People & AI - Dr. Yoshua Bengio on Ethical AI

For this episode, we’re re-airing the first conversation we had on this podcast with Dr. Yoshua Bengio, which took place in February 2022. Dr. Yoshua Bengio is a prominent figure in the field of artificial intelligence, renowned for his foundational contributions to deep learning, which garnered him the 2018 A.M. Turing Award alongside Geoffrey Hinton and Yann LeCun. As a Full Professor at the University of Montréal and the Founder and Scientific Director of Mila – Quebec AI Institute, Yoshua has significantly advanced AI research and its ethical applications, evidenced by his leadership roles and numerous accolades, including the Killam Prize and being the most cited computer scientist in 2022. His work extends beyond academia into shaping policy on responsible AI development.
May 10, 2024
5 min read

Listen on

Apple podcasts:




Karthik Ramakrishnan [00:00:06]

Welcome to another episode of the People and AI podcast, where we explore the forefront of artificial intelligence. I'm your host, Karthik Ramakrishnan. I'm the CEO and co founder of Armilla AI, and join us as we delve into AI's latest advances, ethical considerations, the risks, but more importantly, the exciting capabilities of what this technology can provide, particularly in the context of the enterprise.

In this episode, we revisit my initial conversation with Doctor Yoshua Bengio from February 2022. Doctor Bengio is a leading figure in AI. His influence extends beyond academia, shaping policies for responsible AI development. Even two years later, Doctor Bengio's insights remain pertinent. Enjoy this episode of people and AI.

Karthik Ramakrishnan [00:00:53]

Joining us today is one of the world's most recognized and leading experts in artificial intelligence, Joshua Bengio, most known for his pioneering work in deep learning, earning him the 2018 A.M. Turing award, the Nobel Prize of Computing with Doctor Hilton and Doctor LeCun. He's a full professor at University of Montreal and the founder and scientific director of Miele Quebec AI Institute. He co directs the CIFAR Learning in machines and Brains program as a senior fellow and acts as a senior director of EvAdo. In 2019, he was awarded the prestigious Killam Prize and in 2021 became one of the most cited computer scientists in the world, if not the most cited computer scientist in the world. He's a fellow of both the Royal Society of London and Canada and an officer of the Order of Canada. Doctor Bengio, thank you so much for joining us today. While this is our inaugural episode, we do start this day at a very serious tone with other happenings around the world, especially in Europe. Doctor Bengio, would you like to share some thoughts on and what's happening and how that concerns us?

Dr. Yoshua Bengio [00:02:00]

Yeah. Like many people, I'm very worried and concerned what is going on in Ukraine, and what it brings me back to is the importance of democracy, the importance of an organization of our societies in which human rights and human values and equity and justice are the center. And we all take part in power in some sense. And I think this is important from an AI perspective, because AI being a powerful tool, and it's going to be even more in the future. It can be used in ways that go against these values. It could be used in wartime, you know, with killer drones. That may happen in Ukraine, we don't know. And so we should always go back to what are the objectives, the values, the principles that should guide our decisions on our world.

Karthik Ramakrishnan [00:02:48]

Absolutely. And we wish the very best for the people of Ukraine in their stance for democracy and as you mentioned, Doctor Bengio, stance for the world, and very concerning in terms of how we lead ourselves into the future. From there, Doctor Bengio, I think everyone's curious. As a start, how did you get into deep learning?

Dr. Yoshua Bengio [00:03:07]

You know, my first research subject was neural networks in around 1985. I started reading some of the early papers of the early eighties on that topic, including those of Jeff Hinton. That then became a kind of role model for me, and I got very excited reading these early Neil net papers. The connectionist movement of the eighties, a lot of people there were not just computer scientists, but also cognitive scientists, neuroscientists trying to understand intelligence in general, and the idea that there would be general principles that could explain intelligence, both ours, the intelligence of animals, and would allow us to build intelligent machines. I found that so exciting that I've been writing that way since then.

Karthik Ramakrishnan [00:03:55]

And we're really glad that you kept at it through all of its winters and successes. And here we are actually bringing it to life, mostly through a lot of the contributions from your team at Mila. So what is top of mind for you today? What's the area of focus that you think is most important in the next level of research in AI?

