Yoshua Bengio on intelligent machines (17-02-2016)

Canadian computer scientist Yoshua Bengio on artificial intelligence and how we can create thinking and learning machines through algorithms.

More videos with Yoshua Bengio

automatically generated captions
00:00:00 Yoshua: Yes, my name is Yoshua Bengio. And I am a professor here at the University of Montreal.
00:00:05 I also lead an institute called the Montreal Institute for Learning Algorithms,
00:00:10 that is specializing in my area of science, which is machine learning, how computers learn from examples.
00:00:18 Speaker 2: And what is the difference between, you say, machine learning?
00:00:27 Yoshua: Yes.
00:00:27 Speaker 2: But there's also this new thing called deep learning.
00:00:27 Yoshua: Right.
00:00:27 Speaker 2: What's the easiest way to,
00:00:31 Yoshua: Yes, so deep learning is inside machine learning, it's one of the approaches to machine learning.
00:00:40 Machine learning is very general, it's about learning from examples.
00:00:45 And scientists over the last few decades have proposed many approaches for allowing computers to learn from examples.
00:00:51 Deep learning is introducing a particular notion that the computer learns to represent information
00:01:03 and to do so at multiple levels of abstraction.
00:01:07 What I'm saying is a bit abstract, but to make it easier,
00:01:10 you could say that deep learning is also heavily inspired by what we know of the brain, of how neurons compute.
00:01:17 And it's a follow up on decades of earlier work, on what's called neural networks, or artificial neural networks.
00:01:26 Speaker 2: So, what is your background that you can relate to this?
00:01:31 Yoshua: I got interested in neural networks and machine learning, right at the beginning of my graduate studies.
00:01:40 So when I was doing my master's, I was looking for a subject
00:01:43 and I started reading some of these papers on neural networks.
00:01:46 And this was the early days of the so-called Connectionist Movement.
00:01:51 And I got really, really excited and I started reading more.
00:01:54 And I told the professor who was gonna supervise me that this is what I want to do.
00:02:02 And that's what I did, and I continued doing it and I'm still doing it.
00:02:04 Speaker 2: And do you think with your research, that you are on a route or a mainline, main thinking line,
00:02:19 which will get you somewhere?
00:02:22 Yoshua: So, say it's funny that you ask this question, cuz it depends.
00:02:24 It's like some days I feel very clearly that I know where I'm going and I can see very far.
00:02:34 I have the impression that I'm seeing far in the future and I see also where I've been and there's a very clear path.
00:02:44 And sometimes maybe I get more discouraged and I feel, where am I going? [LAUGH]
00:02:50 Yoshua: It's all exploration, I don't know where the future, what the future holds, of course.
00:02:56 So I go between these two states, which you need.
00:02:59 Speaker 2: Where are you now?
00:03:01 Yoshua: Right now I'm pretty positive about a particular direction.
00:03:11 I've moved to some fundamental questions that I find really exciting, and that's kind of driving a lot of my thinking,
00:03:21 looking forward.
00:03:22 Speaker 2: Can you tell me, I'm a not a scientist, most of our viewers are not as well.
00:03:27 But can you describe for me where you think your path leads to?
00:03:39 Because you sometimes you have a clear goal, you know where you're going.
00:03:42 Yoshua: Right.
00:03:42 Speaker 2: Where are you going?
00:03:43 Yoshua: So,
00:03:45 Yoshua: My main quest is to understand the principles that underlie intelligence.
00:03:53 And I believe that this happens through learning, that intelligent behavior arises in nature
00:03:59 and in the computers that we're building through learning.
00:04:02 The machine, the animal, the human becomes intelligent because it learns.
00:04:10 And understanding the underlying principles is like understanding the laws of aerodynamics for building airplanes,
00:04:19 right?
00:04:21 So I and others in my field are trying to figure out what is the equivalent of the laws of aerodynamics
00:04:29 but for intelligence.
00:04:31 So that's the quest, and we are taking inspiration from brains,
00:04:38 we're taking inspiration from a lot of our experiments that we're doing with computers trying to learn from data.
00:04:46 We're taking inspiration from other disciplines, from physics, from psychology, neuroscience.
00:05:00 And other fields, even electrical engineering, of course statistics, I mean, it's a very multi-disciplinary area.
00:05:09 Speaker 2: So you must have a clue?
00:05:11 Yoshua: Yes, I do. [LAUGH] So, one of the, well, may not be so easy to explain.
00:05:21 But one of the big mysteries about how brains manage to do what they do,
00:05:25 is what scientists have called for many decades the question of credit assignment.
00:05:30 That is, how do neurons in the middle of your brain, hidden somewhere, get to know how they should change.
00:05:38 What they should be doing that will be useful for the whole collective, that is, the brain.
00:05:44 And we don't know how brains do it, we now have algorithms that do a pretty good job at it.
00:05:52 They have their limitations
00:05:55 but one of the things I'm trying to do is to better understand this credit assignment question.
00:06:01 And it's crucial for deep learning, because deep learning is about having many levels of neurons talking to each other.
00:06:09 So that's why we call them deep, there are many layers of neurons. That's what gives them their power.
00:06:15 But the challenge is, how do we train them, how do they learn? And it gets harder the more layers you have.
00:06:23 So, in the 80s people found how train networks with a single, hidden layer.
00:06:30 So just not very deep, but they were already able to do interesting things.
00:06:35 And about ten years ago we started discovering ways to train much deeper networks,
00:06:40 and that's what led to this current revolution called deep learning.
00:06:44 Speaker 2: And this revolution, I didn't read it in the papers, so it's not front page news,
00:06:51 but for the science world it's a breakthrough.
00:06:53 Yoshua: Yes, so in the world of artificial intelligence there has been a big shift brought by deep learning.
00:07:04 So there has been some scientific advances but then it turned into advances in application.
00:07:11 So very quickly these techniques turned out to be very useful for improving how computers understand speech for example,
00:07:20 that speech recognition.
00:07:21 And then later much bigger, I would say, in terms of impact, effect fact happened
00:07:27 when we discovered that these algorithms could be very good for object recognition from images.
00:07:34 And now many other tasks in computer vision are being done using these kinds of networks.
00:07:40 These deep networks
00:07:41 or some specialized version of deep networks called convolutional networks that work well for images.
