Chapter 2. YOSHUA BENGIO

SCIENTIFIC DIRECTOR, MONTREAL INSTITUTE FOR LEARNING ALGORITHMS AND PROFESSOR OF COMPUTER SCIENCE AND OPERATIONS RESEARCH, UNIVERSITY OF MONTREAL

Yoshua Bengio is a professor of computer science and operations research at the University of Montreal and is widely recognized as one of the pioneers of deep learning. Yoshua was instrumental in advancing neural network research, in particular “unsupervised” learning where neural networks can learn without relying on vast amounts of training data.

MARTIN FORD: You are at the forefront of AI research, so I want to begin by asking what current research problems you think we’ll see breakthroughs in over the next few years, and how those will help us on the road to AGI (artificial general intelligence)?

YOSHUA BENGIO: I don’t know exactly what we’re going to see, but I can tell you that there are some really hard problems in front of us and that we are far from human-level AI. Researchers are trying to understand what the issues are, such as, why is it that we can’t build machines that really understand the world as well as we do? Is it just that we don’t have enough training data, or is it that we don’t have enough computing power? Many of us think that we are also missing the basic ingredients needed, such as the ability to understand causal relationships in data—an ability that actually enables us to generalize and to come up with the right answers in settings that are very different from those we’ve been trained in.

A human can imagine themselves going through an experience that is completely new to them. You might have never had a car accident, for example, but you can imagine one and because of all the things you already know you’re actually able to roleplay and make the right decisions, at least in your head. Current machine learning is based on supervised learning, where a computer essentially learns about the statistics of the data that it sees, and it needs to be taken through that process by hand. In other words, humans have to provide all of those labels, possibly hundreds of millions of correct answers, that the computer can then learn from.

A lot of current research is in areas where we’re not doing so well, such as unsupervised learning. This is where the computer can be more autonomous in the way that it acquires knowledge about the world. Another area of research is in causality, where the computer can not only observe data, like images or videos, but also act on it and see the effect of those actions in order to infer causal relationships in the world. The kinds of things that DeepMind, OpenAI, or Berkeley are doing with virtual agents, for example, are going in the right direction to answer those types of questions, and we’re also doing these kinds of things in Montreal.

MARTIN FORD: Are there any particular projects that you would point to as being really at the forefront of deep learning right now? The obvious one is AlphaZero, but what other projects really represent the leading edge of this technology?

YOSHUA BENGIO: There are a number of interesting projects, but the ones that I think are likely in the long run to have a big impact are those that involve virtual worlds in which an agent is trying to solve problems and is trying to learn about their environment. We are working on this at MILA, and there are projects in the same area in progress at DeepMind, OpenAI, Berkeley, Facebook and Google Brain. It’s the new frontier.

It’s important to remember, though, that this is not short-term research. We’re not working on a particular application of deep learning, instead we’re looking into the future of how a learning agent makes sense of its environment and how a learning agent can learn to speak or to understand language, in particular what we call grounded language.

MARTIN FORD: Can you explain that term?

YOSHUA BENGIO: Sure, a lot of the previous effort in trying to make computers understand language has the computer just read lots and lots of text. That’s nice and all, but it’s hard for the computer to actually get the meaning of those words unless those sentences are associated with real things. You might link words to images or videos, for example, or for robots that might be objects in the real world.

There’s a lot of research in grounded language learning now trying to build an understanding of language, even if it’s a small subset of the language, where the computer actually understands what those words mean, and it can act in correspondence to those words. It’s a very interesting direction that could have a practical impact on things like language understanding for dialog, personal assistants, and so on.

MARTIN FORD: So, the idea there is basically to turn an agent loose in a simulated environment and have it learn like a child?

YOSHUA BENGIO: Exactly, in fact, we want to take inspiration from child development scientists who are studying how a newborn goes through a series of stages in the first few months of life where they gradually acquire more understanding about the world. We don’t completely understand which part of this is innate or really learned, and I think this understanding of what babies go through can help us design our own systems.

