more bank tellers today than we did when the ATM was introduced, because that too allows banks to open more branches and hire more tellers. I mean, is it really as simple as, “Well, you’ve built this tool, now there’s a machine doing a job a human did and now you have an unemployed human.” Is that kind of the only force at work?
Of course it’s not simple, there are many forces at work. And there are forces that resist change, as we’ve seen from Luddites in 18th century. And now there are people, for example coal mining districts, who want to go back to coal mining. Of course, it’s not that simple. What I’m saying is we only had a few examples of industrial revolutions, and as data scientists say, it’s very hard to generalize from few examples. It’s true that past technologies have generated more work. It doesn’t follow that this new technology, which is different, will generate more work for all the people. It may very well be different. We cannot rely on three or four past examples to generalize for the future.
Fair enough. So let’s talk, if we can, about how you spend your days, which is in data science, what are some recent advances that you think have materially changed the job of a data scientist? Are there ones? And are there more things that you can kind of see that are about to change and begin? Like how is that job evolving as technology changes?
Yes, well data scientists now live in the golden age of the field. There are now more powerful tools that make data science much easier, tools like Python and R. And Python and R both have a very large ecosystem of tools, like scikit-learn for example in the case of Python, or whatever Hadley Wickham comes up in the case of R. There are tools like Spark and various things on top of that that allow data scientists to access very large amount of data. It’s much easier and much faster for data scientists to build models. The danger for data scientists, again, is automation, because as those tools make it easier and easier, and soon they make the work, you know, a large part of it automated. In fact, there are already companies like DataRobot and others that allow business users who are not data scientists just to plug their data, and DataRobot or their competitors just generate the results. No data scientist needed. That is already happening in many areas. For example, ads on the internet are automatically placed, and there are algorithms that make millions of decisions per second and build lots of models. Again, no human involvement because humans just cannot do millions of models a second. There are many areas where this automation is already happening. And recently I had a poll in KDnuggets asking, when do you think data science work will be automated? Then the median answer was about 20 or 25. So although this is a golden age for data scientists, I think they should enjoy it because who knows what will happen in the next 8 to 10 years.
So, when Mark Cuban was talking about the first – he gave a talk earlier this year – he said the first trillionaires will be in businesses that utilize AI. But he said something very interesting, which is, he said that if he were coming up through university again, he would study philosophy. That’s the last thing that’s going to be automated. What would you suggest to a young person today listening to this? What do you think they should study, in the cognitive area, that is either blossoming or is it likely to go away?
I think what will be very much in demand is at the intersection of humanities and technology. If I was younger I would still study machine learning and databases, which is actually what I studied for my PhD 30 years ago. I probably would study more mathematics. The deep learning algorithms that are making tremendous advances are very mathematically intensive. And the other aspect is, kind of maybe the hardest to automate is human intuition and empathy, understanding what other people need and want, and how to best connect with them. I don’t know how much that can be studied, but if philosophy or social studies or poetry is the way to it, then I would encourage young people to study it. I think we need a balanced approach, not just technology but humanities as well.
So, I’m intrigued that our DNA is– I’m going to be off here, whatever I say. I think is about is about 740 meg, it’s on that order. But when you look at how much of it we share with, let’s say, a banana, it’s 80-something percent, and then how much we share with a chimp, it’s 99%. So somewhere in that 1%, that 7 or 8 meg of code that tells how to build you, is the secret to artificial general intelligence, presumably. Is it possible that the code to do an AGI is really quite modest and simple? Not simple – you know, there’s two different camps in the AGI world. And one is that humans are a hack of 100 or 200 or 300 different skills that you put them all together and that’s us. Another one is, we had Pedro Domingos on the show and he had a book called The Master Algorithm, which posits that there is an algorithm that can solve any problem, or any solvable problem, the way human is. Where on that spectrum would you fall? And do you think there is a simple answer to an AGI?
I don’t think there is a simple answer. Actually, I’m a good friend with Pedro and I moderated his webcast on his book last year. But I think that the master algorithm that he looks for may exist, but it doesn’t exclude having lots of additional specialized skills. I think there is very good evidence that there is such a thing as general intelligence in humans, that people, for example, make have different scores on SAT on verbal and math. I know that my verbal score would be much lower than my math score. But usually if you’re above average on one, you would be above average on the other. And likewise, if you’re below average on one, you will be below average. People seem to have some general skills, and in addition there are a lot of specialized skills. You know, you can be a great chess player but have no idea how to play music, or vice versa. I think there are some general algorithms, and there are lots of specialized algorithms that leverage special structure of the domain. You can think of it this way, that when people were developing chess-playing programs, they initially applied some general algorithms, but then they found that they could speed up these programs by building specialized hardware that was very specific to chess. Likewise, people when they start new skills they approach it generally, then they develop the specialized expertise which speeds up their work. I think likewise it could be with intelligence. There may be some general algorithm, but it would have ways to develop lots of special skills that would leverage whatever specific or particular tasks.
Broadly speaking, I guess data science relies on three things: it relies on hardware, faster and faster hardware; better and better data, more of it and labeled better; and then better and better algorithms. If you kind of had to put those three things side by side, where are we most efficient? Like if you could really amp one of those three things way up, what would it be? 
That’s a very good question. With current algorithms, it seems that more data produces much better results than a smarter algorithm, especially if it is relevant data. For example, for image recognition there was a big quantitative jump when deep learning trained on millions of images as opposed to thousands of images. But I think what we need for next big advances is having somewhat smarter algorithms. One big shortcoming for deep learning is, again, it requests so much data. People seem to be able to learn from very few examples. And the algorithms that we have are not yet able to do that. In algorithm’s defense, I have to say that when I say people can learn from very few examples, we assume those are adults and they’ve already spent maybe 30 or 40 years of training interacting with the world. So maybe if algorithms can spend some years training and interacting with the world, they’ll acquire enough knowledge so they’ll be able to generalize to other similar examples. Yes, I think probably data, then algorithms, and then hardware. That would be my order.
So, you’re alluding to transfer learning, which is something humans seem to be able to do. Like you said, you could show a person who’s never seen an Academy Award, what that little statue that looks like, and then you could show them photographs of it in the dark, on its side, underwater, and they could pick it out. And what you just said is very interesting, which is, well yeah, we only had one photo of this thing, but we had a lifetime of learning how to recognize things underwater and in different lighting and all that. What do you think about transfer learning for computers? Do you think we’re going to be able to use the datasets that we have that are very mature, like the image one, or handwriting recognition, or speech translation, are we going to be able to use those to solve completely unrelated problems? Is there some kind of meta-knowledge buried in those things we’re doing really well now, that we can apply to things we don’t have good data on?
I think so. I think because the world itself is the best representation. So recently I read a paper that applied this negative transformation to ImageNet, and it turns out that now a deep learning system that was trained to recognize, I don’t remember exactly what it was, but let’s say cats, would not be able to recognize negatives of cats, because the negative transformation is not part of its repertoire. But that is very easy to remedy if you just add negative vocabulary image to the training. I think there is maybe a large but finite number of such transformations that humans are familiar with, like the negative and rotated and other things. And it’s quite possible that by doing such transformation to very large existing databases, we could teach those machine learning systems to achieve and exceed human levels. Because humans themselves are not perfect in recognition.
Earlier, this conversation we’re having, we’re taking human knowledge and how people do things and we’re kind of applying that to computers. Do you think AI researchers learn much from brain science? Do they learn much from psychology? Or is it more that’s handy for telling stories or helping people understand things? But as you started at the very beginning with airplanes and birds we were talking, there really isn’t a lot of mapping between how humans do things and how machines do them.
Yes, by the way, the airplanes and birds analogy I think is due to Yann LeCun. And I think some AI researchers are inspired by how humans do things, and the prime example is Geoff Hinton who is an amazing researcher, not only because of what he achieved, but he has extremely good understanding of both computers and human consciousness. And several talks that I’ve heard of him and some conversation afterwards, he suggested he uses his knowledge of how human brain works as an inspiration for coming up with new algorithms. Again, not copying them but inspiring the algorithms. So to answer your question, yes, I think human consciousness is very relevant to understanding how intelligence could be achieved, and as Geoff Hinton says, that’s the only working example we have at the moment.
We were able to kind of do chess in AI so easily because there were so many – not so easily, obviously people worked very hard on it – but because there were so many well-kept records of games that would be training data. We can do handwriting recognition well because we have a lot of handwriting and it’s been transcribed. We do translation well because there is a lot of training data. What are some problems that would be solvable if we just had the data for them, and we just don’t have it nor do we have any good way of getting it? Like, what’s a solvable problem that really our only impediment is that we don’t have the data?