00:07:48 And then it moves on, so now people are doing a lot of work on natural language.
00:07:53 Trying to have the computer to understand English sentences, what you mean Being able to answer some questions
00:08:03 and so on. So these are applications but they have a huge economic impact and even more in the future.
00:08:11 That has attracted a lot of attention from other scientists, from the media,
00:08:20 and from of course business people who are investing billions of dollars into this right now.
00:08:26 Speaker 2: Yeah, is it exciting for you to be in the middle of this new development?
00:08:31 Yoshua: It is, it is very exciting and it's not something I had really expected.
00:08:37 Because ten years ago when we started working on this there were very few people in the world,
00:08:43 maybe a handful of people interested in these questions.
00:08:46 And initially it started very slowly we, it was difficult to get money for these kinds of things.
00:08:53 It was difficult to convince students to work on these kinds of things.
00:08:59 Speaker 2: Well maybe you can explain to me the ten years, or whatever,
00:09:05 12 years ago you were three people because it was not popular [CROSSTALK]
00:09:06 Yoshua: Right, that's right, that's right. Yes, that's right.
00:09:09 So there has been a decade before the last decade where this kind of research essentially went out of fashion.
00:09:18 People moved on to other interests.
00:09:20 They lost the ambition to actually get AI, to get machines to be as intelligent as us,
00:09:30 and also the connection between neuroscience and machine learning, it got divorced.
00:09:36 But a few people including myself and Jeff Hinton and Yann Lecun continued doing this and we started to have good results.
00:09:46 And other people in the world were also doing this and more people joined us.
00:09:53 And in a matter of about five years it started to be a more accepted area and then the applications,
00:10:02 the success in applications started to happen, and now it's crazy.
00:10:07 We get hundreds of applicants, for example, for doing grad studies here and companies are hiring like crazy
00:10:16 and buying scientists for their research labs.
00:10:20 Speaker 2: Do you notice that. Do they approach you as well?
00:10:24 Yoshua: Yeah.
00:10:25 Speaker 2: Big companies.
00:10:26 Yoshua: Yes. [LAUGH]
00:10:27 Yoshua: So I could be much richer. [LAUGH]
00:10:31 Yoshua: But I chose to stay in academia.
00:10:33 Speaker 2: So you've made some good thinking? And now it has become popular.
00:10:43 Yoshua: Yes.
00:10:44 Speaker 2: But, it has become valuable as well.
00:10:46 Yoshua: Yes, very valuable, yes.
00:10:49 Speaker 2: Why? Maybe-
00:10:51 Yoshua: Basically it's at the heart of what companies like Google, Microsoft, IBM, Facebook, Samsung, Amazon, Twitter.
00:11:05 Speaker 2: Why?
00:11:05 Yoshua: All of these companies they see this as a key technology for their future products
00:11:14 and some of the existing products already.
00:11:16 Speaker 2: And
00:11:17 Speaker 2: Are they right?
00:11:18 Yoshua: Yeah, they are. Of course, I don't have a crystal ball.
00:11:30 So there are a lot of research questions which remain unsolved, and it might take just a couple of years
00:11:38 or decades to solve them, we don't know. But even if, say, scientific research on the topic stopped right now.
00:11:44 And you took the current state of the arts in terms of the science, and you just applied it, right,
00:11:53 collecting lots of data sets because these items need a lot of data.
00:11:59 Just applying the current science would already have a huge impact on society.
00:12:04 So I don't think they're making a very risky bet,
00:12:10 but it could be even better because we could actually approach human level intelligence.
00:12:14 Speaker 2: You know that or you think so?
00:12:17 Yoshua: We could.
00:12:18 I think that we'll have other challenges to deal with and some of them we currently know are in front of us,
00:12:31 others we probably will discover when we get there.
00:12:34 Speaker 2: So now you're in the middle of a field of exciting research.
00:12:43 Yoshua: Yeah.
00:12:44 Speaker 2: That you know you're right and you have the goal and sometimes you see it clearly,
00:12:47 and it has become popular around people who want to study here.
00:12:48 Yoshua: Yep.
00:12:48 Speaker 2: And the companies want to invest in you.
00:12:50 Yoshua: Yes.
00:12:51 Speaker 2: So you must feel a lot of tension or a lot of-
00:12:55 Yoshua: It's true, it's true. Sudden.
00:12:58 Speaker 2: How does it feel to be in the middle of this development?
00:13:02 Yoshua: So initially it's exhilarating to have all this attention, and it's great to have all this recognition.
00:13:11 And also, its great to attract really the best minds that are coming here for doing PhD's and things like that.
00:13:21 It's absolutely great. But sometimes I feel that it's been too much, that I don't deserve that much attention.
00:13:30 And that all these interactions with media and so on are taking time away from my research
00:13:41 and I have to find the right balance here.
00:13:47 I think It is really important to continue to explain what we're doing so that more people can learn about it
00:13:55 and take advantage of it, or become researchers themselves in this area.
00:13:59 But I need to also focus my main strength which is not speaking to journalists.
00:14:07 My main strength is to come up with new ideas, crazy schemes, and interacting with students to build new things.
00:14:17 Speaker 2: Have you thought of the possibility that you're wrong?
00:14:20 Yoshua: Well, of course, science is an exploration. And I'm often wrong.
00:14:33 I propose ten things, nine of which end up not working.
00:14:39 But we make progress, so I get frequent positive feedback that tells me that we're moving in the right direction.
00:14:51 Speaker 2: If your right enough to go on.
00:14:53 Yoshua: Yes, yes, yes and these days because the number of people working on this has grown really fast,
00:15:01 the rate at which advances come is incredible.
00:15:06 The speed of progress in this field has greatly accelerated and mostly because there are more people doing it.
00:15:15 Speaker 2: And this is also reflected in what the companies do with it.
00:15:17 Yoshua: Yes, so companies are investing a lot in basic research in this field which is unusual.
00:15:25 Typically companies would invest in applied research where they take existing algorithms
00:15:31 and try to make them use them for products.
00:15:34 But right now there's a big war between these big IT companies to attract talent.
00:15:40 And also they understand that there is the potential impact,
00:15:47 the potential benefit of future research is probably even greater than what we have already achieved.