Dr. Yoshua Bengio [00:04:13]

Well, first of all, as a caveat, I really believe in the importance of a diversity of research paths and research directions and respecting other people's choice and what they think is important. So I'm only going to speak for myself. And what really excites me these days is the perspective of bridging the gap between human intelligence and current state of the art and machine learning. And the latter is interestingly dominated by approaches that I've contributed to. So deep learning, typically in a nutshell, that correspond to some kinds of abilities we see in the brain that I like to call system one abilities. But humans also have these system two abilities that have to do with more deliberate thinking and conscious processing that give us actually some powers of generalization and understanding and imagination and modeling causality that are still lacking in state of the art machine learning. So that's the objective. But what's exciting is I see a path to bridge that gap.

Karthik Ramakrishnan [00:05:23]

And your 2017 paper, consciousness Pyre, you talk about how you have to have some assumptions about the world, even minimal ones, in order to learn efficiently. This, to me, feels a lot like what transfer learning is attempting to do. You train a model in Wikipedia to teach in English, and then further train it on the task you actually care about. Are you envisioning something similar but more general when you talk about conscious priors.

Dr. Yoshua Bengio [00:05:49]

The idea, I mean, what machine learning rests on, both in theory and in practice, is putting some constraints, preferences, priors, regularizers, architectures, choices of objective functions. It's all about not just the data, but also some extra bits of information that allow to learn efficiently with that data. And in a way, the idea of the consciousness prior is that it looks like brains, especially human brains, are easier to understand at that level, are exploiting some preferences, some ways, some architectures, choices and so on that we're currently leaving on the table. It's like we're leaving money on the table. If you want their tricks that brains use, that I think we can make sense of from a machine learning perspective and that we can incorporate in neural nets. And when we try to understand these technical term inductive biases, it makes sense that they help us to reason about the world at the level where we think, at the level where we talk, at the level where we are conscious of different aspects of how the world works and how we take decisions and so on. That emerged part of what's going on in our brain, which is a tiny, tiny bit of what's going on, exploits assumptions, if you want inductive biases that haven't been used. And we have good reasons to think that they could really make a big difference in terms of some of the current weaknesses of machine learning, including the difficulty in transfer.

So when you transfer to a new task, it's a different task, it's a different distribution. So why should you expect that anything to work? It's because they have something in common and how to extract the things, the aspects of the world that are in common between different tasks. Actually, this is a question that a causal understanding helps us to put some light on. And if you look at cognition, you also get some clues about how humans manage to do that. For example, they break knowledge into small pieces that can be reused in novel ways. And so it's not like things have changed that much. It's just that we're using them in novel ways.

Karthik Ramakrishnan [00:08:01]

That sounds a lot like having a model understand the world it exists in and how the world operates, right? And how the various parameters operate. Back 2030 years ago, expert systems were a means to that end, like, let's understand the dictionary. If we understand all the words in a dictionary, then we could create sentences that didn't really work. It sounds like we're moving again into that sort of a thing, thinking, but with the new technology. Could you elaborate on that?

Dr. Yoshua Bengio [00:08:27]

Yeah, yeah, totally. I mean, there were a lot of problems with the classical AI, symbolic, rule based systems they didn't learn, which we know now is super important, because there's a lot of information about the world that we simply don't express verbally or are hard to put in a computer. So they had, like, knowledge engineers to try to tease that from humans. It turned out we can't express a lot of our intuitive knowledge, so it failed, and so you have to rely on data. Right. But the modern machine learning is essentially only using data. I mean, not, we're using things like inconvenience, like about spatial invariances and equivariances and things like that. But there's a lot of inductive biases that we're, as I said, leaving some money on the table.

Now, there are good things in the good old fashioned AI that we would like to incorporate in neural nets, like modularization of knowledge in the right way, like the compositional abilities of putting together pieces of knowledge on the fly as needed to explain what we're seeing or to plan something new. And even the use of symbols is not such a bad thing. Now, we need to adapt our nets to incorporate all these things, but I don't think it's so far off.

Karthik Ramakrishnan [00:09:38]

And you've also said in past that the work of Daniel Kahneman in his book has really influenced your thinking around system one and system two, sort of how we take those conscious versus subconscious or unconscious ways of thinking about the world. For the layperson would love to understand how that's being brought in, in a very tangible way into the research.

