Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
A verbatim transcript of a Lex Fridman interview with Yoshua Bengio discussing core limitations and future directions for deep learning. Bengio highlights the mystery of biological credit assignment over long timescales, the need for training objectives that induce causal/world models and joint language–vision learning, and the importance of disentangled, compositional representations to improve generalization and avoid catastrophic forgetting. He also addresses AI safety and societal risks (short/medium-term harms like surveillance, discrimination, and autonomous weapons), the role of reinforcement/agent learning and generative models, and ideas about machine teaching and human–machine interaction.
Chapters
What difference between biological neural networks and artificial neural networks is most mysterious, captivating, and profound for you?
First of all, there's so much we don't know about biological neural networks, and that's very mysterious and captivating because maybe it holds the key to improving artificial neural networks. One of the things I studied recently, something that we don't know how biological neural networks do, but would be really useful for artificial ones, is the ability to do credit assignment through very long time spans. There are things that we can in principle do with artificial neural nets, but it's not very convenient and it's not biologically plausible. And this mismatch, I think this kind of mismatch may be an interesting thing to study to A, understand better how brains might do these things because we don't have good corresponding theories with artificial neural nets, and B, maybe provide new ideas that we could explore about things that brain do differently and that we could incorporate in artificial neural nets.
So let's break credit assignment up a little bit. It's a beautifully technical term, but it could incorporate so many things. So is it more on the RNN memory side that thinking like that, or is it something about knowledge building up common sense knowledge over time, or is it more in the reinforcement learning sense that you're picking up rewards over time for a particular to achieve a certain kind of goal?
So I was thinking more about the first two meanings whereby we store all kinds of memories, episodic memories in our brain, which we can access later in order to help us both infer causes of things that we are observing now and assign credit to decisions or interpretations we came up with a while ago when those memories were stored. And then we can change the way we would have reacted or interpreted things in the past. And now that's created assignment used for learning.
So in which way do you think artificial neural networks, the current lstm, the current architectures, are not able to capture the. Presumably you're thinking of very long term.
Yes. So current nets are doing a fairly good job for sequences with dozens or say hundreds of time steps. And then it gets sort of harder and harder and depending on what you have to remember and so on, as you consider longer durations. Whereas humans seem to be able to do credit assignment through essentially arbitrary times. Like I could remember something I did last year, and then now, because I see some new evidence, I'm going to change my mind about the way I was thinking last year and hopefully not do the same mistake again.
I think a big part of that is probably forgetting you're only remembering the really important things. It's very efficient forgetting.
Yes. So there's a selection of what we remember and I think there are really cool connection to higher level cognition here regarding consciousness deciding and emotions like sort of deciding what comes to consciousness and what gets stored in memory, which, which are not trivial either.
So you've been at the forefront there all along, showing some of the amazing things that neural networks, deep neural networks can do in the field of artificial intelligence is just broadly in all kinds of applications, but we can talk about that forever. But what in your view, because we're thinking towards the future, is the weakest aspect of the way deep neural networks represent the world? What is in your view is missing?
So current state of the art neural nets trained on large quantities of images or texts have some level of understanding of what explains those data sets. But it's very basic, it's very low level, and it's not nearly as robust and abstract and general as our understanding. Okay, so that doesn't tell us how to fix things. But I think it encourages us to think about how we can maybe train our neural nets differently so that they would focus, for example, on causal explanation, something that we don't do currently with neural net training. Also, one thing I'll talk about in my talk this afternoon is instead of learning separately from images and videos on one hand and from text on the other hand, we need to do a better job of jointly learning about language and about the world to which it refers so that, you know, both sides can help each other. We need to have good world models in our neural nets for them to really understand sentences which talk about what's going on in the world. And I think we need language input to help provide clues about what high level concepts, like semantic concepts, should be represented at the top levels of these neural nets. In fact, there is evidence that the people purely unsupervised learning of representations doesn't give rise to high level representations that are as powerful as the ones we're getting from supervised learning. And so the clues we're getting just with the labels, not even sentences, is already very powerful.
Do you think that's an architecture challenge or is it a data set challenge?
Neither.