00:15:52 So for these two reasons, they have invested a lot in basic research and they are basically making offers to.
00:16:00 Professors
00:16:00 and students in the field to come work with them in an environment that looks a little bit like what you have in
00:16:07 universities where they have a lot of freedom, they can publish, they can go to conferences and talk with their peers.
00:16:13 So it's a good time for the progress of science because companies are working in the same direction as universities
00:16:20 towards really fundamental questions.
00:16:23 Speaker 2: But then they own it, that's the difference?
00:16:25 Yoshua: Yeah, that's something that's one of the reasons why I'm staying in academia.
00:16:32 I want to make sure that what I do is going to be, not owned by a particular person, but available for anyone.
00:16:40 Speaker 2: But is that the risk?
00:16:42 Is it really a risk that because the knowledge is owned by a company that, why would it be a risk?
00:16:49 Yoshua: I don't think it's a big deal right now, so the major research, industrial research centers,
00:17:04 they publish a lot of what they do.
00:17:10 And they do have patents, but they say that these patents are protective so in case somebody would sue them.
00:17:15 But they won't prevent other people, other companies using their technologies. At least that's what they say.
00:17:20 So right now there's a lot of openness in the business environment for this field.
00:17:30 We'll see how things are in the future.
00:17:32 There's always a danger of companies coming to a point where they become protective.
00:17:38 But then what I think is that companies who pull themselves out of the community,
00:17:43 and not participate to the scientific progress and exchange with the others. They will not progress as fast.
00:17:50 And I think that's the reason, they understand that, if they want to see the most benefits from this progress,
00:17:58 they have to be part of the public game of exchanging information and not keeping information secret.
00:18:04 Speaker 2: Part of the mind of the universe.
00:18:06 Yoshua: Yes, exactly. Part of the collective that we're building of all our ideas and our understanding of the world.
00:18:17 There is something about doing it personally into in that enables us to be more powerful and understanding.
00:18:24 If we're just trying to be consumers of ideas.
00:18:27 We're not mastering those ideas as well as if we're actually trying to improve them.
00:18:32 So
00:18:32 when we do research we get on top of things much more than if we're simply trying to understand some existing paper
00:18:43 and trying to use it for some product.
00:18:45 So there's something that is strongly enabling for companies to do that kind of thing, but that's new.
00:18:54 One decade ago for example many companies were shutting down their research labs and so on,
00:19:01 so it was a different spirit.
00:19:03 But right now, the spirit is openness, sharing, and participating in the common development of ideas through science
00:19:14 and publication and so on.
00:19:16 Speaker 2: It's funny that you said basic research is the same thing as fundamental research
00:19:22 Yoshua: Yes, yes. Yes.
00:19:24 Speaker 2: And that it becomes popular in some way.
00:19:27 Yoshua: Well, I think first of all it's appealing.
00:19:30 I mean as a person, I find researchers, PhD's candidate or professor or something.
00:19:39 It's much more appealing to me to know that what I do will be a contribution to humanity, right,
00:19:45 rather than something secret that only I and a few people would know about
00:19:49 and maybe some people will make a lot of money out of it that. I don't think it's as satisfying.
00:19:56 And as I said I think there are circumstances right now, that even from purely economic point of view,
00:20:01 is more interesting for companies to share right now. And be part of the research.
00:20:06 Speaker 2: So I think first to understand what you're really into I would like to know from you some basic definitions.
00:20:25 Yoshua: Yes.
00:20:26 Speaker 2: For example.
00:20:28 Speaker 2: What in your way of thinking is, and would you describe thinking?
00:20:37 Yoshua: Yes.
00:20:38 Speaker 2: What is thinking?
00:20:39 Yoshua: Right, well obviously we don't know. Because the brain-
00:20:43 Speaker 2: What do we don't know?
00:20:44 Yoshua: We don't know how the brain works. We have a lot of information about it.
00:20:51 Too much maybe, but not enough of the kind that allows us to figure out the basic principles of how we think,
00:20:59 and what does it mean at a very abstract level. But of course, I have my own understanding, so I can share that.
00:21:07 And with the kinds of equations I drew on the board there, and other people in my field.
00:21:16 There's this notion that what thinking is about is adjusting your mental configuration to be more coherent,
00:21:32 more consistent with everything you have observed, right?
00:21:38 And more typically, the things you're thinking about, or what you are currently observing.
00:21:44 So if I observe a picture, my neurons change their state to be in agreement with that picture and agreement,
00:21:53 given everything that the brain already knows, means that they are looking or an interpretation for that image.
00:21:59 Which may be related to things I could do that are related like I see this,
00:22:04 I need to go there because it tells me a message that matters to me.
00:22:08 So everything we know is somehow built in this internal model of the world that our brain has
00:22:14 and you get all these pieces of evidence each time we hear something, we listen to something
00:22:21 and our brain is actuating all of that stuff and then what it does is try to make sense of it,
00:22:30 reconcile the pieces like a piece of a puzzle. And so sometimes you know, it happens to you, something clicks right.
00:22:39 Suddenly you see a connection that explains different things.
00:22:44 Your brain does that all the time and not always that you get at this conscious impression, and thinking is this,
00:22:52 according to me, it's finding structure, and meaning, and the things that we observing and we've seen,
00:23:06 and that's also what science does, right?
00:23:08 Science is about finding explanations for what is around us,
00:23:13 but thinking it's happening in our head where science is a social thing.
00:23:18 Speaker 2: It's outside your head.
00:23:20 Yoshua: Science has a part inside.
00:23:25 Yeah, science has a part inside of course, because we are thinking when we do science. But science has a social aspect.
00:23:33 Science is a community of minds working together,
00:23:37 and the history of minds having discovered concepts that explain the world around us,
00:23:45 and sharing that in ways that are efficient. [talk in Dutch]
00:23:48 Yoshua: One thing I could talk about too is learning, right.
00:24:23 You asked me about thinking but I think a very important concept in my area is learning, I think.
00:24:37 I can explain how that can happen in those models or brains. [talk in Dutch] Yeah, yeah.
00:24:43 Speaker 2: Okay [Dutch] So you explained what thinking is. Now we'd like to know what is intelligence?
00:24:56 Yoshua: That's a good question. I don't think that there's a consensus on that either.
00:25:01 Speaker 2: On what?