One idea I introduced a few years ago in machine learning that is very common in training animals is curriculum learning. The idea is that we don’t just show all the training examples as one big pile in an arbitrary order. Instead, we go through examples in an order that makes sense for the learner. We start with easy things, and once the easy things are mastered, we can use those concepts as the building blocks for learning slightly more complicated things. That’s why we go through school, and why when we are 6 years old we don’t go straight to university. This kind of learning is becoming more important in training computers as well.

MARTIN FORD: Let’s talk about the path to AGI. Obviously, you believe that unsupervised learning—essentially having a system learn like a person—is an important component of it. Is that enough to get to AGI, or are there other critical components and breakthroughs that have to happen for us to get there?

YOSHUA BENGIO: My friend Yann LeCun has a nice metaphor that describes this. We’re currently climbing a hill, and we are all excited because we have made a lot of progress on climbing the hill, but as we approach the top of the hill, we can start to see a series of other hills rising in front of us. That is what we see now in the development of AGI, some of the limitations of our current approaches. When we were climbing the first hill, when we were discovering how to train deeper networks, for example, we didn’t see the limitations of the systems we were building because we were just discovering how to go up a few steps.

As we reach this satisfying improvement that we are getting in our techniques—we reach the top of the first hill—we also see the limitations, and then we see another hill that we have to climb, and once we climb that one we’ll see another one, and so on. It’s impossible to tell how many more breakthroughs or significant advances are going to be needed before we reach human-level intelligence.

MARTIN FORD: How many hills are there? What’s the timescale for AGI? Can you give me your best guess?

YOSHUA BENGIO: You won’t be getting that from me, there’s no point. It’s useless to guess a date because we have no clue. All I can say is that it’s not going to happen in the next few years.

MARTIN FORD: Do you think that deep learning or neural networks generally are really the way forward?

YOSHUA BENGIO: Yes, what we have discovered in terms of the scientific concepts that are behind deep learning and the years of progress made in this field, means that for the most part, many of the concepts behind deep learning and neural networks are here to stay. Simply put, they are incredibly powerful. In fact, they are probably going to help us better understand how animal and human brains learn complex things. As I said, though, they’re not enough to get us to AGI. We’re at a point where we can see some of the limitations in what we currently have, and we’re going to improve and build on top of that.

MARTIN FORD: I know that the Allen Institute for AI is working on Project Mosaic, which is about building common sense into computers. Do you think that kind of thing is critical, or do you think that maybe common sense emerges as part of the learning process?

YOSHUA BENGIO: I’m sure common sense will emerge as part of the learning process. It won’t come up because somebody sticks little bits of knowledge into your head, that’s not how it works for humans.

MARTIN FORD: Is deep learning the primary way to get us to AGI, or do you think it’s going to require some sort of a hybrid system?

YOSHUA BENGIO: Classical AI was purely symbolic, and there was no learning. It focused on a really interesting aspect of cognition, which is how we sequentially reason and combine pieces of information. Deep learning neural networks, on the other hand, have always been about focusing on a sort of bottom-up view of cognition, where we start with perception and we anchor the machine’s understanding of the world in perception. From there, we build distributed representations and can capture the relationship between many variables.

I studied the relationships between such variables with my brother around 1999. That gave rise to a lot of the recent progress in natural language, such as word embeddings, or distributed representations for words and sentences. In these cases, a word is represented by a pattern of activity in your brain—or by a set of numbers. Those words that have a similar meaning are then associated with similar patterns of numbers.

What’s going on now in the deep learning field is that people are building on top of these deep learning concepts and starting to try to solve the classical AI problems of reasoning and being able to understand, program, or plan. Researchers are trying to use the building blocks that we developed from perception and extend them towards these higher-level cognitive tasks (sometimes called System 2 by psychologists). I believe in part that’s the way that we’re going to move towards human-level AI. It’s not that it’s a hybrid system; it’s like we’re trying to solve some of the same problems that classical AI was trying to solve but using the building blocks coming from deep learning. It’s a very different way of doing it, but the objectives are very similar.