Dr. Yoshua Bengio [00:10:00]

Right. So, first of all, maybe contrary to some people who are also seeking similar goals, who are exploring what's called, like, hybrid models that are partially symbolic and partially like neural nets, I'm trying to do it all with neural nets, except that they would have different ways of being trained and being organized, because your brain is just a big neural net now, even though it's just one big neural net, there is something different between what Kahneman calls system one and system two abilities. So an easy way to understand this is everything you can do in less than half a second without having to think about it. That's system one. Everything that requires thinking and time because it's not instantaneous, might take seconds, sometimes hours or days of thinking about things. So that's more like system two. Everything that the part of what our brain does which goes through consciousness, what essentially goes to your working memory and can be verbalized, that's the system two part. But it's not like there are two different parts. It's just the bits that emerge because they go through what bars called the global workspace, the working memory, which is a bottleneck. So the bits of information from system one that are selected out of millions to occupy your thoughts for half a second, that's the part that's you're conscious of, that you can communicate. It's all part of one thing. It's just that it's the emerged part of the iceberg, if you want. And the top of the iceberg wouldn't work alone unless you're doing very simple things like arithmetic, like add three and five. Even if you're doing math, you need your intuition to find the theorem to find the proof. Otherwise, it's just exponentially hard. So one of the classical problems with good old fashioned AI is search.

If you're just trying to solve everything in a symbolic way, you are facing this exponential search problem all the many ways you could explain something or find a proof for something or whatever plan. It's just exponentially large, and it's never been solved, but now it's basically solved by using neural nets, by using deep learning. So you look at AlphaGo, you learn a value function that somehow is trained to mimic the result of trying. What would happen if I tried the exponentially large number of possible continuations of the current board into an endgame? There's no way I can go through all of them, but I can train a neural net which generalizes, so it only needs to be trained on a finite number of these trajectories of games. And then it learns to guess what should you do? And that's the system one thing. And then the system two thing is able to maybe try a few things and see how they work and compare what's coming out of your imagination, your system one, with the verbalizable facts and rules and social norms and so on, that you want to be, or your goals that you are aware of.

Karthik Ramakrishnan [00:12:57]

It feels to me that the system two first makes conscious decisions of which parts to choose. When it doesn't know something, as it tries and chooses these different paths, it then gets internalized into a system one, so that next time it sees that appear, system one takes, it's just an unconscious decision. Like driving a car in a known road versus an unknown road.

Dr. Yoshua Bengio [00:13:17]

Exactly. If you're in a new city, initially you have to concentrate a lot, and you can't just rely on your habit because you might do something stupid. Or even worse, if you're in a city where the traffic laws are different, then you really need to concentrate and have your assistant to sort of watchdog tell you, no, no, no. You have the impulse to go left, but you have to go right. And another, like, nice examples that have to do with AI bias and ethics is like racial bias. If I absorb a lot of texts, it's going to include a lot of racial bias or gender bias or whatever. And we don't want our AI systems to just sponge us, drink all that, and then behave in a way that's aligned with the data, but not aligned with what we think they should be doing. And humans, if I was raised in a racist family, but I don't think this is the right thing to do, my impulse is going to be to interpret things or to act in a way that embodies these racist priors. But if my values tell me no, that's not what I want for myself, that's not what I want for my society, I can inhibit myself from acting in that way and teach my system one to react differently. And after a few years, your immediate impulsive response will be adapted through practice of doing the right thing, not based on the data, but based on the system two level verbalizable goals and norms and so on, that we would like our machines to have the system one, system two division is also a path towards making more moral machines.

Karthik Ramakrishnan [00:14:55]

Thank you for taking us there. I think this is a very important topic, particularly around managing risks of AI. Now, in your research, I'm sure you're taking this into consideration. Data scientists in the real world who are building these systems and where these intentional or unintentional, most likely unintentional biases creep in, what should they be doing to mitigate that as they're building these systems?

Dr. Yoshua Bengio [00:15:17]

Well, there are already things you don't need high tech or like, you know, what I and others will discover in the next ten years. There are things you can do already. I mean, step number one is already choose your data wisely, if you can, and often you can't document things. So we know where the data came from, how the system was trained. And sometimes there are things you can do to mitigate, right? So you can, even with small amounts of data, train a system to detect that maybe an utterance may have a racist bias or something. I mean, it might not be perfect, but it might raise a flag. And then you could use that to decide to not act or just say, I don't know, I'm about to say something stupid, so I'm going to shut up.