00:25:02 Yoshua: On what is intelligence.
00:25:04 Speaker 2: If you reframe my question that I can.
00:25:07 Yoshua: Okay. So what is intelligence?
00:25:09 That's a good question and I don't think that there's a consensus but in my area of research people generally,
00:25:17 understand intelligence as the ability to take good decisions. And what good decisions.
00:25:24 Speaker 2: What's good?
00:25:24 Yoshua: Good for me. Right?
00:25:27 Speaker 2: Okay.
00:25:28 Yoshua: Good in the sense that they allow me to achieve my goals, to, If I was a animal to survive my predators,
00:25:37 to find food, to find mates. And for humans good might be achieving social status, or being happy, or whatever.
00:25:45 It's hidden in your mind. What is it that's good for you.
00:25:49 But somehow we are striving to take decisions that are good for us and, in order to do that,
00:25:58 it's very clear that we need some form of knowledge.
00:26:02 So, even a mouse that's choosing to go left or right in a maze is using knowledge,
00:26:10 and that kind of knowledge is not necessarily the kind of knowledge you find in the book, right?
00:26:16 A mouse cannot read a book, cannot write a book,
00:26:19 but in the mouse's brain there is knowledge about how to control the mouses' body in order to survive in order to find
00:26:28 food and so on. So the mouse is actually very intelligent in the context of the life of a mouse.
00:26:35 If you were suddenly teleported in a mouse, you would probably find it difficult to do the right things.
00:26:44 So, intelligence is about taking the right decision and it requires knowledge.
00:26:48 And now the question is to build intelligent machines or to understand how humans and animals are intelligent,
00:26:54 where are we getting the knowledge? Where can we get the knowledge?
00:26:59 And some of it is hard-wired in your brain from birth.
00:27:04 And some of it is going to be learned through experience, and that's the thing that we're studying in my field.
00:27:10 How do we learn or rather what are the mathematical principles for learning that could be applied to computers
00:27:18 and not just trying to figure out what animals, how animals learn.
00:27:22 Speaker 2: Then we get to the point the learning.
00:27:25 Yoshua: Right.
00:27:26 Speaker 2: So can you explain To me, because for everybody else, you think of learning, you learn at school?
00:27:38 Yoshua: Yeah.
00:27:39 Speaker 2: You read books, and there's someone telling you how the world works.
00:27:47 So what, in your concept, is the definition of learning?
00:27:50 Yoshua: Yes, my definition of learning is not the kind of learning that people think about when they're in school
00:27:57 and listening to a teacher. Learning is something we do all the time.
00:28:01 Our brain is changing all the time in response to what we're seeing, experiencing. And it's an adaptation.
00:28:08 And we are not just storing in our brain our experiences, it's not learning by heart, that's easy,
00:28:18 a file in a computer is like learning by heart. You can store facts.
00:28:23 But that's trivial, that's not what learning really is about.
00:28:28 Learning is about integrating the information we are getting through experience into some more abstract form that
00:28:38 allows us to take good decisions. That allow us to predict what will happen next.
00:28:43 That allow us to understand the connections between things we've seen. So, that's what's learning is really about.
00:28:52 In my field, we talk about the notion of generalization.
00:28:56 So, the machine can generalize from things it has seen and learned from, to new situations.
00:29:05 That's the kind of learning we talk about in my field.
00:29:09 And the way we typically do it in machines and how we think it's happening in the brain is that it's a slow,
00:29:17 gradual process. Each time you live an experience, one second of your life, there's gonna be some changes in your brain.
00:29:26 Small changes. So it's like your whole system is gradually shifting towards what would make it take better decisions.
00:29:38 So that's how you get to be intelligent, right?
00:29:40 Because you learn, meaning you changed the way you perceive and act, so that next time you would see something,
00:29:49 you will have some experience similar to what happened before, you would act better
00:29:54 or you would predict better what would have happened.
00:29:57 Speaker 2: So, it's very experienced based.
00:29:59 Yoshua: Yes, learning is completely experienced based.
00:30:03 Of course, in school we think of learning as, teaching knowledge from a book or some blackboard.
00:30:13 But, that's not the really the main kind of learning.
00:30:18 There is some learning happening when the student integrates all that information and tries to make sense of it.
00:30:23 But just storing those facts is kind of useless.
00:30:30 Speaker 2: It's a difference that you have to have an interest in it.
00:30:32 Yoshua: Well motivation for humans is very important. Because we are wired like this.
00:30:36 The reason we are wired like this is there are so many things happening around us that emotions help us to filter
00:30:45 and focus on some aspects more than others, those that matter to us, right?
00:30:50 That's a motivation, might be fear as well sometimes.
00:30:53 But for computers, basically they will learn what we ask them to learn, we don't need to introduce motivation
00:31:02 or emotions. These, up to now, we haven't needed to do that.
00:31:05 Speaker 2: But when you explain this deep learning.
00:31:08 Yoshua: Yes, yes.
00:31:12 Speaker 2: Maybe from the perspective of a machine and a human, you can learn computer experience
00:31:20 I think, but not interest or.
00:31:29 Yoshua: Well you can, emotions are something you're born with.
00:31:35 We're born with circuits that make us experience emotions because some situations matter more to us.
00:31:47 So, in the case of the computer, we also, in a sense,
00:31:51 hardwire these things by telling the computer Well this matters more than that
00:31:57 and you have to learn well to predict well here and here it matters less.
00:32:01 So we don't call that emotions but it could play a similar role.
00:32:06 Speaker 2: It looks like emotions.
00:32:08 Yoshua: Right.
00:32:08 Speaker 2: But then it's still program.
00:32:10 Yoshua: Absolutely so AI is completely programmed.
00:32:13 Speaker 2: Yeah.
00:32:14 But as I understand it well, you are reaching searching in this area where this program, which is beyond programming.
00:32:27 That they start to think for themselves.
00:32:27 Yoshua: Okay. So there's an interesting connection between learning and programming.
00:32:29 So the traditional way of putting knowledge into computers,
00:32:33 Is to write a program that essentially contains all our knowledge.
00:32:37 And step by step you tell the computer, if this happens you do this, and then you do that, and then you do that,
00:32:43 and then this happens you do that, and so on and so on. That's what a program is.