MARTIN FORD: Your prediction, then, is that it’s all going to be neural networks, but with different architectures?

YOSHUA BENGIO: Yes. Note that your brain is all neural networks. We have to come up with different architectures and different training frameworks that can do the kinds of things that classical AI was trying to do, like reasoning, inferring an explanation for what you’re seeing and planning.

MARTIN FORD: Do you think it can all be done with learning and training or does there need to be some structure there?

YOSHUA BENGIO: There is structure there, it’s just that it’s not the kind of structure that we use to represent knowledge when we write an encyclopedia, or we write a mathematical formula. The kind of structure that we put in corresponds to the architecture of the neural net, and to fairly broad assumptions about the world and the kind of task that we’re trying to solve. When we put in a special structure and architecture that allows the network to have an attention mechanism, it’s putting in a lot of prior knowledge. It turns out that this is central to the success of things like machine translation.

You need that kind of tool in your toolbox in order to solve some of those problems, in the same way that if you deal with images, you need to have something like a convolutional neural network structure in order to do a good job. If you don’t put in that structure, then performance is much worse. There are already a lot of domain-specific assumptions about the world and about the function you’re trying to learn, that are implicit in the kind of architectures and training objectives that are used in deep learning. This is what most of the research papers today are about.

MARTIN FORD: What I was trying to get at with the question on structure was that, for example, a baby can recognize human faces right after it is born. Clearly, then, there is some structure in the human brain that allows the baby to do that. It’s not just raw neurons working on pixels.

YOSHUA BENGIO: You’re wrong! It is raw neurons working on pixels, except that there is a particular architecture in the baby’s brain that recognizes something circular with two dots inside it.

MARTIN FORD: My point is that the structure pre-exists.

YOSHUA BENGIO: Of course it does, but all the things that we’re designing in those neural networks also pre-exist. What deep learning researchers are doing is like the work of evolution, where we’re putting in the prior knowledge in the form of both the architecture and the training procedure.

If we wanted, we could hardwire something that would allow the network to recognize a face, but it’s useless for an AI because they can learn that very quickly. Instead, we put in the things that are really useful for solving the harder problems that we’re trying to deal with.

Nobody is saying that there is no innate knowledge in humans, babies, and animals, in fact, most animals have only innate knowledge. An ant doesn’t learn much, it’s all like a big, fixed program, but as you go higher up in the intelligence hierarchy, the share of learning keeps increasing. What makes humans different from many other animals is how much we learn versus how much is innate at the start.

MARTIN FORD: Let’s step back and define some of those concepts. In the 1980s, neural networks were a very marginalized subject and they were just one layer, so there was nothing deep about them. You were involved in transforming that into what we now call deep learning. Could you define, in relatively non-technical terms, what that is?

YOSHUA BENGIO: Deep learning is an approach to machine learning. While machine learning is trying to put knowledge into computers by allowing computers to learn from examples, deep learning is doing it in a way that is inspired by the brain.

Deep learning and machine learning are just a continuation of that earlier work on neural networks. They’re called “deep” because they added the ability to train deeper networks, meaning they have more layers, and each layer represents a different level of representation. We hope that as the network gets deeper, it can represent more abstract things, and so far, that does seem to be the case.

MARTIN FORD: When you say layers, do you mean layers of abstraction? So, in terms of a visual image, the first layer would be pixels, then it would be edges, followed by corners, and then gradually you would get all the way up to objects?

YOSHUA BENGIO: Yes, that’s correct.

MARTIN FORD: If I understand correctly, though, the computer still doesn’t understand what that object is, right?