I'm tempted to just end it there. Can you elaborate slightly?
Of course, data sets and architectures are something you want to always play with. But I think the crucial thing is more the training objectives, the training frameworks, for example, going from passive observation of data to more active agents which learn by intervening in the world, the relationships between causes and effects the sort of objective functions which could be important to allow the highest level explanations to rise from the learning, which I don't think we have now, the kinds of objective functions which could be used to reward exploration, the right kind of exploration. So these kinds of questions are neither in the data set nor in the architecture, but more in how we learn under what objectives, and so on.
Yeah, I've heard you mention in several contexts the idea of sort of the way children learn, they interact with objects in the world. And it seems fascinating because in some sense, except with some cases in reinforcement learning, that idea is not part of the learning process. In artificial neural networks. It's almost like, do you envision something like an objective function saying, you know what, if you poke this object in this kind of way, it would be really helpful for me to further learn. Right. Sort of almost guiding some aspect of learning.
So I was talking to Rebecca Sachs just an hour ago and she was talking about lots and lots of evidence from infants seem to clearly pick what interests them in a directed way. And so they're not passive learners. They focus their attention on aspects of the world which are most interesting, surprising, in a non trivial way that makes them change their theories of the world.
So that's a fascinating view of the future progress. But on a more maybe boring question, do you think going deeper, do you think just increasing the size of the things that have been increasing a lot in the past few years will, will also make significant progress? So some of the representational issues that you, you mentioned, they're kind of shallow in some sense.
Oh, shallow.
Prior.
In the sense of abstraction.
In the sense of abstraction. They're not getting some.
I don't think that having more depth in the network in the sense of instead of 100 layers we have 10,000 is going to solve our problem.
You don't think so? Is that obvious to you?
Yes. What is clear to me is that engineers and companies and labs grad students will continue to tune architectures and explore all kinds of tweaks to make the current state of the art slightly, ever slightly better. But I don't think that's going to be nearly enough. I think we need some fairly drastic changes in the way that we are considering learning to achieve the goal that these learners actually understand in a deep way the environment in which they are, you know, observing and acting.
But I guess I was trying to ask a question that's more interesting than just more layers is basically, once you figure out a way to learn through interacting, how many parameters does it take to store that information. So I think our brain is quite bigger than most neural networks.
Right, right. Oh, I see what you mean. Oh, I'm with you there. So I agree that in order to build neural nets with the kind of broad knowledge of the world that typical adult humans have, probably the kind of computing power we have now is going to be insufficient. So, well, the good news is there are hardware companies building neural net chips and so it's going to get better. However, the good news, in a way which is also a bad news, is that even our state of the art deep learning methods fail to learn models that understand even very simple environments, like some grid worlds that we have built, even these fairly simple environments. I mean, of course if you train them with enough examples, eventually they get it. But it's just like instead of what humans might need, just dozens of examples, these things will need millions for very, very, very simple tasks. And so I think there's an opportunity for academics who don't have the kind of computing power that say Google has to do really important and exciting research to advance the state of the art in training frameworks, learning models, agent learning, in even simple environments that are synthetic, that seem trivial, but yet current machine learning fails on.
We talked about priors and common sense knowledge. It seems like we humans take a lot of knowledge for granted. So what's your view of these priors of forming this broad view of the world, this accumulation of information and how we can teach neural networks or learning systems to pick that knowledge up? So knowledge, you know, for a while the artificial intelligence was maybe in the 80s, like there's a time where knowledge representation, knowledge acquisition, expert systems, I mean symbolic AI was a view, was an interesting problem set to solve and it was kind of put on hold a little bit, it seems like, because it doesn't work. It doesn't work. That's right, but.
That's right, but the goals of that remain important.
Yes, remain important. And how do you think those goals can be addressed?