00:32:47 But when we allow the computer to learn we also program it, but the program that is there is different.
00:32:55 It's not a program that contains the knowledge we want a computer to have.
00:33:00 We don't program the computer with the knowledge of doors and cars and images and sounds.
00:33:05 We program the computer with the ability to learn and then the computer experiences.
00:33:11 You know, images, or videos, or sounds, or texts and learns the knowledge from those experiences.
00:33:21 So you can think of the learning program as a meta program and we have something like that in our brain.
00:33:28 If one part of your cortex dies you have an accident,
00:33:31 that part used to be doing some job like maybe interpreting music or some types of songs or something.
00:33:40 Well, if you continue listening to music then some other part will take over
00:33:47 and that function may have been sort of impaired for some time
00:33:52 but then it will be taken by some other part of your cortex. What does that mean?
00:33:57 It means that the same program that does the learning, works there in those two regions of your cortex.
00:34:04 The one that used to be doing the job, and the one that does it now.
00:34:08 And that means that your brain has this general purpose learning recipe that it can apply to different problems
00:34:19 and that this different parts of your brain will be specialized on different tasks.
00:34:25 Depending on what you do and which how the brain is connected.
00:34:29 If we remove that part of your brain then some other parts will start doing the job,
00:34:35 if the job is needed because you do those experiences, right?
00:34:39 So if I had a part of my brain that was essentially dealing with playing tennis and that part dies,
00:34:47 I'm not gonna be able to play tennis anymore. But if I continue practicing it's gonna come back.
00:34:55 And that means that the same learning, general purpose learning recipe is used everywhere at least in the cortex.
00:35:04 And this is important not just for understanding brains,
00:35:07 but for companies building products because we have this general purpose recipe
00:35:11 or family recipes that can be applied for many tasks.
00:35:17 The only thing that really differs between those different tasks is the data, the examples that the computer sees.
00:35:23 So that's why companies are so excited about this because they can use this for many problems that they wanna solve so
00:35:28 long as they can teach the machine by showing it examples.
00:35:30 Speaker 2: Is it always, is learning always positive?
00:35:34 Yoshua: Learning is positive by construction in the sense that it's moving the learner towards a state of understanding
00:35:51 of its experiences. So in general, yes, because learning is about improving something.
00:36:00 Now, if the something you're improving is not the thing you should be improving, you could be in trouble.
00:36:06 People can be trained into a wrong understanding of the world and they start doing bad things,
00:36:14 so that's why education is so important for humans.
00:36:18 And for machines right now the things we are asking the machines to do are very simple like understanding the content
00:36:25 of images and texts and videos and things like that.
00:36:26 Speaker 2: So learning is not per se positive because also you can learn wrong things.
00:36:30 Yoshua: Right but if you're just observing things around you and taken randomly then it's just what the world is right.
00:36:40 Speaker 2: And that's the state of the some kind of primitive learning of computers right now or?
00:36:45 Yoshua: Right now, yeah the learning the computers do is very primitive. It's mostly about perception.
00:36:53 And in the case of language some kind of semantic understanding, but it's still a pretty low level understanding.
00:37:00 Speaker 2: Is it possible for you to explain that in a simple way how is it possible for a computer to learn?
00:37:11 Yoshua: So the way that the computer is learning is by small iterative changes, right?
00:37:22 So let's go back to my artificial neural network, which is a bunch of neurons connected to each other,
00:37:29 and they're connected through these synaptic connections.
00:37:33 At each of these connections there is the strength of the connection which controls how a neuron influences another
00:37:39 neuron. So you can think that strength as a knob. And what happens during learning is those knobs change.
00:37:48 We don't know how they change in the brain, but in our algorithms, we know how they change.
00:37:52 And we understand mathematically why it makes sense to do that and they change little bit each time you see an example.
00:37:58 So i show the image of a cat but the computer says it's a dog.
00:38:03 So, I'm going to change those knobs so that it's going to be more likely that the computer is going to say cat.
00:38:10 Maybe the computer outputs a score for dog and a score for cat.
00:38:15 And so what we want to do is decrease the score for dog and increase the score for cat.
00:38:21 So that the computer, eventually, after seeing many millions of images, starts seeing the right class more often
00:38:32 and eventually gets it as well as humans.
00:38:35 Speaker 2: That still sounds like putting just enough data or less data for a computer to recognize something.
00:38:43 But how do you know that the computer is learning? How do you know
00:38:48 Yoshua: Well, you can test it on new images.
00:38:51 So if the computer was only learning by heart, copying the examples that it has seen,
00:38:56 it wouldn't be able to recognize a new image of say, new breed of dog, or a new angle, new lighting.
00:39:04 At the level of pixels, those images could be very, very different.
00:39:11 But, if the computer really figured catness,
00:39:15 at least from the point of view of images it will be able to recognize new images of new cats, taking on new postures
00:39:25 and so on and that's what we call generalization.
00:39:28 So we do that all the time, we test the computer to see if it can generalize to new examples, new images,
00:39:35 new sentences Can you show that to us, not right now but-
00:39:40 Speaker 2: Yeah. You can show that proof of learning skills.
00:39:43 Yoshua: Yeah, yeah I'll try to show you some examples of that, yeah.
00:39:49 Speaker 2: Great, so is there something, I'm missing that right now for understanding deep learning?
00:39:57 Yoshua: Yes.
00:39:59 Speaker 2: Okay, tell me.
00:40:00 Yoshua: I thought this was a statement, not a question.
00:40:04 Well, but yes, of course I [LAUGH] think there are many things that you are missing.
00:40:10 So there are many, many interesting questions in deep learning
00:40:14 but one of the interesting challenges has to do with the question of supervised learning versus unsupervised learning.
00:40:25 Right now, the way we teach the machine to do things
00:40:30 or to recognize things is we use what's called supervised learning where we tell the computer exactly what it should do
00:40:37 or what output it should have for a given input.
00:40:41 So let's say I'm showing it the image of a cat again, I tell the computer, this is a cat.
00:40:49 And I have to show it millions of such images.
00:40:53 That's not the way humans learn to see and understand the world or even understand language.
00:41:00 For the most part, we just make sense of what we observe without having a teacher that is sitting by us
00:41:11 and telling us every second of our life. This is a cow, this is a dog.
00:41:16 Speaker 2: A supervisor.