YOSHUA BENGIO: The computer has some understanding, it’s not a black-and-white argument. A cat understands a door, but it doesn’t understand it as well as you do. Different people have different levels of understanding of the many things around them, and science is about trying to deepen our understanding of those many things. These networks have a level of understanding of images if they’ve been trained on images, but that level is still not as abstract and as general as ours. One reason for this is that we interpret images in the context of our three-dimensional understanding of the world, obtained thanks to our stereo vision and our movements and actions in the world. This gives us a lot more than just a visual model: it also gives us a physical model of objects. The current level of computer understanding of images is still primitive but it’s still good enough to be incredibly useful in many applications.

MARTIN FORD: Is it true that the thing that has really made deep learning possible is backpropagation? The idea that you can send the error information back through the layers, and adjust each layer based on the final outcome.

YOSHUA BENGIO: Indeed, backpropagation has been at the heart of the success of deep learning in recent years. It is a method to do credit assignment, that is, to figure out how internal neurons should change to make the bigger network behave properly. Backpropagation, at least in the context of neural networks, was discovered in the early 1980s, at the time when I started my own work. Yann LeCun independently discovered it around the same time as Geoffrey Hinton and David Rumelhart. It’s an old idea, but we didn’t practically succeed in training these deeper networks until around 2006, over a quarter of a century later.

Since then, we’ve been adding a number of other features to these networks, which are very exciting for our research into artificial intelligence, such as attention mechanisms, memory, and the ability to not just classify but also generate images.

MARTIN FORD: Do we know if the brain does something similar to backpropagation?

YOSHUA BENGIO: That’s a good question. Neural nets are not trying to imitate the brain, but they are inspired by some of its computational characteristics, at least at an abstract level.

You have to realize that we don’t yet have a full picture of how the brain works. There are many aspects of the brain that are not yet understood by neuroscientists. There are tons of observations about the brain, but we don’t know how to connect the dots yet.

It may be that the work that we’re doing in machine learning with neural nets could provide a testable hypothesis for brain science. That’s one of the things that I’m interested in. In particular, backpropagation up to now has mostly been considered something that computers can do, but not realistic for brains.

The thing is, backpropagation is working incredibly well, and it suggests that maybe the brain is doing something similar—not exactly the same, but with the same function. As a result of that, I’m currently involved in some very interesting research in that direction.

MARTIN FORD: I know that there was an “AI Winter” where most people had dismissed deep learning, but a handful of people, like yourself, Geoffrey Hinton, and Yann LeCun, kept it alive. How did that then evolve to the point where we find ourselves today?

YOSHUA BENGIO: By the end of the ‘90s and through the early 2000s, neural networks were not trendy, and very few groups were involved with them. I had a strong intuition that by throwing out neural networks, we were throwing out something really important.

Part of that was because of something that we now call compositionality: The ability of these systems to represent very rich information about the data in a compositional way, where you compose many building blocks that correspond to the neurons and the layers. That led me to language models, early neural networks that model text using word embeddings. Each word is associated with a set of numbers corresponding to different attributes that are learned autonomously by the machine. It didn’t really catch on at the time, but nowadays almost everything to do with modeling language from data uses these ideas.

The big question was how we could train deeper networks, and the breakthrough was made by Geoffrey Hinton and his work with Restricted Boltzmann Machines (RBMs). In my lab, we were working on autoencoders, which are very closely related to RBMs, and autoencoders have given rise to all kinds of models, such as generative adversarial networks. It turned out that by stacking these RBMs or autoencoders we are able to train deeper networks than we were able to before.

MARTIN FORD: Could you explain what an autoencoder is?

YOSHUA BENGIO: There are two parts to an autoencoder, an encoder and a decoder. The idea is that the encoder part takes an image, for example, and tries to represent it in a compressed way, such as a verbal description. The decoder then takes that representation and tries to recover the original image. The autoencoder is trained to do this compression and decompression so that it is as faithful as possible to the original.