Karthik Ramakrishnan [00:15:58]

That's fair. In fact, you chair the responsible AI stream with the global partnership on AI that you've launched. Could you talk to us a little bit more about what the effort there are that's taking place in that group to mitigate some of these risks?

Dr. Yoshua Bengio [00:16:12]

Certainly. So, GPAI the global partnership on AI is an international body that was created a couple of years ago, and it brings together 20 odd countries. Now it's increasing. There are new countries coming in, mostly democracies, that have at heart to develop together more responsible ways of developing AI and deploying it, and also encourage the development of AI in their economies. So in the working group that I've been co chairing, we had different committees that looked at a particular issue. So one committee worked on social media, which is a big challenge. We had also a committee that worked on how AI could be used to fight climate change, and we were able to show some of our work. And a report at the last cop and I was involved in a committee that worked on how AI could be used to prevent future pandemics and develop new drugs against infectious diseases, antimicrobial resistance. And one of the common themes of these two working groups, these two committees, is that you're thinking that there are really important areas of application of AI, which you could call AI for social good, where important, because society really needs those things. We need those innovations in health, in the environment, in education. But industry doesn't necessarily go there. It might not be commercially sufficiently attractive. So first of all, we need to recognize that, and then governments need to recognize that and see that they have a duty to their goals. Most of the money that say, the canadian government spends is on healthcare and education, the environment. So these are things that governments spend a lot of money on, and they're not trying to innovate in those areas. It's like we're just spending money, but we're not investing to make it cheaper in the future to deal with those problems. So part of what we did was try to wake them up, say, look, you have an opportunity to use innovation in particular AI in those areas and here specific ways that this could happen.

Karthik Ramakrishnan [00:18:18]

And have you seen challenges in, or what are the challenges that you've seen in actually executing on coming up with how we bring AI into these fields?

Dr. Yoshua Bengio [00:18:27]

So there are technical challenges, and that's mostly what I do with my days. And there are political challenges, and that's harder because I'm not a politician, and it's harder because society is complex and politics is complex. So on the political side, one thing I learned that I basically got from talking to people who are closer to, I mean, governing, high ranking civil servants and so on, is that governments in our democracies will move in a direction not only if they think it's good for society, and even if they see evidence it might not be enough. They also will move if the people care. So the reason we see some too slow movement on climate change these days is because there are more and more people who realize this is a really serious issue, and that's putting pressure on government. The fact that we are in this pandemic and hopefully coming out of it, has put in the minds of people the importance of infectious diseases, which probably most people didn't care about just two years ago. So you need awareness, you need people, and awareness that there is an issue, but also awareness that innovation can be an important part of addressing these challenges, these global challenges, or these national challenges. I'm not sure we are totally there yet, public awareness.

And I think scientists have a responsibility to be part of that. It's not their main job. Like, I can't be like talking to the media all the time, but I feel I have a responsibility then, of course, on the technical side where it is. That's where it gets fun, because I feel like I have more control in some sense. And it's exciting because as we're trying to solve these problems, like designing drugs much more efficiently for preventing future infectious diseases, I realize there's difficult machine learning questions, AI questions, that if we can advance them, is also going to advance big questions in AI, like curiosity, exploration, understanding, like the causal structure of the world. These all come up when you try to solve these practical.

Karthik Ramakrishnan [00:20:27]

I would like to talk about your work on this. Climate does not exist. A website that you launched earlier this year to educate people on climate change. Can you talk a little bit more about the project and the impact it will have?

Dr. Yoshua Bengio [00:20:38]

This was actually started before GPAI, and it's the work of a whole team of people, led in particular by Sasha Luccioni, Victor Schmidt and many others. And it's really about how AI can be used to, in a way, play the game of awareness and politics. So one reason why we haven't been acting so much on climate is simply because people don't put a high priority on it. And it's interesting to ask why don't they? Because rationally, we should quickly be changing, we shouldn't be waiting. And so, talking to experts, it's clear that a major issue is a cognitive bias. In other words, because our human brains have these inductive biases, they're not always good. They can also be bad. For example, we focus on the short term dangers.