Right. So first of all, I believe that one reason why the classical expert systems approach failed is because a lot of the knowledge we have. So you talked about common sense intuition. There's a lot of knowledge like this which is not consciously accessible. There are lots of decisions we're taking that we can't really explain, even if sometimes we make up a story. And that knowledge is also necessary for machines to take good decisions. And that knowledge is hard to codify in expert systems, rule based systems and, you know, classical AI formalism. And there are other issues, of course, with the old AI, like not really good ways of handling uncertainty. I would say something more subtle, which we understand better now, but I think still isn't enough in the minds of people. There's something really powerful that comes from distributed representations. The thing that really makes neural nets work so well. And it's hard to replicate that kind of power in a symbolic world. The knowledge in expert systems and so on is nicely decomposed into like a bunch of rules. Whereas if you think about a neural net, it's the opposite. You have this big blob of parameters which work intensely together to represent everything the network knows and it's not sufficiently factorized. And so I think this is one of the weaknesses of current neural nets that we have to take lessons from classical AI in order to bring in another kind of compositionality which is common in language, for example, and in these rules, but that isn't so native to neural nets.
And on that line of thinking, disentangled representations.
Yes. So let me connect with disentangled representations, if you don't mind. So for many years I've thought, and I still believe that it's really important that we come up with learning algorithms, either unsupervised or supervised but, or reinforcement, whatever, that build representations in which the important factors, hopefully causal factors, are nicely separated and easy to pick up from the representation. So that's the idea of disentangled representations. It says transform the data into a space where everything becomes easy. We can maybe just learn with linear models about the things we care about. And I still think this is important, but I think this is missing out on a very important ingredient which classical AI systems can remind us of. So let's say we have these design technical representations. You still need to learn about the relationships between the variables, those high level semantic variables. They're not going to be independent. I mean, this is like too much of an assumption. They're going to have some interesting relationships that allow to predict things in the future to explain what happened in the past. The kind of knowledge about those relationships in a classical AI system is encoded in the rules. Like a rule is just like a little piece of knowledge that says, oh, I have these 2, 3, 4 variables that are linked in this interesting way, then I can say something about one or two of them given a couple of others. Right? In addition to disentangling the elements of the representation, which are like the variables in rule based system, you, you also need to disentangle the mechanisms that relate those variables to each other. So like the rules, so the rules are neatly separated. Like each rule is living on its own. And when I change a rule because I'm learning it doesn't need to break other rules. Whereas current neural nets, for example, are very sensitive to what's called catastrophic forgetting. Where after I've learned some things and then I learn new things, they can destroy the old things that I had learned. If the knowledge was better factorized and separated, disentangled, then you would avoid a lot of that. Now you can't do this in the sensory domain, but my idea, like in pixel space. But my idea is that when you project the data in the right semantic space, it becomes possible to now represent this extra knowledge beyond the transformation from inputs to representations, which is how representations act on each other and predict the future and so on in a way that can be neatly disentangled. So now it's the rules that are disentangled from each other and not just the variables that are disentangled from each other.
And you draw a distinction between semantic space and pixels. Does there need to be an architectural difference?
Well, yeah. So there's the sensory space, like pixels, where everything is entangled. The information, like the variables are completely interdependent in very complicated ways. And also computation, like, it's not just the variables, it's also how they are related to each other is all intertwined. But I'm hypothesizing that in the right high level representation space, both the variables and how they relate to each other can be disentangled. And that will provide a lot of generalization power.
Generalization power, yes. Distribution of the test set is assumed to be the same as a distribution of the training set.
Right. This is where current machine learning is too weak. It doesn't tell us anything. It's not able to tell us anything about how our neural nets, say, are going to generalize to a new distribution. And people may think, well, but there's nothing we can say if we don't know what the new distribution will be. The truth is humans are able to generalize to new distributions.
How are we able to do that?
Because there is something. These new distributions, even though they could look very different from the training distributions, they have things in common. So let me give you a concrete example. You read a science fiction novel, the science fiction novel maybe brings you in some other planet where things look very different on the surface, but it's still the same laws of physics. And so you can read the book and you understand what's going on. So the distribution is very different. But because you can transport a Lot of the knowledge you had from Earth about the underlying cause and effect, relationships and physical mechanisms and all that, and maybe even social interactions. You can now make sense of what is going on on this planet where like visually, for example, things are totally different.
Taking that analogy further and distorting it, let's enter a science fiction world of say, Space Odyssey 2001 with HAL.
Yeah.