00:41:16 Yoshua: That's right. There is no supervisor.
00:41:19 We do get some feedback but it's pretty rare and sometimes it's only implicit.
00:41:25 So you do something and you get a reward but you don't know exactly what it was you did that gave you that reward.
00:41:37 Or you talk to somebody, the person is unhappy and you're not sure exactly what you did that was wrong
00:41:44 and the persons not gonna tell you in general what you should have done.
00:41:48 So this is called reinforcement learning when you get some feedback but it's a very weak type.
00:41:54 You did well or you didn't do well.
00:41:56 You have an exam and you achieved 65% but you don't know, if you don't know what the errors were
00:42:05 or what the right answers are it's very difficult to learn from that.
00:42:08 But we are able to learn from that, from very weak signals or no reinforcement at all, no feedback,
00:42:16 just by observation and trying to make sense of all of these pieces of information.
00:42:21 That's called unsupervised learning.
00:42:23 And we're not yet, we are much more advanced with supervised learning than with unsupervised learning.
00:42:32 So all of the products that these companies are building right now, it's mostly based on supervised learning.
00:42:38 Speaker 2: So the next step is unsupervised learning?
00:42:41 Yoshua: Yes, yes.
00:42:42 Speaker 2: Does that mean that unsupervised learning that the computer can think for themselves?
00:42:47 Yoshua: That means the computer will be more autonomous, in some sense. That we don't need.
00:42:56 Speaker 2: That's a hard one.
00:42:57 Yoshua: More autonomous.
00:42:59 Speaker 2: Autonomous computer?
00:43:00 Yoshua: Well more autonomous in its learning. We're are not talking about robots here, right?
00:43:04 We are just talking about computers gradually making sense of the world around us by observation.
00:43:13 And we probably will still need to give them some guidance, but the question is how much guidance.
00:43:20 Right now we have to give them a lot of guidance. Basically we have to spell everything very precisely for them.
00:43:27 So we're trying to move away from that so that they can essentially become more intelligent because they can take
00:43:34 advantage of all of the information out there which doesn't come with a human that explains every bits and pieces.
00:43:44 Speaker 2: But when a computer starts to learn.
00:43:47 Yoshua: Yes.
00:43:48 Speaker 2: Is it possible to stop the computer from learning? [LAUGH]
00:43:51 Yoshua: Sure.
00:43:54 Speaker 2: How? It sounds like if it starts to learn, then it learns.
00:43:58 Yoshua: It's just a program running. It's stored in files. There's nothing like, there's no robot.
00:44:05 There is no, I mean at least in the work we do,
00:44:08 it's just a program that contains files that contain those synaptic weights for example.
00:44:19 And as we see more examples we change those files so that they will correspond to taking the right decisions.
00:44:27 But there's no, those computers don't have a consciousness, there's no such thing right now, at least, for a while.
00:44:44 Speaker 2: Is it right when I say, well, deep learning or self learning computer is becoming more autonomous.
00:44:51 Yoshua: Autonomous in its learning, right?
00:44:54 Speaker 2: Yes, free.
00:44:56 Yoshua: Again, it's probably gonna be a gradual thing where the computer requires less and less of our guidance.
00:45:04 That we probably, so, If you think about humans, we still need guidance.
00:45:09 If you take a human baby nobody wants to do that experiment.
00:45:15 But you can imagine a baby being isolated from society. That child probably would not grow to be very intelligent.
00:45:24 Would not understand the world around us as well as we do. That's because we've had parents, teachers and so on, guide us.
00:45:34 And we've been immersed in culture.
00:45:37 So all that matters, and it's possible that it will also be required for computers to reach our level of intelligence.
00:45:44 The same kind of attention we're giving to humans, we might need to give to computers.
00:45:48 But right now, the amount of attention we have to give to computers for them to learn about very simple things,
00:45:53 is much larger than what we need to give to humans.
00:45:57 Humans are much more autonomous in their learning than machines are right now.
00:46:01 So we have a lot of progress to do in that direction.
00:46:04 Speaker 2: Is the difference also just the simple fact that we have biology?
00:46:09 Yoshua: Well biology is not magical. Biology is, can be understood.
00:46:17 It's what biologists are trying to do
00:46:19 and we understand a lot but there As far as the brain is concerned there's still big holes in our understanding.
00:46:25 Speaker 2: A baby grows but a computer doesn't.
00:46:28 Yoshua: Sure it can, we can give it more memory and so on right? So you can grow the size of the model.
00:46:40 That's not a big obstacle.
00:46:42 I mean computing power is an obstacle, but I'm pretty confident that over the next few years we're gonna see more
00:46:50 and more computing power available as it has been in the past,
00:46:55 that will make it more possible to train models to do more complex tasks.
00:47:00 Speaker 2: So how do you tackle all the people who think this is a horror scenario?
00:47:10 Of course, people start to think about growing computers and it's not about that.
00:47:15 Yoshua: So I think.
00:47:18 Speaker 2: You have to have a stand point.
00:47:20 Yoshua: That's right. I do. So first of all, I think there's been a bit of excessive expression of fear about AI.
00:47:34 Maybe because the progress has been so fast, it has made some people worried.
00:47:40 But if you ask people like me who are into it every day.
00:47:46 They're not worried, because they can see how stupid the machines are right now.
00:47:51 And how much guidance they need to move forward.
00:47:55 So to us, it looks like we're very far from human level intelligence
00:48:01 and even have no idea whether one day computers will be smarter than us. Now that may be a short term view.
00:48:11 What will happen in the future is hard to say, but we can think about it.
00:48:18 And I think it's good that some people are thinking about the potential dangers.
00:48:25 I think it's difficult right now to have a grasp on what could go wrong.
00:48:31 But with the kind of intelligence that we're building in machines right now, I'm not very worried.
00:48:37 It's not the kind of intelligence that I can foresee exploding, becoming more and more intelligent by itself.
00:48:46 I don't think that's plausible for the kinds of deep learning methods and so on.
00:48:51 Even if they were much more powerful and so on, it's not something I can envision.
00:48:56 That being said, it's good that there are people who are thinking about these long term issues.
00:49:02 One thing I'm more worried about is the use of technology now, or in the next couple of years or five or ten years.