We have a very hard time to give importance to things that are going to happen in ten years from now and are going to be very, very bad, or things are going to be bad for our children. Unless somehow we get sufficiently stimulated. It becomes very present in our mind, things that are happening far from us. If I look outside, I don't see climate change. I see trees and grass and so on. Well, now it's snow, but it looks normal. You need scientists to measure in order to really see that something really wrong is going on. And so our common physical sense of danger isn't triggered by the current situation yet. It is very serious. The idea was, what can we do to raise awareness using the ability of machine learning to create visual, I mean, create visuals, create images that might touch us more directly. If I show you your house flooded because of climate change, maybe it's going to have more of an emotional impact and you're going to take it more seriously because you're going to feel, oh, this could be here, this could be happening to me, I need to worry about it. So it's not that people don't understand the messages, it's just that it doesn't get to their decision making part of the brain at an emotional level. There are other things. You're worried about your job, you're worried about your kids not getting to school. So we've done an experiment and we developed a technology to create these personalized images of houses and streets and so on in your neighborhood that could be modified by climate change because of fires, for example, or flooding or smog. What was the idea?

Karthik Ramakrishnan [00:23:12]

Most people deny it or reject it, because if you cannot see what you cannot see, it's hard for them to accept. And I think this is phenomenal. In fact, I went on it and I saw that my front yard is going to be pretty much submerged. I live near the lake. And so that was a bit of a shocker for sure. Now, in terms of the risks of AI is a big topic. There's inherent risks that are there today and risks that are there in the future. Are there tools and technologies? And we talked about data being very important in terms of how we select data and how we use that to train these models. And that's the first step. Are there any technologies that people should be thinking about or working on to further mitigate the risks?

Dr. Yoshua Bengio [00:23:51]

Yeah, but there are also things that are not technology. So a lot of the discussions we had at GPI is about regulation. We haven't been regulating the computing industry very much. We, you know, we regulate the airline industry, we regulate the nuclear industry, regulate the chemicals industry and so on. So there's been said this exception that doesn't make any sense because computers are everywhere and AI is, you know, becoming powerful. We need to make sure it's done in a way that protects the public. So that's one part of the answer. And then on the technology side, there's a lot of work that's going on. You know, there are conferences on the subject of AI, safety and AI, fairness and ethics, and some of that is more social, and a lot of that is new algorithms that may be couldable decisions. And I mentioned the work that I'm doing on how we could build future deep learning systems that can understand what we're saying at the verbal level in a way that's different from like current things like GPT-3 that are responding intuitively, but not necessarily in a coherent, consistent way throughout their behavior because they are lacking this reflexive system, two type of computation. And so I'm hoping that in the future, with the sort of changes I'm working towards, we'll be able to do what Asimov was dreaming of, that we can state the things we want in natural language and computers will be able to take that, not like we currently do, as like training data, but rather as goals or constraints, social norms, the same way that we take laws and rules, you know, whether they are the rules of our home or the rules of our city or the rules of our country, we take them and we are able to see when our behavior could contradict those, and then we can decide to act in a way that's consistent with them and that's currently lacking. And if we want to build moral machines, there are two paths which could be combined. One is the morality is induced from the data. For example, there's this idea of training AI systems by looking at how really wise people would act, would behave in different situations. Well, it's hard to get that kind of data, like for a number of reasons, but still that's a direction. Another is to be able to state what we want, which is also has its problems. Like not everything in morality is easy to state non ambiguously, but that's the way humans for the most part have been doing it. So I think we should try all the avenues we have to build machines that are going to be better aligned with our needs and our values.

Karthik Ramakrishnan [00:26:41]

It's a bit of a more philosophical question, maybe not. When we talk about morality at some level, it is also subjective, right? Some societies hold different things of value versus other societies. So how do we pick a generalized, let's say a model that reflects all of these various different moralities, if you will? Again, I don't think there's an answer here, but certainly it's something to think about.