Or maybe. Which is probably one of my favorite AI movies. And then me too. And then there's another one that a lot of people love that maybe a little bit outside the AI community is Ex Machina. I don't know if you've seen it.
Yes, yes.
By the way, what are your views on that movie? Are you able to include it?
So there are things I like and things I hate.
So you could talk about that in the context of a question I want to ask, which is there's quite a large community of people from different backgrounds, often outside of AI, who are concerned about existential threat of artificial intelligence. You've seen this community develop over time. You've seen, you have a perspective. So what do you think is the best way to talk about AI safety, to think about it, to have discourse about it within AI community and outside and grounded in the fact that Ex Machina is one of the main sources of information for the general public about AI.
So I think you're putting it right. There's a big difference between the sort of discussion we ought to have within the AI community and the sort of discussion that really matter in the general public. So I think the picture of Terminator and AI loose and killing people and super intelligence, that's going to destroy us whatever we try isn't really so useful for the public discussion. Because for the public discussion, the things I believe really matter are the short term and medium term very likely negative impacts of AI on society. Whether it's from security, like Big Brother scenarios with face recognition or killer robots, or the impact on the job market or concentration of power and discrimination. All kinds of social issues which could actually some of them could really threaten democracy.
For example, just to clarify, when you said killer robots, you mean autonomous weapon, like the weapon systems?
Yes. I don't know. That's right. So I think these short and medium term concerns. Concerns should be important parts of the public debate. Now, existential risk for me is a very unlikely consideration, but still worth academic investigation in the same way that you could say, should we study what could happen if meteorite came to Earth and destroyed it? So I think it's very unlikely that this is going to happen or happen in a reasonable future. It's very. The sort of scenario of an AI getting loose goes against my understanding of at least current machine learning and current neural nets and so on. It's not plausible to me. But of course I don't have a crystal ball and who knows what AI will be in 50 years from now. So I think it is worth that scientists study those problems. It's just not a pressing question as far as I'm concerned.
So before I continue down that line, I have a few questions there. But what, what, what, what do you like and not like about Ex Machina as a movie? Because I. I actually watched it for the second time and enjoyed it. I hated it the first time and I enjoyed it quite a bit more the second time when I sort of learned to accept certain pieces of it. See it as a concept movie. How. What was your experience? What were your thoughts?
So the negative is the picture it paints of science is totally wrong. Science in general and AI in particular. Science is not happening in some hidden place by some really smart guy, one person, one person. This is totally unrealistic. This is not how it happens. Even a team of people in some isolated place will not make it. Science moves by small steps, thanks to the collaboration and community of a large number of people interacting. And all the scientists who are expert in their field kind of know what is going on, even in the industrial labs, information flows and leaks and so on. And the spirit of it is very different from the way science is painted in this movie.
Yeah, let me ask on that point. It's been the case to this point that kind of. Even if the research happens inside Google or Facebook, inside companies, it still kind of comes out, ideas come out.
Absolutely.
Do you think that will always be the case with AI? Is it possible to bottle ideas to the point where there's a set of breakthroughs that go completely undiscovered by the general research community? Do you think that's even possible?
It's possible, but it's unlikely.
Unlikely.
It's not how it is done now. It's not how I can foresee it in. In the foreseeable future. But of course I don't have a crystal ball. And so who knows? This is science fiction after all, but. But usually science.
It sounds ominous that the lights went off during. During that discussion.
So the problem, again, one thing is the movie and you could imagine all kinds of science fiction. The problem for me, maybe similar to the question about existential risk, is that this kind of movie paints such a Wrong picture of what is actual, the actual science and how it's going on, that it can have unfortunate effects on people's understanding of current science. And so that's kind of sad. There's an important principle in research which is diversity. So in other words, research is exploration. Research is exploration in the space of ideas. And different people will focus on different directions. And this is not just good, it's essential. So I'm totally fine with people exploring directions that are contrary to mine or look orthogonal to mine. More than fine. I think it's important I and my friends don't claim we have universal truth about what will, especially about what will happen in the future. Now, that being said, we have our intuitions and then we act accordingly according to where we think we can be most useful and where society has the most to gain or to lose. We should have those debates and not end up in a society where there's only one voice and one way of thinking and research money is spread out.