00:49:11 Where the technology could be developed and used in a way that could either be very good for many people
00:49:19 or not so good for many people.
00:49:21 And so for example, military use and other uses, which I think I would consider not appropriate,
00:49:28 are things we need to worry about.
00:49:31 Speaker 2: All right, can you name examples of that?
00:49:35 Yoshua: Yeah, so there's been a fuss
00:49:37 and a letter signed by a number of scientists who tried to tell the world we should have a ban on the use of AI for
00:49:48 autonomous weapons that could essentially take the decision to kill by themselves.
00:49:53 So that's something that's not very far fetched in terms of technology and the given science.
00:49:59 Basically, the science is there, it's a matter of building these things.
00:50:03 But it's not something we would like to see, and there could be an arms race of these things.
00:50:09 So we need to prevent it, the same way that, collectively, the nations decided to have bans on biological weapons
00:50:18 and chemical weapons and, to some extent, on nuclear weapons. The same thing should be done for that.
00:50:25 And then there are other uses of this technology, especially as it matures,
00:50:30 which I think are questionable from an ethical point of view.
00:50:32 So I think that the use of these technologies to convince you to do things, like with publicity,
00:50:40 and trying to influence, maybe think about influencing your vote, right?
00:50:50 As the technology becomes really stronger,
00:50:53 you could imagine people essentially using this technology to manipulate you in ways you don't realize.
00:51:01 That is good for them, but is not good for you.
00:51:05 And I think we have to start being aware of that and all the issues of privacy are connected to that as well.
00:51:14 But in general, because we're training currently, companies are using these systems for advertisements.
00:51:21 Where they're trying to predict what they should show you, so that you will be more likely to buy some product, right?
00:51:29 So it seems not so bad, but if you push it, they might bring you into doing things that are not so good for you.
00:51:41 I don't know, like smoking or whatever, right?
00:51:45 Speaker 2: Well, we just stopped at a point where I was going to ask you about,.
00:51:54 is that why you wrote the manifest about diversity and thinking? Because I'll show you, [FOREIGN] Okay.
00:52:11 Speaker 2: Because computers,
00:52:12 Speaker 2: You can learn them a lot of things, but it's almost unimaginable that you can learn them diversity.
00:52:24 Am I correct that that has a connection?
00:52:26 Yoshua: If you want, I will elaborate now. So you're asking me about diversity,
00:52:34 Yoshua: And I can say several things.
00:52:40 First of all, people who are not aware of the kinds of things we do in AI, with machine learning, deep learning, and so on.
00:52:48 May not realize that the algorithms, the methods we're using already include a lot of what may look like diversity,
00:53:00 creativity. So for the same input, the computer could produce different answers.
00:53:05 And so there's a bit of randomness, just like for us.
00:53:08 Twice in the same situation, we don't always take the same decision.
00:53:12 And there are good reasons for that, both for us and for computers. So that's the first part of it.
00:53:17 But there's another aspect of diversity, which I have studied in a paper a few years ago,
00:53:23 which is maybe even more interesting. Diversity is very important, for example, for evolution to succeed.
00:53:35 Because evolution performs a kind of search in the space of genomes of the blueprint of each individual.
00:53:46 Yoshua: And up to now, machine learning is considered what happens in a single individual, how we learn,
00:53:57 how a machine can learn.
00:53:59 But has not really investigated much the role of having a group of individuals learning together, so a kind of society.
00:54:09 And in this paper a few years ago, I postulated that learning in an individual could get stuck.
00:54:19 That if we were alone learning by observing the world around us, we might get stuck with a poor model of the world.
00:54:26 And we get unstuck by talking to other people and by learning from other people,
00:54:32 in the sense of they can communicate some of the ideas they have, how they interpret the world.
00:54:39 And that's what culture is about. Culture has many meanings, but that's the meaning that I have.
00:54:45 That it's not just the accumulation of knowledge, but how knowledge gets created through communication and sharing.
00:54:54 Yoshua: And what I postulated in that paper is that there is a, it's called an optimization problem,
00:55:02 that can get the learning of an individual to not progress anymore.
00:55:08 In a sense that, as I said before, learning is a lot of small changes,
00:55:13 but sometimes there's no small change that really makes you progress.
00:55:19 So you need some kind of external kick that brings a new light to things.
00:55:25 And another connection to evolution, the connection to evolution, actually,
00:55:32 is that this small kick we get from others is like we are building on top of existing solutions that others have come
00:55:42 up with. And of course, the process of science is very much like this. We're building on other scientists' ideas.
00:55:47 But it's true for culture, in general.
00:55:50 And this actually makes the whole process of building more intelligent beings much more efficient.
00:55:59 In fact, we know that since humans have made progress, thanks to evolution and not just. thanks to culture
00:56:10 and not just to evolution, we've been making. our intelligence has been increasing much faster.
00:56:18 So, evolution is slow whereas you can think of culture,
00:56:24 the evolution of culture as a process that's much more efficient. Because we are manipulating the right objects.
00:56:31 So what does this mean in practice?
00:56:33 It means that just like evolution needs diversity to succeed, because there are many different.
00:56:40 Variants of the same type of genes that are randomly chosen and tried,
00:56:49 and the best ones combine together to create new solutions just like this in cultural evolution.
00:56:56 Which is really important for our intelligence as I was saying, we need diversity,
00:57:02 we need not just one school of thought, we need to allow all kinds of exploration, most of which made fail.
00:57:09 So, in science we need to be open to new ideas, even if it's very likely it's not gonna work,
00:57:16 it's good that people explore, otherwise we're gonna get stuck.
00:57:20 In some, in the space of possible interpretations of the world, it may take forever before we escape.
00:57:27 Speaker 2: It is like doing basic research but you don't have-
00:57:31 Yoshua: Yes.
00:57:33 Speaker 2: A specific goal.
00:57:33 Yoshua: That's right so basic research is exploratory, it's not trying to build a product.
00:57:38 It's just trying to understand and it's going in all possible directions.
00:57:43 According to our intuitions of what may be more interesting but without a strong constraint.
00:57:48 So, yeah basic research is like this, but there's a danger because humans they like fashionable things, and trends,
00:58:00 and compare each other, and so on, that we're not giving enough freedom for exploration.
00:58:10 And it's not just science, it's in general, right in society we should allow a lot more freedom.