I think at the forefront of such problem is medical diagnosis, because there are many diseases where the data already exists, it’s just maybe not collected in electronic form. There is a lot of genetic information that could be collected and correlated with both diseases and treatment, what works. Again, it’s not yet collected, but Google and 23andMe and many other companies are working on that. Medical radiology recently witnessed great success of a startup called Enlitic, where they were able to identify tumors using deep learning on almost the same quality as human radiologists. So I think in medicine and health care we will see big advances. And in many other areas where there is a lot of data, we can also see big advances. But the flipside of data, or what we can touch on it, is people, at least in some part of the political spectrum, are losing connection on whether it’s actually true or not. Last year’s election saw a tremendous amount of fake news stories that seemed to have significant influence. So while on one hand we’re training machines to do a better and better job in recognizing what is true, many humans are losing their ability to recognize what is true and what is happening. Just to witness denial of climate change by many people in this country.
You mention text analysis on your LinkedIn profile. I just saw that that was something that you evidently know a lot about. Is the problem you’re describing solvable? If you had to say the number one problem of the worldwide web is you don’t know what to believe, you don’t know what’s true, and you just don’t have a way necessarily of sorting results by truthiness, do you think that that is a machine learning problem, or is that not one? Is it going to require moderation in humans? Or is truth not a defined enough concept on which to train 50 billion web pages?
I think the technical part certainly can be solved from machine learning point of view. But the worldwide web does not exist in vacuum, it is embedded in human society. And as such, it suffers from all the advantages and problems of humans. If there are human actors that will find it beneficial to bend the truth and use the worldwide web to convince other people what they want to convince them of, they will find some ways to leverage the algorithms. The operator by itself is not a panacea as long as there are humans with all of our good and evil intentions around it.
But do you think it’s really solvable? Because I remember this Dilbert comic strip I saw once where Dilberts on a sales call and the person that he’s talking to says, “Your salesmen says your product cures cancer!” And Dilbert says, “That is true.” And the guy says, “Wait a minute! It’s true that it cures cancer or it’s true that he said that?” And so it’s like that, that statement, “Your salesperson said your product cures cancer,” is a true statement. But that subtlety, that nuance, that it’s-true-but-it’s-not-true aspect of it, I just wonder, it doesn’t feel like chess, this very clear-cut win/lose kind of situation. And I just wonder even if everybody wanted the true results to rise to the top, could we actually do that?
Again, I think technically it is possible. Of course, you know nothing will work perfectly, but humans also do not do perfect decisions. For example, Facebook already has an algorithm that can identify clickbait. And one of the signals is relatively simple, just look at the number of people, let’s say, who look at a particular headline, click on a particular link, and then how much time they spend there or whether they return and click backwards. The headline like, “Nine amazing things you can do to cure X,” and you go to that website and it’s something completely different, then you quickly return. Your behavior will be different than if you go to a website that matches the headline. And you know, Facebook and Google and other sites, they can measure those signals and they can see which type or which headlines are deceptive. The problem is that the ecosystem that has evolved seems to reward capturing attention of people, and headlines are more likely to be shared, are worth capturing attention of people, generate emotion in either anger or some cute things. We’re evolving toward internet of anger, partisan anger, and cute kittens. That’s the two extreme axes of what gets attention. I think the technical part is solvable. The problem is that, again, there are humans around it that make a very different motivation from you and me. It’s very hard to work when your enemy is using various cyber-weapons against you.
Do you think nutrition may be something that would be really hard as well? Because no two people – you eat however many times a day, however many every different foods, and there is nobody else who does that same combination on the planet, even for seven consecutive days or something. Do you think that nutrition is a solvable thing, or there are too many variables for there to ever be a dataset that would be able to say, “If you eat broccoli, chocolate ice cream, and go to the movie at 6:15, you’ll live longer?
I think that is certainly solvable. Again, the problem is that humans are not completely logical. That’s our duty and our problem. People know what is good for them, but sometimes they just want something else. We sort of have our own animal instinct that is very hard to control. That’s why all the diets work, but just not for a very long time. People who go on diets very frequently and then you know, find that it didn’t work and go on it again. Yes, for information, nutrition can be solved. How motivation to convince people to follow good nutrition, that is a much, much harder problem.
All right! Well it looks like we are out of time. Would you go ahead and tell the listeners how they can keep up with you, go on your website, and any ways they can follow you, how to get hold of you and all of that?
Yes. Thank you, Byron. You can find me on Twitter @KDnuggets, and visit the website KDnuggets.com. It’s a magazine for data scientists and machine learning professionals. We publish only a few interesting articles a day. And I hope you can read it, or if you have something to say, contribute to it! And thank you for the interview, I enjoyed it.
Thank you very much.
Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here
[voices_in_ai_link_back]

Voices in AI – Episode 9: A Conversation with Soumith Chintala

[voices_in_ai_byline]
In this episode, Byron and Soumith talk about transfer learning, child development, pain, neural networks, and adversarial networks.
[podcast_player name=”Episode 9: A Conversation with Soumith Chintala” artist=”Byron Reese” album=”Voices in AI” url=”https://voicesinai.s3.amazonaws.com/2017-10-16-(00-46-22)-soumith-chintala.mp3″ cover_art_url=”https://voicesinai.com/wp-content/uploads/2017/10/voices-headshot-card-1.jpg”]
[voices_in_ai_link_back]
Byron Reese: This is Voices in AI, brought to you by Gigaom. I’m Byron Reese. Today our guest is Soumith Chintala. He is an Artificial Intelligence Research Engineer over at Facebook. He holds a Master’s of Science and Computer Science from NYU. Welcome to the show, Soumith.
Soumith Chintala: Thanks, Byron. I am glad to be on the show.
So let’s start out with your background. How did you get to where you are today? I have been reading over your LinkedIn, and it’s pretty fascinating.
It’s almost accidental that I got into AI. I wanted to be an artist, more of a digital artist, and I went to intern at a visual effects studio. After the summer, I realized that I had no talent in that direction, so I instead picked something closer to where my core strength lies, which is programming.
I started working in computer vision, but just on my own in undergrad. And slowly and steadily, I got to CMU to do robotics research. But this was back in 2009, and still deep learning wasn’t really a thing, and AI wasn’t like a hot topic. I was doing stuff like teaching robots to play soccer and doing face recognition and stuff like that.
And then I applied for master’s programs at a bunch of places. I got into NYU, and I didn’t actually know what neural networks were or anything. Yann LeCun, in 2010, was more accessible than he is today, so I went, met with him, and I asked him what kind of computer vision work he could give me to do as a grad student. And he asked me if I knew what neural networks were, and I said no.
This was a stalwart in the field who I’m sitting in front of, and I’m like, “I don’t know, explain neural networks to me.” But he was very kind, and he guided me in the right direction. And I went on to work for a couple of years at NYU as a master’s student and simultaneously as a junior research scientist. I spent another year, almost a year there as a research scientist while also separately doing my startup.
I was part of a music and machine learning startup where we were trying to teach machines to understand and play music. That startup went south, and I was looking for new things. And at the same time, I’d started maintaining this tool called Torch, which was the industry-wide standard for deep learning back then. And so Yann asked me if I wanted to come to Facebook, because they were using a lot of Torch, and they wanted some experts in there.
That’s how I came about, and once I was at Facebook, I did a lot of things—research on adversarial networks, engineering, building PyTorch, etc.
Let’s go through some of that stuff. I’m curious about it. With regard to neural nets, in what way do you think they are similar to how the brain operates, and in what way are they completely different?
I’d say they’re completely different, period. We think they’re similar in very high-level and vague terms like, “Oh, they do hierarchical learning, like humans seem to think as well.” That’s pretty much where the similarity ends. We think, and we hypothesize, that in some very, very high-level way, artificial neural networks learn like human brains, but that’s about it.
So, the effort in Europe—the well-funded effort—The Human Brain Project, which is deliberately trying to build an AGI based on the human brain… Do you think that’s a worthwhile approach or not?
I think all scientific approaches, all scientific explorations are worthwhile, because unless we know… And it’s a reasonably motivated effort, right? It’s not like some random people with bad ideas are trying to put this together; it’s a very well-respected effort with a lot of experts.
I personally wouldn’t necessarily take that direction, because there are many approaches to these things. One is to reverse-engineer the brain at a very fundamental level, and try to put it back together exactly as it was. It’s like investigating a car engine… not knowing how it works, but taking X-ray scans of it and all that, and trying to put it back together and hoping it works.
I’m not sure if that would work with as complicated a system as the brain. So, in terms of the approach, I’m not sure I would do it the same way. But I think it’s always healthy to explore various different directions.
Some people speculate that a single neuron is as complicated in its operations as a supercomputer, which either implies we won’t get to an AGI, or we certainly won’t get it by building something like the human brain.  
Let’s talk about vision for just a minute. If I show a person just one sample of some object, a statue of a raven, and then I show them a hundred photos with it partially obscured, on its side, in the dark or half underwater, weirdly lit—a person could just boom, boom, boom, pick it all out.
But you can’t train computers anything like that. They need so many symbols, so many examples. What do you think is going on? What are humans doing that we haven’t taught computers how to do?