Disagreement is a sign of good research, good science. The idea of bias in the human sense of bias. How do you think about instilling in machine learning something that's aligned with human values? In terms of bias, we intuitively as human beings have a concept of what bias means, of what fundamental respect for other human beings means. But how do we instill that into machine learning systems? Do you think so?
I think there are short term things that are already happening and then there are long term things that we need to do. In the short term, there are techniques that have been proposed and I think will continue to be improved. And maybe alternatives will come up to take data sets in which we know there is bias. We can measure it. Pretty much any data set where humans are being observed taking decisions will have some sort of bias, discrimination against particular groups and so on. And we can use machine learning techniques to try to build predictors, classifiers that are going to be less biased. We can do it, for example, using adversarial methods to make our systems less sensitive to these variables we should not be sensitive to. So these are clear, well defined ways of trying to address the problem. Maybe they have weaknesses and more research is needed and so on. But I think in fact they're sufficiently mature that governments should start regulating companies where it matters, say like insurance companies, so that they use those techniques, because those techniques will probably reduce the bias, but at a cost, for example, maybe their predictions will be less accurate. And so companies will not do it until you force them. All right, so this is short term, long term, I'm really interested in thinking of how we can instill moral values into computers. Obviously, this is not something we'll achieve in the next five or 10 years. How can we, you know, there's already work in detecting emotions, for example, in images, in sounds, in texts, and also studying how different agents interacting in different ways may correspond to patterns of, say, injustice, which could trigger anger. So these are things we can do in the medium term and eventually train computers to model, for example, how humans react emotionally. I would say the simplest thing is unfair situations which trigger anger. This is one of the most basic emotions that we share with other animals. I think it's quite feasible within the next few years, so we can build systems that can detect these kinds of things to the extent, unfortunately, that they understand enough about the world around us, which is a long time away. But maybe we can initially do this in virtual environments. So you can imagine, like a video game, where agents interact in some ways and then some situations trigger an emotion. I think we could train machines to detect those situations and predict that the particular emotion will likely be felt. If a human was playing one of.
The characters, you have shown excitement and done a lot of excellent work with unsupervised learning. But on the supervised, you know, there's been a lot of success on the supervised learning side.
Yes, yes.
And one of the things I'm really passionate about is how humans and robots work together. And in the context of supervised learning, that means the process of annotation. Do you think about the problem of annotation, put in a more interesting way as humans teaching machines? Is there?
Yes, I think it's an important subject. Reducing it to annotation may be useful for somebody building a system tomorrow. But longer term, the process of teaching, I think, is something that deserves a lot more attention from the machine learning community. So there are people who've coined the term machine teaching. So what are good strategies for teaching a learning agent? And can we design, train a system that is going to be a good teacher? So in my group, we have a project called Baby AI or Baby AI game, where there is a game or a scenario where there's a learning agent and a teaching agent. Presumably the teaching agent would eventually be a human, but we're not there yet. And the role of the teacher is to use its knowledge of the environment which it can acquire using whatever way, brute force, to help the learner learn as quickly as possible. So the learner is going to try to learn by itself, maybe be using some exploration and whatever, but the teacher can choose can have an influence on the interaction with the learner so as to guide the learner, maybe teach it the things that the learner has most trouble with, or just add the boundary between what it knows and doesn't know and so on. So there's a tradition of these kind of ideas from other fields, like tutorial systems, for example, and AI. And of course, people in the humanities have been thinking about these questions, but I think it's time that machine learning people look at this, because in the future we'll have more and more human machine interaction with a human in the loop. And I think understanding how to make.
This work better, all the problems around that are very interesting and not sufficiently addressed. You've done a lot of work with language too. What aspect of the traditionally formulated Turing Test, a test of natural language understanding and generation, in your eyes, is the most difficult of conversation? What in your eyes is the hardest part of conversation to solve for machines?