Autoencoders have changed quite a bit since that original vision. Now, we think of them in terms of taking raw information, like an image, and transforming it into a more abstract space where the important, semantic aspect of it will be easier to read. That’s the encoder part. The decoder works backwards, taking those high-level quantities—that you don’t have to define by hand—and transforming them into an image. That was the early deep learning work.

Then a few years later, we discovered that we didn’t need these approaches to train deep networks, we could just change the nonlinearity. One of my students was working with neuroscientists, and we thought that we should try rectified linear units (ReLUs)—we called them rectifiers in those days—because they were more biologically plausible, and this is an example of actually taking inspiration from the brain.

MARTIN FORD: What did you learn from all of that?

YOSHUA BENGIO: We had previously used a sigmoid function to train neural nets, but it turned out that by using ReLUs we could suddenly train very deep nets much more easily. That was another big change that occurred around 2010 or 2011.

There is a very large dataset—the ImageNet dataset—which is used in computer vision, and people in that field would only believe in our deep learning methods if we could show good results on that dataset. Geoffrey Hinton’s group actually did it, following up on earlier work by Yann LeCun on convolutional networks—that is, neural networks which were specialized for images. In 2012, these new deep learning architectures with extra twists were used with huge success and showed a big improvement on existing methods. Within a couple of years, the whole computer vision community switched to these kinds of networks.

MARTIN FORD: So that’s the point at which deep learning really took off?

YOSHUA BENGIO: It was a bit later. By 2014, things were lining up for a big acceleration in the community for the take-up of deep learning.

MARTIN FORD: That’s when it transitioned from being centered in universities to being in the mainstream domain at places like Google, Facebook, and Baidu?

YOSHUA BENGIO: Exactly. The shift started slightly earlier, around 2010, with companies like Google, IBM, and Microsoft, who were working on neural networks for speech recognition. By 2012, Google had these neural networks on their Android smartphones. It was revolutionary for the fact that the same technology of deep learning could be used for both computer vision and speech recognition. It drove a lot of attention toward the field.

MARTIN FORD: Thinking back to when you first started in neural networks, are you surprised at the distance things have come and the fact that they’ve become so central to what large companies, like Google and Facebook, are doing now?

YOSHUA BENGIO: Of course, we didn’t expect that. We’ve had a series of important and surprising breakthroughs with deep learning. I mentioned earlier that speech recognition came around 2010, and then computer vision around 2012. A couple of years later, in 2014 and 2015, we had breakthroughs in machine translation that ended up being used in Google Translate in 2016. 2016 was also the year we saw the breakthroughs with AlphaGo. All of these things, among a number of others, were really not expected.

I remember back in 2014 I looked at some of our results in caption generation, where the computer is trying to come up with a caption for an image, and I was amazed that we were able to do that. If you had asked me just one year earlier if we’d be able to do that in a year, I would have said no.

MARTIN FORD: Those captions are pretty remarkable. Sometimes they’re way off the mark, but most of the time they’re amazing.

YOSHUA BENGIO: Of course, they’re way off sometimes! They’re not trained on enough data, and there are also some fundamental advances in basic research that need to be made for those systems to really understand an image and really understand language. We’re far away from achieving those advances, but the fact that they were able to reach the level of performance that they have was not something we expected.

MARTIN FORD: Let’s talk about your career. What was your own path into the field of AI?

YOSHUA BENGIO: When I was young, I would read a lot of science fiction, and I’m sure that had an impact on me. It introduced me to topics such as AI and Asimov’s Three Laws of Robotics, and I wanted to go to college and study physics and mathematics. That changed when my brother and I became interested in computers. We saved our money to buy an Apple IIe and then an Atari 800. Software was scarce in those days, so we learned to program them ourselves in BASIC.

I got so excited with programming that I went into computer engineering and then computer science for my Master’s and PhD. While doing my Master’s around 1985, I started reading some papers on early neural nets, including some of Geoffrey Hinton’s papers, and it was like love at first sight. I quickly decided that this was the subject I wanted to do my research in.