I think it’s just the diversity of tasks we handle every day. If we had a machine learning model that was also handling so many diverse tasks as humans do, it would be able to just pick out a raven out of a complicated image just fine. It’s just that when machines are being trained to identify ravens, they’re being trained to identify ravens from a database of images that don’t look very much like the complicated image that they’ve been given.
And because they don’t handle a diverse set of tasks, they’re doing very specific things. They kind of over-fit to that dataset they have been given, in some way. I think this is just a matter of increasing the number of tasks we can make a single machinery model do, and over time, they will get as smart. Of course, the hard problem is we haven’t figured out how to make the same model do a wide variety of tasks.
So that’s transfer learning, and it’s something humans seem to do very well.
Yes.
Does it hinder us that we take such an isolated, domain-specific view when we’re building neural AIs? We say, “Well, we can’t teach it everything, so let’s just teach it how to spot ravens,” and we reinvent the wheel each time? Do you have a gut intuition where the core, the secret of transfer learning at scale is hiding?
Yeah. It’s not that we don’t want to build models that can do a wide variety of tasks. It’s just that we haven’t figured it out yet. The most popular research that you see in media, that’s being highlighted, is the research that gets superhuman abilities in some specific niche task.
But there’s a lot of research that we deal with day-to-day, that we read about, that is not highlighted in popular media, which tries to do one-shot learning, and smarter transfer learning and stuff. And as a field, we’re still trying to figure out how to do this properly. I don’t think, as a community of AI researchers, we’re restricting ourselves to just do the expert systems. It’s just like we haven’t figured out as well how to do more diverse systems.
Well, you said neural nets aren’t much like the human brain. Would you say just in general, mechanical intelligence is different than human intelligence? Or should one watch how children learn things, or study how people recognize what they do, and cognitive biases and all of that?
I think there is a lot of value in doing cognitive science, like looking at how child development happens, and we do that a lot. A lot of inspiration and ideas, even in machine learning and neural networks, does come from looking at such aspects of human learning and human intelligence. And it’s being done.
We collaborate, for example at FAIR—Facebook AI Research—with a few researchers who do try to understand child development and child learning. We’ve been building projects in that direction. For example, children learn things like object permanence between certain ages. If you hide something from a child and then make it reappear, does the child understand that you just put it behind your back and then just showed it to them again? Or does a child think that that object actually just disappeared and then appeared again?
So, these kinds of things are heavily-studied, and we try to understand how the mechanisms of learning are… And we’ve been trying to replicate these for neural networks as well. Can a neural network understand what object permanence is? Can a neural network understand how physics works? Children learn how physics works by playing a lot, playing with blocks, playing with various things in their environment. And we’re trying to see if neural networks can do the same.
There’s a lot of inspiration that can be taken from how humans learn. But there is slight separation between whether we should exactly replicate how neurons work in a human brain, versus neurons work in a computer thing; because human brain neurons, their learning mechanisms and their activation mechanisms are using very different chemicals, different acids and proteins.
And the fundamental building blocks in a computer are very different. You have transistors, and they work bit-wise and so on. At a fundamental block level, we shouldn’t really look for exact inspirations, but at a very high level, we should definitely look for inspiration.
You used the word ‘understand’ several times, in that “Does the computer understand?” Do computers actually understand anything? Is that maybe the problem, that they don’t actually have an experiencing self that understands?
There’s—as they say in the field—‘nobody home’, and therefore there are just going to be these limits of things that come easy to us because we have a self, and we do understand things. But all a computer can do is sense things. Is that a meaningful distinction?
We can sense things, and a computer can sense things in the sense that you have a sensor. You can consume visual inputs, audio inputs, stuff like that. But understanding can be as simple as statistical understanding. You see something very frequently, and you associate that frequency with this particular association of a term or an object. Humans have a statistical understanding of things, and they have a causal understanding of things. We have various different understanding approaches.
And machines can, at this point, with neural networks and stuff… We take a statistical or frequentist approach to things, and we can do them really well. There’s other aspects of machine learning research as well that try to do different kinds of understanding. Causal models try to consume data and see if there’s a causal relationship between two sets of variables and so on.
There’s various levels of understanding, and understanding itself is not a magical word that can be broken down. I think we can break it down into what kinds and what approaches of understanding. Machines can do certain types of understanding, and humans can do certain more types of understanding that machines can’t.
Well, I want to explore that for just a moment. You’re probably familiar with Searle’s Chinese Room thought experiment, but for the benefit of the listeners…
The philosopher [Searle] put out this way to think about that word [‘understanding’]. The setup is that there’s a man who speaks no Chinese, none at all, and he’s in this giant room full of all these very special books. And people slide questions written in Chinese under the door. He picks them up, and he has what I guess you’d call an algorithm.
He looks at the first symbol, he finds the book with that symbol on the spine, he looks up the second symbol that directs him to a third book, a fourth book, a fifth book. He works his way all the way through until he gets to the last character, and he copies down the characters for the answer. Again, he doesn’t know what they are talking about at all. He slides it back under the door. The Chinese speaker [outside] picks it up, reads it, and it’s perfect Chinese. It’s a perfect answer. It rhymes, and it’s insightful and pithy.  
The question that Searle is trying to pose is… Obviously, that’s all a computer does. It’s a deterministic system that runs these canned algorithms, that doesn’t understand whether it’s talking about cholera or coffee beans or what have you. That there really is something to understanding.  
And Weizenbaum, the man who wrote ELIZA, went so far as to say that when a computer says, “I understand,” that it is just a lie. Because not only is there nothing to understand, there’s just not even an ‘I’ there to understand. So, in what sense would you say a computer understands something?
I think the Chinese Room thing is an interesting puzzle. It’s a thought-provoking situation, rather. But I don’t know about the conclusions you can come to. Like, we’ve seen a lot of historical manuscripts and stuff that we’ve excavated from various regions of the world, and we didn’t understand that language at all. But, over time, through certain statistical techniques, or certain associations, we did understand which words—what the fundamental letters in these languages are, or what these words mean, and so on.
And no one told us exactly what these words mean, or what this language exactly implies. We definitely don’t know how those languages are actually pronounced. But we do understand them by making frequentist associations with certain words to other words, or certain words to certain symbols. And we understand what the word for a ‘man’ is in a certain historical language, or what the word for a ‘woman’ is.
With statistical techniques, you can actually understand what a certain word is, even if you don’t understand the underlying language beforehand. There is a lot of information you can gain, and you can actually understand and learn concepts by using statistical techniques.
If you look at one example in recent machine learning time… is this thing called word2vec. It’s a system, and what it does is you give it a sentence, and it replaces the center word of the sentence with a random other word from the dictionary… And it uses that sentence with this random word in the middle as a negative example, and without replacing the random word—[using] the sentence as-is—is a positive example.
Just using this simple technique, you’ll learn embeddings of words; that is, numbers associated with each word that will try to give some statistical structure to the word. With just a simple model which doesn’t understand anything about what these words mean, or in what context these words are used, you can do simple things like [ask], “Can you tell me what ‘king’, minus ‘man’, plus ‘woman’ is?”
So, when you think of ‘king’, you think, “Okay, it’s a man, a head of state.” And then you say “minus man,” so “king minus man” will try to give you a neutral character of a head of state; and then you add ‘woman’ up, and then you expect ‘queen’… And that’s exactly what the system returns, without actually understanding what each of these words specifically mean, or how they’re spelled, or what context they’re in.
So I think there is more to the story than we actually understand. That is, I think there is a certain level of understanding we can get [to] even without the prior context of knowing how things work. In the same way, computers, I think, can learn and associate certain things without knowing about the real world.
One of the common arguments is like, “Well, but computers haven’t been there and seen that, just like humans did, so they can’t actually make full associations.” That’s probably true. They can’t make full associations, but I think with partial information, they can understand certain concepts and infer certain things just with statistical and causal models that they have to learn [from].
Let me try my question a little differently, and we will get back to the here and now… But this, to me, is really germane because it speaks to how far we’re going to be able to go—in terms of using our present techniques and our present architectures, to build things that we deem to be intelligent.  
In your mind, could a computer ever feel pain? Surely, you can put a sensor on a computer that can take the temperature, and then you write a program so that when it hits 500 degrees, it should start playing this mp3 of somebody screaming in agony. But could a computer ever feel pain? Could it ever experience anything?
I don’t think so. Pain is something that’s been baked into humans. If you bake pain into computers, then yeah, maybe, but not without it evolving to learn what pain is, or like baking that in ourselves. I don’t think it will—
—But is knowing what pain is really the same thing as experiencing it? You can know everything about it, but the experience of stubbing your toe is something different than the knowledge of what pain is.
Yeah, it probably doesn’t know exactly what pain is. It just knows how to associate with certain things about pain. But, there are certain aspects of humans that a computer probably can’t exactly relate to… But a computer, at this stage of machines, has a visual sensor, has an audio sensor, has a speaker, and has a touch sensor. Now we’re getting to smell sensors.
Yes, the computer probably can experience every single thing that humans experience, in the same way; but I think that’s largely dissociative from what we need for intelligence. I think a computer can have its own specific intelligence, but not necessarily have all [other] aspects of humans covered. We’re not trying to replicate a human; we’re trying to replicate intelligence that the human has.