So I would say it's everything having to do with the non linguistic knowledge, which implicitly you need in order to make sense of sentences, things like the Winograd schema. So these sentences that are semantically ambiguous, in other words, you need to understand enough about the world in order to really interpret properly those sentences. I think these are interesting challenges for machine learning because they point in the direction of building systems that both understand how the world works and its causal relationships in the world and, and associate that knowledge with how to express it in language, either for reading or writing.
You speak French?
Yes, it's my mother tongue.
It's one of the Romance languages. Do you think passing the Turing test and all the underlying challenges we just mentioned depend on language? Do you think it might be easier in French than it is in English or as independent of language?
I think it's independent of language. I would like to build systems that can use the same principles, the same learning mechanisms to learn from human agents, whatever their language.
Well, certainly us humans can talk more beautifully and smoothly in poetry. So I'm Russian originally. I know poetry in Russian is maybe easier to convey complex ideas than it is in English, but maybe I'm showing my bias and some people could say that about French. But of course, the goal ultimately is our human brain is able to utilize any kind of those languages, to use them as tools to convey meaning.
Yeah, of course there are differences between languages and maybe some are slightly better at some things, but in the grand scheme of things, where we're trying to understand how the brain works and language and so on, I think these Differences are minute.
So you've lived perhaps through an AI winter of sorts.
Yes.
How did you stay warm and continue with your research?
Stay warm with friends.
With friends. Okay, so it's important to have friends. And what have you learned from the experience?
Listen to your inner voice. Don't be trying to just please the crowds and the fashion. And if you have a strong intuition about something that is not contradicted by actual evidence, go for it. I mean, it could be contradicted by.
People, but not your own instinct based on everything you've learned.
Of course you have to adapt your beliefs when your experiments contradict those beliefs. But you have to stick to your beliefs. Otherwise it's what allowed me to go through those years, it's what allowed me to persist in directions that took time. Whatever other people think took time to mature and bring fruits.
So history of AI is marked with these. Of course it's marked with technical breakthroughs, but it's also marked with these seminal events that capture the imagination of the community. Most recent, I would say AlphaGo beating the World champion human go player was one of those moments. What do you think the next such moment might be?
Okay, sir, first of all, I think that these so called seminal events are overrated. As I said, science really moves by small steps. Now what happens is you make one more small step and it's like the drop that fills the bucket and then you have drastic consequences because now you're able to do something you were not able to do before or now, say the cost of building some device or solving a problem becomes cheaper than what existed and you have a new market that opens up. Right. So especially in the world of commerce and applications, the impact of a small scientific progress could be huge. But in the science itself, I think it's very, very gradual.
And where are these steps being taken now? So there is unsupervised.
So if I look at one trend that I like in my community, so for example, at Milan, my institute, what are the two hottest topics? Gans and reinforcement learning. Even though in Montreal in particular, reinforcement learning was something pretty much absent just two or three years ago. So there's really a big interest from students and there's a big interest from people like me. So I would say this is something where we're going to see more progress even though it hasn't yet provided much in terms of actual industrial fallout. Like Even though there's AlphaGo, Google is not making money on this right now. But I think over the long term this is really, really important for many reasons. So in other words, agent. I would say reinforcement learning may be more generally agent learning because it doesn't have to be with rewards. It could be in all kinds of ways that an agent is learning about its environment.
Now, reinforcement learning you're excited about. Do you think GANS could provide something, some moment in.
Well, GANS or other generative models I believe will be crucial ingredients in building agents that can understand the world. A lot of the successes in reinforcement learning in the past has been with policy gradient, where you just learn a policy, you don't actually learn a model of the world. But there are lots of issues with that. And we don't know how to do model based RL right now. But I think this is where we have to go in order to build models that can generalize faster and better to new distributions that capture to some extent at least, the underlying causal mechanisms in the world.
Last question. What made you fall in love with artificial intelligence? If you look back, what was the first moment in your life when you were fascinated by either the human mind or the artificial mind?
You know, when I was an adolescent, I was reading a lot, and then I started reading science fiction.
There you go.
That's it. That's where I got hooked. And then I had one of the first personal computers, and I got hooked in programming. And so it just, you know, start.
With fiction and then make it a reality.
That's right.
Yosha, thank you so much for talking today.
My pleasure.