MARTIN FORD: Is there any particular advice you’d give to someone who wants to get into the field of being a deep learning expert or researcher?

YOSHUA BENGIO: Just jump in the water and start swimming. There’s a ton of information in the form of tutorials, videos, and open source libraries at all levels because there’s so much interest in this field. And there is the book I co-authored, called Deep Learning, which helps newcomers into the field and is available for free online. I see many undergrad students training themselves by reading lots and lots of papers, trying to reproduce those papers, and then applying to get into the labs which are doing this kind of research. If you’re interested in the area, there’s no better time to start than now.

MARTIN FORD: In terms of your career, one thing I noticed is that of the key people in deep learning, you’re the only one that remains entirely in the academic world. Most others are part-time at companies like Facebook or Google. What made you take that career pathway?

YOSHUA BENGIO: I’ve always valued academia and the freedom to work for the common good or the things that I believe would have more impact. I also value working with students both psychologically and in terms of the efficiency and productivity of my research. If I went into the industry, I would be leaving a lot of that behind.

I also wanted to stay in Montreal, and at that time, it was the case that going into the industry meant going to either California or New York. It was then that I thought that maybe we could build something in Montreal that could become a new Silicon Valley for AI. As a result, I decided to stay and create Mila, The Montreal Institute for Learning Algorithms.

Mila carries out basic research, and also plays a leadership role in the AI ecosystem in Montreal. This role involves working in partnership with the Vector Institute in Toronto, and Amii, in Edmonton, as part of the Canadian strategy to really push AI forward—in terms of science, in terms of the economy, and in terms of positive social impact.

MARTIN FORD: Since you mention it, let’s talk more about AI and the economy, and some of the risks there. I have written a lot about the potential for artificial intelligence to bring on a new Industrial Revolution, and potentially to lead to a lot of job losses. How do you feel about that hypothesis, do you think that it is overhyped?

YOSHUA BENGIO: No, I don’t think it’s overhyped. The part that is less clear is whether this is going to happen over a decade or three decades. What I can say is that even if we stop basic research in AI and deep learning tomorrow, the science has advanced enough that there’s already a huge amount of social and economic benefit to reap from it simply by engineering new services and new products from these ideas.

We also collect a huge amount of data that we don’t use. For example, in healthcare, we’re only using a tiny, tiny fraction of what is available, or of what will be available as even more gets digitized every day. Hardware companies are working hard to build deep learning chips that are soon going to be easily a thousand times faster or more energy-efficient than the ones we currently have. The fact that you could have these things everywhere around you, in cars and phones, is clearly going to change the world.

What will slow things down are things like social factors. It takes time to change the healthcare infrastructure, even if the technology is there. Society can’t change infinitely fast, even if the technology is moving forward.

MARTIN FORD: If this technology change does lead to a lot of jobs being eliminated, do you think something like a basic income would be a good solution?

YOSHUA BENGIO: I think a basic income could work, but we have to take a scientific view on this to get rid of our moral priors that say if a person doesn’t work, then they shouldn’t have an income. I think it’s crazy. I think we have to look at what’s going to work best for the economy and what’s going to work best for people’s happiness, and we can do pilot experiments to answer those questions.

It’s not like there’s one clear answer, there are many ways that society could take care of the people who are going to be left behind and minimize the amount of misery arising from this Industrial Revolution. I’m going to go back to something that my friend Yann LeCun said: If we had had the foresight in the 19th century to see how the Industrial Revolution would unfold, maybe we could have avoided much of the misery that followed. If in the 19th century we had put in place the kind of social safety net that currently exists in most Western nations, instead of waiting until the 1940s and 1950s, then hundreds of millions of people would have led a much better and healthier life. The thing is, it’s going to take probably much less than a century this time to unfold that story, and so the potential negative impacts could be even larger.

I think it’s really important to start thinking about it right now and to start scientifically studying the options to minimize misery and optimize global well-being. I think it’s possible to do it, and we shouldn’t just rely on our old biases and religious beliefs in order to decide on the answer to these questions.