Do you believe that the techniques that we’re using today, the way we look at machine learning, the algorithms we use, basic architectures… How long is that going to fuel the advance of AI? Do you think the techniques we have now—if just given more data, faster computers, tweaked algorithms—we’ll eventually get to something as versatile as a human?  
Or do you think to get to an AGI or something like it, something that really can effortlessly move between domains, is going to require some completely unknown and undiscovered technology?
I think what you’re implying is: Do we need a breakthrough that we don’t know about yet, that we need AGI for?
And my honest answer is we probably do. I just don’t know what that thing looks like, because we just don’t know ahead of time, I guess. I think we are going in certain directions that we think can get us to better intelligence. Right now, where we are is that we collect a very, very large dataset, and then we throw it into a neural network model; and then it will learn something of significance.
But we are trying to reduce the amount of data the neural network needs to learn the same thing. We are trying to increase the number of tasks the same neural network can learn, and we don’t know how to do either of [those] things properly yet. Not as properly as [we do] if we want to train some dog detector by throwing large amounts of dog pictures at it.
I think through scientific process, we will get to a place where we understand better what we need. Over this process, we’ll probably have some unknown models that will come up, or some breakthroughs that will happen. And I think that is largely needed for us to get to a general AI. I definitely don’t know what the timelines are like, or what that looks like.
Talk about adversarial AI for a moment. I watched a talk you gave on the topic. Can you give us a broad overview of what the theory is, and where we are at with it?
Sure. Adversarial networks are these very simple ways of [using] neural networks that we built.
We’ve realized that one of the most common ways we have been training neural networks is: You give a neural network some data, and then you give it an expected output; and if the neural network gives an output that is slightly off from your expected output, you train the neural network to get better at this particular task. Over time, as you give it more data, and you tune it to give the correct output, the neural network gets better.
But adversarial networks are these slightly different formulations of machines, where you have two neural networks. And one neural network tries to synthesize some data. It takes in no inputs, or it takes some random noise as input, and then it tries to generate some data. And you have another neural network that takes in some data, whether it’s real data or data that is generated by this generator neural network. And this [second] neural network, its job is to discriminate between the real data and the generated data. This is called a discriminator network.
[So] you have two networks: the generator network that tries to synthesize artificial data; and you have a discriminator network that tries to tell apart the real data and the artificially-generated data. And the way these things are trained, is that the generator network gets rewards if it can fool the discriminator—if it can make the discriminator think that the data it synthesized is real. And the discriminator only gets rewards when it can accurately separate out the fake data from the real data.
There’s just a slightly different formulation in how these neural networks learn; and we call this an unsupervised learning algorithm, because they’re not really hooking onto any aspects of what the task at hand is. They just want to play this game between each other, regardless of what data is being synthesized. So that’s adversarial networks in short.
It sounds like a digital Turing test, where one computer is trying to fool the other one to think that it’s got the real data.
Yeah, you could see it that way.
Where are we at, practically speaking… because it’s kind of the hot thing right now. Has this established itself? And what kinds of problems is it good at solving? Just general unsupervised learning problems?
Adversarial networks have gotten very popular because they seem to be a promising method to do unsupervised learning. And we think unsupervised learning is one of the biggest things we need to crack before we get to more intelligent machines. That’s basically the primary reason. They are a very promising method to do unsupervised learning.
Even without an AGI, there’s a lot of fear wrapped up in people about the effects of artificial intelligence, specifically automation, on the job market.
People fall into one of three groups: There’s people who think that we’re going to enter kind of a permanent Great Depression, where there’s a substantial portion of the population that’s not able to add economic value.
And then another group says, “Well, actually that’s going to happen to all of us. Anything a human can do, we’re going to be able to build a machine to do.”
And then there are people who say, “No, we’ve had disruptive technologies come along, like electricity and machines and steam power, and it’s never bumped unemployment. People have just used these new machines to increase productivity and therefore wages.”
Of those three camps, where do you find yourself? Or is there a fourth one? What are your thoughts on that?
I think it’s a very important policy and social question on how to deal with AI. Yes, we have in the past had technology disruptions and adapted to them, but they didn’t happen just by market forces, right? You had certain policy changes and certain incentives and short-term boosts for the Depression. And you had certain parachutes that you had to give to people during these drastically-changing times.
So it’s a very, very important policy question on how to deal with the progress that AI is making, and what that means for the job market. I follow the camp of… I don’t think it will just solve itself, and there’s a big role that government and companies and experts have to play in understanding what kind of changes are coming, and how to deal with them.
Organizations like the UN could probably help with this transition, but also, there’s a lot of non-profit companies and organizations coming up who have the mission of doing AI for good, and they also have policy research going on. And I think this will play more and more of a big role, and this is very, very important to deal with—our transition into a technology world where AI becomes the norm.
So, to be clear, it sounds like you’re saying you do think that automation or AI will be substantially disruptive to the job market. Am I understanding you correctly? And that we ought to prepare for it?  
That is correct. I think, even if we have no more breakthroughs in AI as of now, like if we have literally no significant progress in AI for the next five years or ten years, we will still—just with the current AI technology that we [already] have—we will still be disrupting large domains and fields and markets—
—What do you mean, specifically? Such as?
One of the most obvious is transportation, right? We largely solved the fundamental challenges in building self-driving vehicles—
—Let me interrupt you real quickly. You just said in the next five years. I mean, clearly, you’re not going to have massive displacement in that industry in five years, because even if we get over the technological hurdle, there’s still the regulatory hurdle, there’s still retrofitting machinery. That’s twenty years of transition, isn’t it?  
Umm, what I—
—In which time, everybody will retire who’s driving a truck now, and few people will enter into the field—
—What I specifically said was that even if we have no AI breakthroughs in the next five or ten years. I’m not saying that the markets themselves will change in five years. What I specifically said and meant is that even if you have no AI research breakthroughs in five years, we will still see large markets be disrupted, regardless. We don’t need another AI breakthrough to disrupt certain markets.
I see, but don’t you take any encouragement from the past? You can say transportation, but when you look at something like the replacement of animal power with mechanical power, and if you just think of all of the technology, all of the people that displaced… Or you think of the assembly line, which is—if you think about it—a kind of AI, right?
If you’re a craftsperson who makes cars or coaches or whatever one at a time, and this new technology comes along—the assembly line—that can do it for a tenth of the price and ten times the quality. That’s incredibly disrupting. And yet, in those two instances, we didn’t have upticks in unemployment.
Yes,—
—So why would AI be different?
I think it’s just the scale of things, and the fact that we don’t understand fully how things are going to change. Yes, we can try to associate something similar in the past with something similar that’s happening right now, but I think the scale and magnitude of things is very different. You’re talking about in the past over… like over [the course of] thirty years, something has changed.
And now you’re talking about in the next ten years something will change, or something even sooner. So, the scale of things and the number of jobs that are affected, all these things are very different. It’s going to be a hard question that we have to thoroughly investigate and take proper policy change. Because of the scale of things, I don’t know if market forces will just fix things.
So, when you weigh all of the future, as you said—with the technology we have now—and you look to the future and you see, in one column, a lot of disruption in the job market; and then you see all the things that artificial intelligence can do for us, in all its various fields.
To most people, is AI therefore a good thing? Are you overall optimistic about the future with regard to this technology?
Absolutely. I think AI provides us benefits that we absolutely need as humans. There’s no doubt that the upsides are enormous. You accelerate drug discovery, you accelerate how healthcare works, you accelerate how humans transport from one place to another. The magnitude of benefits is enormous if the promises are kept, or the expectations are kept.
And dealing with the policy changes is essential. But my definite bullish view is that the upsides are so enormous that it’s totally worth it.
What would you think, in an AI world, is a good technology path to go [on], from an employment status? Because I see two things. I saw pretty compelling things that say ‘data scientist’ is a super in-demand thing right now, but that’ll be one of the first things we automate, because we can just build tools that do a lot of what that job is.
Right.
And you have people like Mark Cuban, who believes, by the way, [that] the first trillionaires will come from this technology. He said if he had it to do all over again, if he were coming up now, he would study philosophy and liberal arts, because those are the things machines won’t be able to do.
What’s your take on that? If you were getting ready to enter university right now, and you were looking for something to study, that you think would be a field that you can make a career in long-term, what would you pick?
I wouldn’t pick something based on what’s going to be hot. The way I picked my career now, and I think the way people should pick their careers is really what they’re interested in. Now if their only goal is to find a job, then maybe they should pick what Mark Cuban says.
But I also think just being a technologist of some kind, whether they try to become a scientist, or just being an expert in something technology-wise, or being a doctor… I think these things will still be helpful. I don’t know how to associate…
The question is slightly weird to me, because it’s like, “How do I make the most successful career?” And I’ve never thought about it. I’ve just thought about what do I want to do, that’s most interesting. And so I don’t have a good answer, because I’ve never thought about it deeply.