Dr. Yoshua Bengio [00:27:06]

Well, so far, we don't have to pick a single model for everyone or for every country. If you go to a different country, maybe there are things that are different. And if you had a robot in this other country, maybe it should conform to the laws and norms in that country. That's not a big deal because we can have different robots in different countries. Now, that being said, I believe there's no consensus on this, but I do believe that there are some built in moral instincts. In other words, they're genetically written in our DNA that all humans share. In fact, more than that, I mean, one of the reasons to think this is true is that we share some of that with other mammals and primates in particular. Like, you know, the sense of injustice is something that you can see other primates really display, and that's the basis of morality. Like, you know, injustice and being angry when something unfair is going on. So I also think that we could have fairly universal base for morality and what is right and wrong. And then on top of that, of course, different societies have all kinds of rules, their legal system that if you want to operate in those societies, you need to be somewhat aware of. But, you know, even then, like probably you and I not being lawyers, we only have like, a very superficial understanding of all of those books of laws. But we do get the general principle in machines, and it might be different from one country to another. So I think machines could be using that in order to behave properly in a legal way. But I'm more concerned about morality than legality of the actions of an AI system. And I think legality is codification of some of the morality that we hold and more about justice and fairness. To your point?

Karthik Ramakrishnan [00:28:41]

That's an excellent point. Last question on this topic. Do you think that, how important is it to solve that problem right now as we build these nascent AI's? And the reason I ask that, if we don't solve it now, do you think we're going to go down this path where it's going to be very hard to solve for it in the future?

Dr. Yoshua Bengio [00:29:04]

I don't know. Again, I'm a big believer in diversity of research. So for some people, it's very important. If you have been a victim of injustice, and of course, that's the case of many people, you might give a lot more weight on solving the discrimination issues that may come out of AI systems. It is important. The technology that can make AI not a danger for our democracy, or even helping it around things like social media, for example, is a different concern, but also something that we might want to move quickly. You could have other concerns, like recently, because of the pandemic, I've been concerned about our health, and I think the area of healthcare is one where AI has the greatest potential of touching positively human beings in the coming years and really saving a lot of lives. So, in a way, it's urgent, because each year that we don't make progress, people may be dying unnecessarily. And I can think of very specific threats that are global threats, medically, like antimicrobial resistance. The fact that you have these mutations of bugs that are all around us that become resistant to the drugs we currently use, just like we're seeing with COVID-19 how do we prepare? And it turns out AI could be really good to accelerate and improve the exploration, to find solutions to these and other medical problems. The environment can be very pressing. I mean, it looks like, okay, the sky is not going to fall tomorrow morning, but it's a race. Whatever we do is going to take years or decades to have an impact. So it's really important to get started. You see, there are so many challenges. I don't know. I wouldn't want to say, oh, we need to put all our eggs here or there. Every person should go where they're most motivated to make a difference. But most important, I think, is to feel like you can make a difference and go where you can think you can make the most difference.

Karthik Ramakrishnan [00:30:53]

That's an excellent point. Maybe onto some fun questions, if you don't mind. You do a lot of interviews and podcasts. What's a question that you wish you had been asked that you don't get asked often?

Dr. Yoshua Bengio [00:31:04]

Well, it's very rare that the interviewer would have sufficiently read technical papers to ask the questions that I would like to talk about to explain what is going on under the hood and how we might get to the next generation of AI systems. And also, it's difficult to answer. So it's not just like they don't ask it, it's also, it's difficult to answer in a way that's going to be accessible. But of course, as a researcher, I'm much more excited by research I'm doing than by anything else.

Karthik Ramakrishnan [00:31:35]

And so maybe give us a try. If you could talk about that a little bit. What would you like to share?

Dr. Yoshua Bengio [00:31:41]