MARTIN FORD: I agree, but as you say, it could unfold fairly rapidly. It’s going to be a staggering political problem, too.

YOSHUA BENGIO: Which is all the more reason to act quickly!

MARTIN FORD: A valid point. Beyond the economic impact, what are the other things we should worry about in terms of artificial intelligence?

YOSHUA BENGIO: I have been very active in speaking against killer robots.

MARTIN FORD: I noticed you signed a letter aimed at a university in Korea which seemed to be headed towards research on killer robots.

YOSHUA BENGIO: That’s right, and this letter is working. In fact, KAIST, The Korea Advanced Institute of Science and Technology, has been telling us that they will avoid going into the development of military systems which don’t have a human in the loop.

Let me go back to this question about a human in the loop because I think this is really important. People need to understand that current AI—and the AI that we can foresee in the reasonable future—does not, and will not, have a moral sense or moral understanding of what is right and what is wrong. I know there are differences across cultures, but these moral questions are important in people’s lives.

It’s true, not just for killer robots but all kinds of other things, like the work that a judge does deciding on the fate of a person—whether that person should return to prison or be freed into society. These are really difficult moral questions, where you have to understand human psychology, and you have to understand moral values. It’s crazy to put those decisions in the hands of machines, which don’t have that kind of understanding. It’s not just crazy; it’s wrong. We have to have social norms or laws, which make sure that computers in the foreseeable future don’t get those kinds of responsibilities.

MARTIN FORD: I want to challenge you on that. I think a lot of people would say that you have a very idealistic view of human beings and the quality of their judgment.

YOSHUA BENGIO: Sure, but I’d rather have an imperfect human being as a judge than a machine that doesn’t understand what it’s doing.

MARTIN FORD: But think of an autonomous security robot that would be happy to take a bullet first and shoot second, whereas a human would never do that, and that could potentially save lives. In theory, an autonomous security robot would also not be racist, if it were programmed correctly. These are actually areas where it might have an advantage over a human being. Would you agree?

YOSHUA BENGIO: Well, it might be the case one day, but I can tell you we’re not there yet. It’s not just about precision, it’s about understanding the human context, and computers have absolutely zero clues about that.

MARTIN FORD: Other than the military and weaponization aspects, is there anything else that we should be worried about with AI?

YOSHUA BENGIO: Yes, and this is something that hasn’t been discussed much, but now may come more to the forefront because of what happened with Facebook and Cambridge Analytica. The use of AI in advertising or generally in influencing people is something that we should be really aware of as dangerous for democracy—and is morally wrong in some ways. We should make sure that our society prevents those things as much as possible.

In Canada, for example, advertising that is directed at children is forbidden. There’s a good reason for that: We think that it’s immoral to manipulate their minds when they are so vulnerable. In fact, though, every one of us is vulnerable, and if it weren’t the case, then advertising wouldn’t work.

The other thing is that advertising actually hurts market forces because it gives larger companies a tool to slow down smaller companies coming into their markets because those larger companies can use their brand. Nowadays they can use AI to target their message to people in a much more accurate way, and I think that’s kind of scary, especially when it makes people do things that may be against their well-being. It could be the case in political advertising, for example, or advertising that could change your behavior and have an impact on your health. I think we should be really, really careful about how these tools are used to influence people in general.

MARTIN FORD: What about the warnings from people like Elon Musk and Stephen Hawking about an existential threat from super intelligent AI and getting into a recursive improvement loop? Are these things that we should be concerned about at this point?

YOSHUA BENGIO: I’m not concerned about these things, I think it’s fine that some people study the question. My understanding of the current science as it is now, and as I can foresee it, is that those kinds of scenarios are not realistic. Those kinds of scenarios are not compatible with how we build AI right now. Things may be different in a few decades, I have no idea, but that is science fiction as far as I’m concerned. I think perhaps those fears are detracting from some of the most pressing issues that we could act on now.