Do you enjoy science fiction? Is there anything in the science fiction world, like movies or books or TV shows, that you think represents how the future is going to turn out? You look at it and think, “Oh, yes, things could happen that way.”
I do enjoy science fiction. I don’t necessarily have specific books or movies that exactly would depict how the future looks. But I think you can take various aspects from various movies and say, “Huh, that does seem like a possibility,” but you don’t necessarily have to buy into the full story.
For example, if you look at the movie Her: You have an OS that talks to you by voice, has a personality, and evolves with its experience and all that. And that seems very reasonable to me. You probably will have voice assistance that will be smarter, and will be programmed to develop a personality and evolve with their experiences.
Now, will they go and make their own OS society? I don’t know, that seems a bit weird. In popular culture, there are various examples like this that seem like they’re definitely plausible.
Do you keep up with the OpenAI initiative, and what are your thoughts on that?
Well, OpenAI seems to be a very good research lab that does fundamental AI research, tries to make progress in the field, just like all of the others are doing. They seem to have a specific mission to be non-profit, and whatever research they do, they want to try to not tie it to a particular company. I think they’re doing good work.
I guess the traditional worry about it is that an AGI, if we built one is, is of essentially limitless value, if you can make digital copies of it. If you think about it, all value is created, in essence, by technology—by human thought and human creativity—and if you somehow capture that genie in that bottle, you can use it for great good or great harm.
I think there are people who worry that by kind of giving ninety-nine percent of the formula away to everybody, no matter how bad their intentions are, you increase the likelihood that there’ll be one bad actor who gets that last little bit and has, essentially, control of this incredibly powerful technology.  
It would be akin to the Manhattan Project being open source, except for the very last step of the bomb. I think that’s a worry some people have expressed. What do you think?
I think AI is not going to be able to be developed in isolation. We will have to get to progress in AI collectively. I don’t think it will happen in a way where you just have a bunch of people secretly trying to develop AI, and suddenly they come up with this AGI that’s eternally powerful and something that will take over humanity, or something like that.
I don’t think that fantasy—which is one of the most popular ways you see things in fiction and in movies—will happen. The way I think it will happen is: Researchers will incrementally publish progress, and at some point… It will be gradual. AI will get smarter and smarter and smarter. Not just like some extra magic bit that will make it inhumanly smart. I don’t think that will happen.
Alright. Well, if people want to keep up with you, how do they follow you personally, and that stuff that you’re working on?
I have a Twitter account. That’s how people usually follow what I’ve been up to. It’s twitter.com/soumithchintala.
Alright, I want to thank you so much for taking the time to be on the show.
Thank you, Byron.
Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here
[voices_in_ai_link_back]

Voices in AI – Episode 5: A Conversation with Daphne Koller

[voices_in_ai_byline]
In this episode, Byron and Daphne talk about consciousness, personalized medicine, and transfer learning.
[podcast_player name=”Episode 5: A Conversation with Daphne Koller” artist=”Byron Reese” album=”Voices in AI” url=”https://voicesinai.s3.amazonaws.com/2017-09-28-(00-56-17)-daphne-koller.mp3″ cover_art_url=”https://voicesinai.com/wp-content/uploads/2017/09/voices-headshot-card-4.jpg”]
[voices_in_ai_link_back]
Byron Reese: This is Voices in AI, brought to you by Gigaom. I’m Byron Reese. Today our guest is Daphne Koller. She’s the Chief Computing Officer over at Calico. She has a PhD in Computer Science from Stanford, which she must have liked a whole lot, because she shortly thereafter became a professor there for eighteen years. And it was during that time that she founded Coursera with Andrew Ng. She is the recipient of so many awards, I would do them an injustice to try to list them all. Two of them that just stick out are the Presidential Early Career Award for Scientists and Engineers, and, famously, The MacArthur Foundation Fellowship.
Welcome to the show, Daphne.
Daphne Koller: Good to be here, Byron. Thank you for inviting me.
I watched a number of your videos, and you do a really interesting thing where you open up by defining your terms often, so that everybody has, as you say, a shared vocabulary. So what is ‘artificial intelligence’ when you use that term?
Well, I think artificial intelligence is one of the harder things to define because in many ways, it’s a moving target. Things that used to be considered artificial intelligence twenty years ago are now considered so mundane that no one even thinks of them as artificial intelligence—for instance, optical character recognition.
So, there is the big lofty AI goal of general artificial intelligence, building a single agent that achieves human-level type intelligence, but I actually think that artificial intelligence should—and in many people’s minds I hope still does—encompass the very many things that five years ago would have been considered completely out of reach, and now are becoming part of our day-to-day life. For instance, the ability to type a sentence in English and have it come out in Spanish or Chinese or even Swahili.
With regard to that, there isn’t an agreed-upon definition of intelligence to begin with. So what do you think of when you think of intelligence, and secondly, in which sense is it artificial? Is it artificial like artificial turf, is it really turf, or it just pretends to be? Do you think AI is actually “intelligent,” or is it a faux imitation intelligence?
Boy, that’s a really good question, Byron. I think intelligence is a very broad spectrum that ranges from very common sense reasoning that people just take for granted, to much more specialized tasks that require what people might consider to be a deeper level of intelligence, but in many cases are actually simpler for a computer to do. I think we should have a broad umbrella of all of these as being manifestations of the phenomenon of intelligence.
In terms of it being false intelligence, no; I think what makes artificial intelligence “artificial” is that it’s humanly-constructed. That is, it didn’t organically emerge as a phenomenon, but rather we built it. Now you could question whether the new machine learning techniques are in fact organic growth, and I would say that you could make the case that if we build an architecture, that you put it in the world with the same level of intelligence as a newborn infant, and it really learns to become intelligent—maybe we shouldn’t call it artificial intelligence at that point.
But I think, arguably, the reason for the use of the word “artificial” is because it’s human-constructed as opposed to biologically-constructed.
Interestingly, McCarthy, the man who coined the phrase, later regretted it. And that actually brings to mind another question, which is: When five scientists convened at Dartmouth for the summer of 1956, to “solve the problem with artificial intelligence,” they really thought they could do it in a summer of hard work.
Because they assumed that intelligence was like, you know, in physical laws… We found just a few laws that explained all physical phenomenon, and electricity just a few, and magnetism just a few, and there was a hope that intelligence was really something quite simple. You know, iteratively-complex but had just a few overriding laws. Do we still think that? Do you think that? Is it not like that at all?
That was the day of logical AI, and I think people thought that one could reason about the world using the rules of logic, where you have a whole bunch of facts that you know—dogs are mammals; Fido is a dog, therefore Fido is a mammal—and that all you would need is to write down those facts, and the laws of logic would then take care of the rest. I think we now understand that that is just not the case, and that there is a lot of complexity both on the fact side, and then how you synthesize those facts to create broader conclusions, and how do you deal with the noise, and so on and so forth.
So I don’t think anyone thinks that it’s as simple as that. As to whether there is a single, general architecture that you can embed all of intelligence in, I think some of the people who believe that deep neural networks are the solution to the future of AI would advocate that point of view. I’m agnostic about that. I personally think that that’s probably not going to be quite there, and you’re probably going to need at least one or two other big ideas, and then a heck of a lot of learning to fine-tune parts of the model to very different use models—in the same way that our visual system is quite different from our common sense reasoning system.
And I do want to get on, in a minute, to the here and now, but just in terms of thinking through what you just said, it sounds like you don’t necessarily think that an AGI is something that we are on the way towards. You know, we can make one percent of it, and when algorithms get a little better, and computers get a little faster, and we get a little more data, we’ll evolve our way there.
It sounds like what you said is that there is some breakthrough that we need that we don’t yet have; that AGI is something very different than the, kind of, weak AI we have today. Would you agree with that?
I would agree with that. I wouldn’t necessarily agree with the fact that we are not on the right path. I think there has been a huge amount of progress in the last, well, not only in the last few years, but across the evolution of AI. But it is definitely putting us on the path there. I just think that we need additional major breakthroughs to get us there.
So with regard to the human genome, you know it’s x-number of billions of base pairs, which map to something like 700 megabytes. But most of that we share with all life, even like plants and bananas and all of that, and if you look at the part that makes us different than say a bonobo or a chimp, it may only be half of one percent.
So it may only be like three megabytes. So does that imply to you that to build an AGI, the code might be very… We are an AGI, and our intelligence is evidently built with those three megabytes of code. When working to build an AGI computer, is that useful, or is that a fair way to think about it? Or is that apples and oranges in your view?
Boy! Well, first of all, I think I would argue that a bonobo is actually quite intelligent, and a lot of the things that make us generally intelligent are shared with a bonobo. Their visual system, their ability to manipulate objects, to create tools and so on is something that certainly we share with monkeys.
Fair enough.
I think there is that piece of it. I also think that there is an awful lot of complexity that happens as part of the learning process, that we as well as monkeys and other animals go through as we encounter the world. It evolves our neural system, and so that part of it is something that emerges as well, and could be shared. So I think it’s more nuanced than that, in terms of counting the number of bits.