Right. So let me talk about the challenge of out of distribution generalization. This is a technical term, and it refers to difficulty that we have in AI. But in a way, humans have the same problem, that we may be trained with an experience that comes from some distribution, some type of examples, and then we could be confronted with a fairly new situation. So maybe you've been seeing things around you and understanding visual worlds in your city, your neighborhood, and suddenly, if you get transported to the moon, things look very different. The distribution of images is very different, but somehow your brain is able to make sense of it. How is that even possible? And intuitive way to understand how it is possible is that your brain has made sense of the world by breaking down your understanding of it into little pieces, which you could call causal mechanisms, like the laws of physics. I mean, it doesn't have to be the actual laws of physics. Like, most people don't understand the laws of physics, but the way that they're taught in school. But the laws of physics, as in what we call intuitive physics, like, intuitively, a two year old understands gravity. In fact, I think it's even maybe around 14 months or something. So, very early on, we understand the causal structure of the world at a level that's intuitive. It's a system, one thing, and we're able to port that knowledge. Two settings where it's the same laws of physics, but the inputs are different. So if you're, you know, on the moon, it's still gravity. The constant of gravity is different, but it's the same formula. The landscape looks very different, but that's just because, well, there's a different history and there's no life and there's, you know, no atmosphere and so on. But the physics of how light bounces around and, you know, comes from rocks on the moon to your eyes is the same, and you don't need to retrain yourself for that. So somehow we have a way to decouple the parts about how the world works. These causal mechanisms from the particulars of where you are or the settings or the conditions where you are, that combined with those mechanisms, give rise to what you see, what you observe by decoupling these things, like the laws of physics are the same on the moon and earth, but the context, you know, what grows and doesn't grow or what sort of rocks there are, there is different. And so the images that you see with your eyes are different, but by decoupling them, it means that you're robust to those changes in distribution. So if I go to a different place, if I go. So I'm going to go back to my driving example. Everything I know about cars, about how humans behave on the road, how sort of the social norms about, you know, you don't run onto people and things like this when you're driving. It's the same whether you are in Montreal or you're in London. There's only this wonderful rule that changed. Right, you drive on the left or on the right. And because your brain has organized knowledge into these different kind of independent pieces, if I change one of those pieces, I change one of the rules. It's okay, you need some adaptation. But a, you're able to notice if you're about to do something that contradicts this new rule, and b, you're able to adapt yourself fairly quickly using your internal like watchdog so that you don't run onto people on the streets in London when you're driving. So I think that we can introduce these kinds of structures. For example, for many years, I've been working on modular neural nets. Instead of having one big homogeneous neural net, we break it down into smaller pieces that can talk to each other in a way that, inspired by this communication between parts of your brain in different theories of conscious processing, where there's a bottleneck of communication between those parts. And so, in other words, it's not fully connected. These modules can talk to each other, but there's a. A little channel that only allows a few bits of information to go back and forth, and that sort of thing actually helps. But it's just the beginning of the path towards integrating these inductive biases that humans exploit to allow them to generalize to new distributions, new tasks very quickly compared to current AI systems.

Karthik Ramakrishnan [00:36:17]

For more adaptive than the brittle systems that we are building today.

Dr. Yoshua Bengio [00:36:20]


Karthik Ramakrishnan [00:36:21]

That's excellent, I think. How far along are you in coming to a breakthrough with this?

Dr. Yoshua Bengio [00:36:26]

Well, I'm a natural enthusiast for the projects that I'm leading, so I think it's a few years down the road. But for me, the most important thing is I see a fairly clear path, and I see the kinds of tools that may get us there. Of course, we may encounter obstacles that are hard to predict on the way. So recently we discovered a new framework for training deep learning systems called GFlowNets, for generative flow networks. And we had a paper at the last neerips. It turns out this framework makes it fairly easy to incorporate the system. Two inductive biases. And that's what makes me very hopeful that we can get there in a few years.

Karthik Ramakrishnan [00:37:12]

Amazing. One of the things that used to be, and you flagged this very early on about how so the research that's happening these days is within corporations and the accessibility of that to the general public for the greater good. Have you made much progress in making research open and do you think we're in the right trajectory?

Dr. Yoshua Bengio [00:37:31]

I would say the slope is positive.

Karthik Ramakrishnan [00:37:34]

That's great.

Dr. Yoshua Bengio [00:37:34]