We’ve talked about killer robots and we’ve talked about political advertising, but there are other concerns, like how data could be biased and reinforce discrimination, for example. These are things that governments and companies can act on now, and we do have some ways to mitigate some of these issues. The debate shouldn’t focus so much on these very long-term potential risks, which I don’t think are compatible with my understanding of AI, but we should pay attention to short-term things like killer robots.

MARTIN FORD: I want to ask you about the potential competition with China and other countries. You’ve talked a lot about, for example, having limitations on autonomous weapons and one obvious concern there is that some countries might ignore those rules. How worried should we be about that international competition?

YOSHUA BENGIO: Firstly, on the scientific side I don’t have any concern. The more researchers around the world are working on a science, the better it is for that science. If China is investing a lot in AI that’s fine; at the end of the day, we’re all going to take advantage of the progress that’s going to come of that research.

However, I think the part about the Chinese government potentially using this technology either for military purposes or for internal policing is scary. If you take the current state of the science and build systems that will recognize people, recognize faces, and track them, then essentially you can build a Big Brother society in just a few years. It’s quite technically feasible and it is creating even more danger for democracy around the world. That is really something to be concerned about. It’s not just states like China where this could happen, either; it could also happen in liberal democracies if they slip towards autocratic rule, as we have seen in some countries.

Regarding the military race to use AI, we shouldn’t confuse killer robots with the use of AI in the military. I’m not saying that we should completely ban the use of AI in the military. For example, if the military uses AI to build weapons that will destroy killer robots, then that’s a good thing. What is immoral is to have these robots kill humans. It’s not like we all have to use AI immorally. We can build defensive weapons, and that could be useful to stop the race.

MARTIN FORD: It sounds like you feel there’s definitely a role for regulation in terms of autonomous weapons?

YOSHUA BENGIO: There’s a role for regulation everywhere. In the areas where AI is going to have a social impact, then we at least have to think about regulation. We have to consider what the right social mechanism is that will make sure that AI is used for good.

MARTIN FORD: And you think governments are equipped to take on that question?

YOSHUA BENGIO: I don’t trust companies to do it by themselves because their main focus is on maximizing profits. Of course, they’re also trying to remain popular among their users or customers, but they’re not completely transparent about what they do. It’s not always clear that those objectives that they’re implementing correspond to the well-being of the population in general.

I think governments have a really important role to play, and it’s not just individual governments, it’s the international community because many of these questions are not just local questions, they’re international questions.

MARTIN FORD: Do you believe that the benefits to all of this are going to clearly outweigh the risks?

YOSHUA BENGIO: They’ll only outweigh the risks if we act wisely. That’s why it’s so important to have those discussions. That’s why we don’t want to move straight ahead with blinkers on; we have to keep our eyes open to all of the potential dangers that are lurking.

MARTIN FORD: Where do you think this discussion should be taking place now? Is it something primarily think tanks and universities should do, or do you think this should be part of the political discussion both nationally and internationally?

YOSHUA BENGIO: It should totally be part of the political discussion. I was invited to speak at a meeting of G7 ministers, and one of the questions discussed was, “How do we develop AI in a way that’s both economically positive and keeps the trust of the people?”, because people today do have concerns. The answer is to not do things in secret or in ivory towers, but instead to have an open discussion where everybody around the table, including every citizen, should be part of the discussion. We’re going to have to make collective choices about what kind of future we want, and because AI is so powerful, every citizen should understand at some level what the issues are.

YOSHUA BENGIO is Full Professor of the Department of Computer Science and Operations Research, scientific director of the Montreal Institute for Learning Algorithms (Mila), CIFAR Program co-director of the CIFAR program on Learning in Machines and Brains, Canada Research Chair in Statistical Learning Algorithms. Together with Ian Goodfellow and Aaron Courville, he wrote Deep Learning, one of the defining textbooks on the subject. The book is available for free from https://www.deeplearningbook.org.