Right. So, we have this brain, the only AGI that we know of…  And we, of course, don’t know how our brains work. We really don’t. We can’t even model a nematode worm’s 302 neurons in a computer, let alone our hundred billion. And then we have something we call the “mind,” which is a set of capabilities that the brain manifests that don’t seem to be—with the emphasis on seem to be—derivable from neurons firing.
And then you have consciousness, which of course… Nobody purports to say they know exactly how it is that hydrogen came to name itself. So, doesn’t that suggest that you need to understand the mind, and you need to understand consciousness, in order to actually make something that is intelligent? And it will also need those things.
You know, that’s a question that artificial intelligence has struggled with a lot. What is the mind, and to what extent does that emerge from neurons firing? And if you really dive into that question, it starts to relate to the notion of soul and religion, and all sorts of things that I’m not sure I am qualified to comment on. Most people wouldn’t necessarily agree with the others’ point of view on this anyway.
I think in this respect, Turing had it right. I don’t know that you’re conscious. All I can see is your observed behavior, and if you behave as if you are conscious, I take it on faith that you are. So if we build a general artificial intelligence that acts intelligent, that is able to interact with us, understand our emotions, express things that look like disappointment or anger or frustration or joy…
I think we should give it the benefit of the doubt that it has evolved a consciousness, regardless of our ability to understand how that came about.
So tell me about your newest gig, the Chief Computing Officer at Calico. Calico, according to their website, are aiming to devise interventions that slow aging and counteract age-related diseases. What’s your mission there, within that?
I came on board to create, at Calico, what you might call a second pillar of Calico’s efforts. One pillar being the science that we’re doing here, that drives toward an understanding of the basic concepts of aging, and the ability to turn that into therapeutics for aging and age-related diseases.
But we all know that biology, like many other disciplines, is turning into data science, where we have—”we” being the community at large—developed a remarkable range of technologies that can measure all sorts of things about biological systems, from the most microscopic level, all the way up to the organismal level—interventions that allow us to perturb single genes or even single nucleotides.
And how do you take this enormity of data and really extract insights from it, is a computational question. There need to be tools developed to do this. And this is not something that biologists can learn on their own. It’s also something computer scientists can’t do on their own. It requires a true partnership between those two communities working together to make sense of the data using computational techniques, and so what I am building here at Calico is an organization within Calico that does exactly that—in partnership with our pre-existing world class biology team.
Do you think there is broad consensus, or any consensus about the bigger questions of what is possible? Like do humans need to have a natural life span? Are we going to be able to better tailor medicines to people’s genomes? What are some of those things that are, kind of, within sight?
I am very excited about the personalized medicine, precision medicine trajectory. I completely agree with you that that is on the horizon. I find it remarkably frustrating that we treat people as one-size-fits-all. And, you know, a patient walks into a doctor’s office, and there is a standard of care that was devised for a population of people that is sometimes very different from the specifics of the person… Even to the point that there are a whole bunch of treatments which were designed largely based on a cohort of men, and you have a woman coming into the doctor’s office, and it might not work for her at all. Or similarly with people of different ethnic origins.
I think that’s clearly on the horizon, and it will happen gradually over the course of the coming years. I think the ability to intelligently design medications, in a way that is geared towards achieving particular biological effects that we’re able to detect using mechanisms like CRISPR, for instance.
CRISPR, by the way, for those of you who’ve not heard of this, is a gene editing system that was developed over the last five or ten years—probably more like five—and is remarkably able to do very targeted interventions in a genome. And then one can measure the effects and say, “Oh, wait a minute, that achieved this phenotypic outcome, let’s now create a therapeutic around that.” And that therapeutic might be a drug, or it could—as we get closer to viral therapies or even gene editing—be something that actually does the exact same thing that we did in the lab, but in the context of real patients. So that’s another thing that is on the horizon, and all of this is something that requires a huge understanding of the amounts of data that are being created, and a set of machine learning artificial intelligence tools.
Now, prior to World War II, I read that we only had about five medicines. You had quinine, you had penicillin—well, you didn’t have penicillin—you had aspirin, you had morphine; and they were all, fortunately, very inexpensive.
And then Jonas Salk develops the Salk vaccine, and they ask him who owns the patent and he says, there is no patent, you can’t patent the sun. And so you know, you get the Salk vaccine, so inexpensive. Now, though, we have these issues that, you know, if you have Hepatitis-C and you need that full treatment, that’s $70,000. Are we not on a path to create ever more and more expensive medications and therapies that will create a huge gulf between the haves and the have-nots?
I think it’s a very important moral dilemma. I tend to be, rightly or wrongly, I guess, an optimist about this, in that I think some medications are really expensive because we don’t have productionized processes for creating a medication. And we certainly don’t have productionized processes, or even a template, for how to come up with a new medication for an indication that’s discovered.
But—and again, I am an optimist—as we get a better understanding of, for instance, the human genome and maybe the microbiome, and how different aspects of that and the environment come together to create both healthy and aberrant phenotypes, it will become easier to construct new drugs that are better able to cure people. And as we get better at it, I hope that costs will come down.
Now, that’s sort of a longer-term solution. In the shorter term, I think that it’s incumbent upon us as a society to help the have-nots who are sick to get access to the best medications, or at least to a certain common baseline of medications that are important to help people stay alive. I think that that’s a place where some societies do this well, and others maybe not so well. And I don’t think that’s fair.
Of course, you know, there are more and more people that hit the age of 100, but the number of supercentenarian—people who hit 110—seems stubbornly fixed. I mean, you can go to Wikipedia and read a list of all of them. And the number of people who’ve hit 125 seems to be, you know, zero. People who’ve hit 130, zero. Why is it that, although the number of centenarians goes way up—and it’s in the hundreds of thousands—the number of people who make it to 125 is zero?
That’s a topic that’s been highly-discussed very recently. There’s been a series of papers that have talked about this. I think there’s a number of hypotheses. One that I find compelling is that what causes people to die, at a certain time in history, changes over time. I mean, there was a time, not that long ago, when women’s life spans were considerably shorter than that of men, because many of them died in childbirth. So the average lifespan of a woman was relatively shorter, until we realized that we needed to sterilize the doctor’s hands when they were delivering the baby, and now it’s different.
We discovered antibiotics, which allowed us to address many of the deaths that are attributed to pathogens, though not all of them. AIDS was a killer, and then we invented retroviral therapy which allows AIDS patients to live a much longer life. So, over time, we get through additional bottlenecks that are killing people at later and later points in time. So right now, for instance, we don’t have a cure for Alzheimer’s and Parkinson’s and other forms of dementia, and that kills a lot of people.
It kills at a much later age than they would have died from in earlier cases, at earlier times in history. But I hope that at some point in the next twenty years, someone will discover a cure for Alzheimer’s, and then people will be able to live longer. So I think over time, we solve the thing that kills you next, and that allows the thing that’s next down the line to kill you next, and then we go ahead and try and cure that one.
You know, when you look at the task before you, if you are trying to do machine learning to help people live longer and healthier lives, it’s got to be frustrating that, like, all the data must be bad, because symptoms generally aren’t recorded in a consistent way. You don’t have a control, like for example twins who, five minutes into the world go down different paths.
Everybody has different genomes. Everybody eats different food, breathes different air. How much of a hurdle is that to us being able to do really good machine learning on things like nutrition, which seems, you know… We don’t even know if eggs are good for you or bad for you and those sorts of things.
It’s a huge hurdle, and I think it was one of the big obstacles to the advancement of machine learning in other domains, up until relatively recently, when people were able to acquire enough data to get around that. If you look at the earlier days of, for instance, computer vision, the data sets were tiny—and that’s not that long ago, we’re talking about less than a decade.
You had data sets with a few hundred, and a few thousand images was considered large, and you couldn’t do much machine learning on that because when you think about the variation of a standard category… Like, a wedding banquet that ranges from photos of a roomful of people milling around to someone cutting a wedding cake.
And so the variability there is extremely large, and if all you have is twenty images of a wedding banquet, you’re not going to get very far training on that. Now, the data is still as noisy—and arguably even noisier when you download it from Google Images or Flickr or such—but there’s enough of it that you get to explore a sufficient part of the space for a machine learning algorithm. So that you can, not counteract the noise, but simply accommodate it as a variability in your models.
If we get enough data on the medical side, the hope is that we’ll be able to get to a similar place where, yes, the variability will remain, but if you have enough of the ethnic diversity, and enough of the people’s lifestyle, and so on, all represented in your data set, then that will allow us to address the variability. But that requires a major data collection effort, and I think we have not done a very good job as a society of making that a priority to collect, consolidate, and to some extent clean medical data so that we can learn from it.
The UK, for instance, has a project that I think is really exciting. It’s the UK Biobank project. It’s 500,000 people that were genotyped, densely-phenotyped, and their records are tied to the UK National Health Service; so you have ongoing outcome data for them. It’s still not perfect. It doesn’t tell you what they eat every day, but they asked them that in the initial survey, so you get at least some visibility into that. I think it’s an incredibly exciting project, and we should have more of those.