There's more awareness that as we share more data, algorithms, source code, we all benefit. Unfortunately, the current economic system, where a lot of the innovation is happening in companies, comes a little bit in contradiction with that, because from the point of view of the survival of a company, it's not profitable to share your insights, your discoveries, your data, your code with your competitors. It makes sense, right? So there's a little bit of a contradiction here, like what is good for society if we share more people to use it in other companies, or in academia or civil society, and everybody wins. That's how research works. By the way, the reason we're making scientific progress is because we are sharing what we find. If we have to reinvent everything all the time, we're in trouble. It's going to be very inefficient. So sharing knowledge makes progress faster. But economic system based on competition between companies goes against that. So how do we go around that? I mean, I'm not an economist, but I think there are things that are within reach. For example, in our GPA reports, we talk about sharing data, but it's not just something we wish. We think that governments have a role there. Let me be more precise. Let's say that a government funds companies for some innovation, like discovering carbon capture materials or new future antivirals against COVID-19 or something, because we feel like we need it. And governments should get into funding that kind of research. And a lot of it is going to happen in industry. So if government is putting money on the table, they could also come. That money could come with conditions. Okay, we're going to pay for 50% of it. You'll be able to use the algorithms for other things, but the results you get on these socially important directions, you have to share, you have to share the code, you have to share the data. It should be easy for other companies or anyone. And then there are many variations. But the point is, governments have a lever here. It doesn't completely change our economic system. I think it's also interesting to think about how could we change our economic system so that innovation is going to be more efficient in the first place, not just in those places like, you know, where social, there's a big incentive for governments to step in.

And then there's also at some point this is also called an arms race between governments too. So they might not be incentivized to even make it open. And that's a whole other kind of worms, if you will, that needs to be encountered, especially, especially in the kind of cold war situation in which we're going back, whether it's, you know, with China or Russia now, it's really sad because it's not going to help progress in any way.

Karthik Ramakrishnan [00:40:27]

Well, I mean, that is their hope. But as you said, the slope is positive. So hopefully we're making more progress in that direction than not. Final couple of really fun questions. What's your favourite AI movie?

Dr. Yoshua Bengio [00:40:38]

My favourite AI movie, so I usually answer 2000 Space Odyssey. This movie is amazing. And you have to consider when it was done, we didn't have the special effects facilities that we have today. But it's not just the visuals and the special effects. The story has AI in it, playing not such a great role, but an interesting one with machines that learn actually. And it's, you know, it spans the origins of humanity to our future. Pretty amazing.

Karthik Ramakrishnan [00:41:09]

Even when I read, as you've brought up as well, Asimov's books, I mean, written in the fifties and sixties, and he's really so prescient in many things that he's talked about as well. I think we still are trying to bring to life what was talked about 50 years ago or 60 years ago. One final question. If you could change one thing about your career up to this point, what would it be?

Dr. Yoshua Bengio [00:41:28]

Well, it's hard to say because it worked out really well. So here's what, you know, I would give as advice to myself when I was 25 and starting to do research, I would say focus even more on trying to understand what is going on. Don't waste time on the low hanging fruits. Go for the hard problems. I mean, in a way I've done a lot of that, but I think I could have done even more and earlier. The most important is the first thing, try to understand what you're doing. That is the number one thing in science. And unfortunately in AI, a lot of people are just toying around with the software tools. I mean, so the software tools are marvellous because they make our algorithms much more accessible to many more people. But it also can make us lazy. So when I was a grad student, there was no open source like neural net code. I had to write everything myself. And c. There was not even c then. So when you write your own code from scratch at the low level, you better understand what you're doing. Otherwise it's just gonna not work. Right? But nowadays, not so much, you know? So there's the flip side of having all this luxury. It's like you don't understand your phone, but it's very useful to you. That's fine because you don't do research on phones. But if you were, you better understand you know what's under the hood. So yeah, at the same time, you have to make choices. You can't understand everything. There's so many things. So you need to be wise in picking what to investigate. You know, pick some battles. That's going to be your strength, your expertise. Don't get dispersed too much. I've dispersed myself quite a bit in my career. I continue to.

Karthik Ramakrishnan [00:43:11]

But that's the fun part, that you can afford to do that now. But in the early days, to your point, focus is everything, right? And understanding the fundamental building blocks of what you're working on. Otherwise you, without a strong foundation, you won't be able to build something great on top of that.

Dr. Yoshua Bengio [00:43:25]

Exactly. Always asking yourself, why am I doing this? Why are they doing this? Is this just some recipe? But okay, it works. But why? And sometimes we don't know the answer. But you should try at least to get the bottom thing.

Karthik Ramakrishnan [00:43:41]

Doctor Bengio, thank you so much for sharing all of that. Your work, what you see in the future, and really tangible advice for someone in this research space. So thank you so much.

Dr. Yoshua Bengio [00:43:49]

My pleasure. And thanks for the great questions.