They don’t necessarily have to use the exact same technique that the UK Biobank is using, but if we have medical data for millions of people, we will be able to learn a lot more. Now we all understand there are serious privacy issues there, and we have to be really thoughtful about how to do this.
But if you talk to your average patient, especially ones who are suffering from a devastating illness, you will find that many of them are eager to share some information about their medical condition to the benefit of science, so that we can learn how to treat their disease better. Even if it doesn’t benefit them, because it might be too late, it will benefit others.
So you just mentioned object recognition, and of course humans do that so well. I could show you a photograph of a little Tiki statue, or a raven, or something… And then you could instantly go through a bunch of photos and recognize it if it’s underwater, or if it’s dark, or if it’s inside, and all of that. And I guess it’s through transferred learning of some kind. How far along are we… Do we know how to do it, and we just don’t have the horsepower to do it, or do we not really even understand how that works yet?
Well, I think it’s not that there is one way to do this. There’s a number of techniques that have been developed for transfer learning, and I agree with you that transfer learning is hugely important. But right now, if you look at models—like the Computer Vision Inception Network that Google has developed, there is a whole set of layers in that neural network that were devised based on a large category of web images that have a broad range of categories. But that same set of layers is now taken, pre-trained, and with a relatively small amount of training data—sometimes even as little as zero training examples—can be used for applications that it was never intended for, like the retinopathy project, for instance, that they recently published. I think that’s happening.
Another example, also from Google, is in the machine translation realm, where they recently showed that you could use a network architecture to translate between two languages for which you didn’t have any examples of those two languages together. The machine was effectively creating an interlingua on its own, so that you’re translating a sentence in Thai into this interlingua and then producing a sentence in Swahili as an output, and you’ve never seen a pair of sentences and Thai in Swahili together. So I think we’re already seeing examples of transfer learning emerging in the context of specific domains and I think it’s incredibly exciting.
You mentioned CRISPR/Cas9 a few minutes ago. And of course it comes with the possibility of actually changing genes in a way that that alters the line, right, where the children and the grandchildren have this new altered gene state. There is no legislative or ruling body that has any authority over any of that? CRISPR is cheap, and so can anybody do that?
I agree with you. I think there’s a very serious set of ethical questions there that we need to start thinking about seriously. So, in some ways, when people say to me, “Oh, we need to come up with legislation regarding the future of AI and the ethical treatment of artificial intelligence agents,” I tell them we have a good long time to think about that. I am not saying we shouldn’t think about it, but it’s not like it’s a burning question.
I think this is a much more burning question, and it comes up with editing the human genome, and I think it comes up at least as much in how do we prevent threats like someone recreating smallpox. That’s not CRISPR, that’s DNA synthesis, which also is a technology that’s here. So I think that’s a set of serious questions that the government ought to be thinking about, and I know that there is some movement towards that, but I think we’re behind the curve there.
Behind the curve in terms of we need to catch up?
Yeah, technology has overtaken our thinking about the legal and ethical aspects of this.
CRISPR would let you do transgenesis on a human. You could take a gene from something that glows in the dark, and make a human that glows in the dark, in theory. I mean, we are undoubtedly on the road to being able to use those technologies in horrific ways, very inexpensively. And it’s just hard to think, like, even if one nation can create legislation for it, it doesn’t mean that it couldn’t be done by somebody else. Is it an intractable problem?
I think all technology can be used for good or evil, or most technology can be used for good or evil. And we have successfully—largely successfully—navigated threats that are also quite significant, like the threat of a nuclear holocaust. Nuclear technology is another one of those examples that, it can be used for good, it has been used for good, it can also be used to great harm. We have not yet, fortunately, had a dirty bomb blow up in Manhattan, making all of Manhattan radioactive, and I am hopeful that will never happen.
So I am not telling you I have the solution to this, but I think that as a society, we should figure out what is morally permissible, and what is not, and then really try and put in guardrails both in terms of social norms, as well as in terms of legal and enforcement questions to try and prevent nations or individuals from doing things that we would consider to be horrific.
And I am not sure we have consensus as a society on what would be horrific. Is it horrific to genetically engineer a child that has a mutation that’s going to make their life untenable, or cut short after a matter of months, and make them better? I would say a lot of people would think that’s totally fine; I think that’s totally fine. Is it as permissible to make your child have superhuman vision, great muscle strength, stamina and so on? I think that’s in the gray zone. Is it permissible to make your child glow in the dark? Yeah, that’s getting beyond the pale, right? But those are discussions that we are not really having as a society, and we should be.
Yeah, and the tricky thing is, there is not agreement on whether you should use transgenesis on seeds, you know? You put Vitamin A in rice, and you can end Vitamin A deficiency, or diminish it, and we don’t seem to be able to get agreement on whether you should even do that.
Yeah. You know I find people’s attitudes here to be somewhat irrational in the sense that we’ve been doing genetic engineering on plants for a very long time, we’ve just been doing it the hard way. Most of the food that we eat comes from species of plants that don’t naturally grow in the wild. They have been very carefully bred to have specific characteristics in terms of resistance to certain kinds of pests, and growing in conditions that require hardier plants, and so on and so forth.
So even genetically engineering plants by very carefully interbreeding them, and doing various other things to create the kinds of food that, for whatever reason, we prefer—tomatoes that don’t spoil when you ship them in the bowels of a ship for three weeks—the fact that we are now doing it more easily doesn’t make it worse. In fact, you could argue that it might make it more targeted, and have fewer side effects.
I think when it comes to engineering other things, it becomes much more problematic, and you really need to think through the consequences of genetic engineering on a human, or genetic engineering on a bug.
Yeah, when x-rays came out, they would take a bunch of seeds and they would irradiate them, and then they would plant them, and very few would grow, but a few would grow poorly, and every now and then you would get some improvement, and that was the technique for much of the produce we eat today.
Indeed, and you don’t know what the radiation did, beyond the stuff that we can observe phenotypically, as in it grows better. So all of these things that are happening to all those other genes went unobserved and unmeasured. Now you are doing a much more precision intervention, in just changing the one gene that you care about. And for whatever reason some people view that as being inferior, and I think that’s a little bit of a misunderstanding of what exactly happened before, and is happening now.
It used to be that the phrase “cure aging” was looked at nonsensically. Is that something that is a valid concept, that we may be able to do?
So we do not use the term “cure aging” at Calico. What we view ourselves as doing is increasing healthspan, which is the amount of time that you live as a healthy, happy, productive human being. I think that we as a society have been increasing healthspan for a very long time. I’ve talked about a couple of examples in the past.
I don’t think that we are on the path to letting people live forever. Some people might think that that’s an achievable goal, but I think it’s definitely a worthy goal to make it so that you live healthy longer, and you don’t have people who spend twenty years of their lives in a nursing home being cared for by others because they are unable to care for themselves.
I think that’s a very important goal for us as a society, both for the people themselves, for their families, but also in terms of the cost that we incur as a society in supporting that level of care.
Well, obviously you’ve had a great impact, you know, presumably in two ways: One, with what you’ve done to promote education, and democratizing that, and then what you are doing in health. What are your goals? What do you hope to accomplish in the field? How do you want to be remembered as?
So, let’s see. I think there’s a number of answers that I could give to that question at different levels. At one level, I would like to be—and not the only one, by any stretch, because there is a whole community of us working here—one of the people that really brought together two fields that it’s critical that we bring together: the field of machine learning and the field of biology, and really turning biology into a data science.
I think that’s a hugely important thing because it is not possible, even today and certainly going forward, to make sense of the data that is being accumulated using simple, statistical methods. You really need to build much deeper models.
Another level of answer is that I would like to do something that made a difference to the lives of individual people. One of the things that I really loved about the work that we did at Coursera was that daily deluge, if you will, of learner stories. Of people who say, “My life has been transformed by the access to education that I would never have had before, and by doing that I am now employed and can feed my children and I was not able to do that before,” for instance.
And so if I can help us get to the point where I get an email from someone who says, “I had a genetic disposition that would have made me die of Alzheimer’s at an early age, but you were able to help create technology that allowed me to avoid that.” To me that would be incredibly fulfilling. Now, that is a very aspirational goal, and I’m not assuming that it’s necessarily achievable by me—and, even if it’s achievable, will definitely involve the work of many others—but that, I think, is what we should aspire to, what I aspire to.
You know, you mentioned the importance of merging machine learning with these other fields, and Pedro Domingos, who actually was on the show not long ago, wrote a book called The Master Algorithm where he proposes that there must exist a master algorithm that can solve all different kinds of problems, that unite the symbolists and the Bayesians and all of the different, what he calls, tribes. Do you think that such a thing is likely to exist? Do you think that neural nets may be that, kind of a one-size-fits-all solution to problems?
I think neural nets are very powerful technology, and they certainly help address, to a certain extent, a very large bottleneck, which is how do you construct a meaningful set of features in domains where it’s really hard for people to extract those, and solve problems really well. I think their development, especially over the last few years, when combined with large data, and the power of really high-end computing, has been transformative to the field.
Do I think they are the universal architecture? Not as of now. I think that there is going to be—and we discussed this earlier—at least one or two big things that would need to be added on top of that. I wish I knew what they were, but I don’t think we are quite there yet.
So you are walking on a beach, and you find a lamp. You rub the lamp, and out pops a genie, and the genie says: “I will give you one of the following three things: new cunning and brilliant algorithms that solve all kinds of problems in more efficient ways, an enormous amount of data that’s clean and accurate and structured, or computers that are vastly faster, way beyond the speed of what we have now.” What would you choose?
Data. I would choose data.
It sounded like, when I set that question up earlier about, “Oh, data, it’s so hard,” you were like, “Tell me about it.” So that is the daily challenge, because I know my doctor still keeps everything in those manila folders that have three letters of my last name, and I think, “Wow, that’s it? That’s what’s going to be driving the future?” So that is your bottleneck?
I think it really is the bottleneck, and it’s not even just a matter of, you know, digitizing the records that are there—which, by the way, it’s not just a matter of they are being kept in manila folders. It’s also a matter of the extent to which different doctors write things in different ways, and some of them don’t write things at all and just leave it to memory, and so on.
But I think even beyond that, there is all the stuff that’s not currently being measured. I think we’re starting to see some glimmers of light in certain ways; for instance, I’m excited by the use of wearable devices to measure things like people’s walking pace and activity and so on. I think that provides us with a really interesting window on daily activity, whereas, otherwise people see the doctor once a year or once every five years, sometimes—and that really doesn’t give us a lot of visibility into what’s going on with their lives the rest of the time.
I think there is a path forward on the data collection, but if you gave me a really beautiful large clean data set that had, you know, genetics and phenotypes and molecular biomarkers, like gene expression and proteomics and so on and so forth… I am not saying I have the algorithms today that can allow me to make sense of all of that but, boy, there is a lot that we can do with that, even today. And it would spur the development of really amazing creative algorithms.
I think we don’t lack creativity in algorithms. There is a lot that would need to happen, but I think we’re, in many cases, stymied by the lack of availability in data as well as just the amount of time and effort in terms of grunge work that’s required to clean what’s there.
So there is a lot of fear wrapped up in some people about artificial intelligence. And just to set the question up, specifically about employment, there’s three views about its effect: There’s one group of people who think we are going to enter into something like a permanent Great Depression, where there are people who don’t have the skills to compete against machines. Then there are those who believe that there’s nothing a machine can’t do eventually, and once they can learn to do things faster than we can, they’ll take every job. Then there is a third camp of people who say, look, every time we’ve had disruptive technologies, even electricity and steam power and machines, people just use those to increase their productivity and that’s how we got a rising standard of living. Which of those three camps—or a fourth one—do you identify with?
I probably would place myself—and again I tend to be an optimist, so factor that in—probably more in the third camp. Which is to say, each time that we’ve had a revolution, it has spurred productivity and people migrated from one job category into another job category that basically moves them in some ways, in many cases, further up the food chain.
So I would hope that that would be the case here; our standard of living will go up, and people will do jobs that are different. I do see the case of people saying that this revolution is different, because, over time, a larger and larger fraction of jobs will disappear and the number of jobs that are left will diminish. That is, you just won’t need that many people to do stuff.
Now, again from the optimist’s perspective, if we really have machines that do everything—from grow crops, to package them and put them in supermarkets, and so on, and basically take care of all of the day-to-day stuff that we need to exist—arguably you could imagine that a lot of us will live a life of partial leisure. And that will allow us to, at least, exist, and have food and water, and some level of healthcare and education, and so on, without having to work, and we will spend our time being creative and being artisans again or something.
Which of those is going to be the case, I think is an interesting question, and I don’t have a firm opinion on that.
So, I followed with a lot of interest Watson, when they took the cancer cases and the treatment that oncologists gave, and then Watson was able to match them ninety-some odd percent of the time, and even offered new ones because it read all of these journals and so forth.
So that’s a case of using artificial intelligence for treatment, but is treatment really fundamentally a much easier problem to solve than diagnosis? Because diagnosis is—you know, my eyes water when I eat potato chips—not very structured data.
I think that if you look back, even in the mid-’90s, which is a long way back now, there were diagnostic models that were actually pretty darn good. People moved away from that, partly because to really scale those out and make those robust, you just needed a lot more data, and also I think there are societal obstacles to the adoption of fully-automated diagnoses.
I think that’s actually an even more fundamental problem, is the extent to which doctors, patients, and insurance companies are willing to take a diagnosis that’s provided by a computer. I don’t think fundamentally, from a technological perspective, that is an unsolvable problem.
So is diagnosis a case for an expert system? I think that’s what you are alluding to—you know, how do you tell the difference between a cold and the flu? Well, do they have a fever; do they have aches and pains?
Is that a set of problems where you would use relatively older technologies to build all that out? And even if we don’t switch to that, being able to have access to just that knowledge base, in some parts of the world, is a huge step forward.
I would agree. And by the way, the thing I was thinking back on is not the earliest version of expert systems, which were all rule-based; but rather the ones that came later, which used a probabilistic model that really incorporated things like the chances of a certain thing manifesting in a somewhat different way, and if you have this predisposing factor, or, like, if you visited a country that has SARS recently, then maybe that changes the probability that what you have is not the cold or the flu but rather something worse.
And so all that needs to be built into the model. And the probabilistic models really did accommodate that, and are easily… In fact, there is a lot of technology that’s already in place for how to incorporate machine learning so that you can make those models better and better over time.
I think that’s an area that one could easily go back to, and construct technology that would be hugely impactful, especially in parts of the world where they lack access to good medical care because there just aren’t enough doctors per capita, or the doctors are all concentrated in big cities. And you have people who are living in some rural village and can’t get to see a doctor.
I agree with you that there is a huge possibility there. I think there is also a huge possibility in treatment of chronic care patients, because those are ones that consume a huge fraction of the resources of a doctor’s time, and there just aren’t enough hours in the day for a doctor to see people as frequently as might be beneficial for keeping track of whether they are deteriorating.
So maybe by the time they come and see the doctor six months later, their condition has already deteriorated, and maybe if it had been caught earlier we could have slowed that down by changing treatment. So I think there are a lot of opportunities to apply a combination of modeling and the machine learning, in medical care, that will really help make people’s lives better.
We’re almost out of time, so I have just two more questions for you. First, what is something that looks, for you, like the kind of problem in health that machine learning is going to be able to solve soon? What’s a breakthrough we can hope to pick up the newspaper and read about in five years, something really potentially big that is within our grasp, but just a little out of our reach?
I think there are a couple of areas that I see emerging which are already happening, and you’re starting to see that. Cancer—I think we talked earlier about the bottlenecks that are being addressed one after the other. And, you know, we have antibiotics and retrovirals and statins; and I think we are starting to see with areas like immuno-oncology, for instance, some actual cures for metastatic cancer which, by and large, is incurable using standard methods, with few exceptions. And I think that’s a big area where I think it’s really exciting.
I am seeing some really interesting developments on things that are in the context of specific diseases, that are more genetically-oriented therapies—be it CRISPR, be it viral therapies. We are seeing some others on the path to being approved in the next few years, and so I think that’s a place where, again, on the therapeutic side, there is a big opportunity.
I think the third one is the use of computers in the context of image-based diagnosis, and that’s an area that I used to work in when I was at Stanford—where you show an image of a tumor biopsy sample, or a radiology image, or a 3D Cat Scan of a patient, and they’re able to discover things that are not visible to a physician. Or maybe only visible to a small subset of truly expert physicians, but in most cases, you’re not going to be lucky enough to be the one that they look at.
So I think that’s an area where we will also see big advancements. These are just three off of the top of my head in the medical space, but I am sure there are others.
And a final question: You seem to be doing a whole lot of things. How do people keep up with you, what’s your social media of choice and so forth?
Boy, I am not much of a social media person, maybe because I am doing so many other things. So I think most of my visibility happens through scientific publications. As we develop new ideas, we subject them to peer review, and when we are confident that we have something to say, that’s when we say it.
Which I think is important because there is so much out there, and I think people rush to talk about stuff that’s half baked, not well-vetted… There is a lot of, unfortunately, somewhat bogus science out there—not to mention bogus news. And I think if we had less stuff, that was higher-quality—and we were not flooded with stuff of dubious correctness through which we had to sift—I think we would all be better off.
All righty. Well thank you so much for taking the time. It was a fascinating hour.
Thank you very much Byron. It was a pleasure for me too. Thank you.
Byron explores issues around artificial intelligence and conscious computers in his upcoming book The Fourth Age, to be published in April by Atria, an imprint of Simon & Schuster. Pre-order a copy here
[voices_in_ai_link_back]