00:58:15 We should allow marginal ways of being and doing things to coexist.
00:58:22 Speaker 2: But if you allow this freedom,
00:58:23 of course most people think well let's don't go that way because then you have autonomous, self-thinking computers
00:58:27 Speaker 2: Creating their own diversity,
00:58:28 and so there are a lot of scenarios which people think of because they don't know, and which scare them, so this.
00:58:46 Yoshua: Well, it's a gamble and I'm more on the positive side.
00:58:52 I think that the rewards we can get by having more intelligence in our machines is immense.
00:58:59 And the way I think about it is, it's not a competition between machines and humans.
00:59:05 Technology is expanding what we are, thanks to technology we're now already much stronger
00:59:14 and more intelligent than we were.
00:59:18 in the same way that the industrial revolution has kinda increased our strength and our ability to do things physically.
00:59:26 The sort of computer revolution and now the AI revolution is gonna increase, continue to increase our cognitive abilities.
00:59:34 Speaker 2: That sounds very logical, but I can imagine you must get tired of all those people who don't,
00:59:41 who fear this development.
00:59:42 Yoshua: Right,
00:59:44 but I think we should be conscious that a lot of that fear is due to a projection into things we are familiar with.
00:59:53 So, we are thinking of AI like we see them in movies,
00:59:58 we're thinking of AI like we see some kind of alien from another planet, like we see animals.
01:00:03 When we think about another being, we think that other being is like us and so we're greedy.
01:00:10 We want to dominate the rest and if our survival is at stake, we're ready to kill right.
01:00:16 So, we project that some machine is gonna be just like us, and if that machine is more powerful then we are,
01:00:25 then we're in deep trouble, right?
01:00:26 So, it's just because we are making that projection, but actually the machines are not some being that has an ego
01:00:35 and a survival instinct. It's actually something we decide to put together.
01:00:41 It's a program and so we should be smart enough
01:00:44 and wise enough to program these machines to be useful to us rather than go towards the wrong needs.
01:00:52 They will cater to our needs because we will design them that way.
01:00:56 Speaker 2: I understand that, but then there's also this theory of suppose you can develop machines
01:01:03 or robots that can self-learn. So, if that grows with this power of.
01:01:17 Yoshua: Yes.
01:01:18 Speaker 2: There is some acceleration in their intelligence or that's.
01:01:27 Yoshua: Maybe, maybe not, I don't, that's not the way I,
01:01:32 what you're saying is appealing if I was to read a science fiction book.
01:01:36 But it doesn't correspond to how I see AI, and the kind of AI we're doing, I don't see such acceleration,
01:01:47 in fact what I see is the opposite. What I foresee is more like barriers than acceleration. So our-
01:01:56 Speaker 2: Slowing you down?
01:01:57 Yoshua: Yes, so our experience in research is that we make progress.
01:02:01 And then we encounter a barrier, a difficult challenge, a difficulty, the algorithm goes so far
01:02:07 and then can't make progress. Even if we have more computer power, that's not really the issue.
01:02:13 The issue are more, are basically computer science issue that things get Harder as you try to solve,
01:02:21 exponentially harder, meaning much, much harder as you try to solve more complex problems.
01:02:27 So, it's actually the opposite I think that happens that.
01:02:30 And I think that would also explain maybe to some extent why we're not super intelligent ourselves.
01:02:37 I mean, the sense that our intelligence is kind of limited. There are many things for which we make the wrong decision.
01:02:44 And then it's true also of animals.
01:02:46 Why is it like that some animals have much larger brains than we do and they're not that smart?
01:02:54 You could come up with a bunch of reasons but it's not they have a bigger brain.
01:02:59 And their brain, a mammal's brain is very very close to ours. So it's hard to say.
01:03:08 Now I think it's fair to consider the worst scenarios and to study it
01:03:14 and have people seriously considering what could happen and how we could prevent any dangerous thing.
01:03:21 I think it's actually important that some people do that.
01:03:24 But, right now I see this as a very long term potential, and the most plausible scenario is not that,
01:03:31 according to my vision.
01:03:32 Speaker 2: Does it have to do with the fact that you tried to develop this deep learning That if you know how it works,
01:03:42 then you also know how to deal with it. Is that why you are confident in not seeing any problem?
01:03:49 Yoshua: You're right that I think we are more afraid of things we don't understand.
01:03:54 And scientists who are working with deep learning everyday don't feel that they have anything to fear because they
01:04:03 understand what's going on.
01:04:04 And they can see clearly that there is no danger that's foreseeable, so you're right that's part of it.
01:04:11 There's the psychology of seeing the machine as some other being. There's the lack of knowledge.
01:04:17 There's influence of science fiction.
01:04:18 So all these factors come together and also the fact that the technology has been making a lot of progress recently.
01:04:23 So all of that I think creates kind of an exaggerated fear.
01:04:27 I'm not saying we shouldn't have any fear I'm just saying it's exaggerated right now.
01:04:31 Speaker 2: Is your main part of life, or your, how you fill the day, is it thinking? Is your work thinking?
01:04:52 What do you physically do?
01:04:54 Yoshua: I'm thinking all the time, yes.
01:04:57 And whether I'm thinking on the things that matter to me the most, maybe not enough.
01:05:03 Managing a big institute, with a lot of students, and so on, means my time is dispersed, but.
01:05:10 When I can focus, or when I'm in a scientific discussion with people, and so on.
01:05:18 Of course there's a lot of thinking, and it's really important, that's how we move forward.
01:05:24 Speaker 2: Yeah, what does that mean? The first question I asked you was about what is thinking.
01:05:30 Yoshua: Yes.
01:05:31 Speaker 2: And now we are back to that question.
01:05:33 Yoshua: Yeah, yeah, so, so.
01:05:34 Speaker 2: You are a thinker so what happens.
01:05:40 Yoshua: Okay.
01:05:40 Speaker 2: During the day?
01:05:41 Yoshua: Yes.
01:05:42 Speaker 2: With you?
01:05:43 Yoshua: So when I listen to somebody explaining something.
01:05:48 Maybe one of my students talking about an experiment, or another researcher talking about their idea.
01:05:55 Something builds up in my mind to try to understand what is going on.
01:06:03 And that's already thinking but then things happen so other pieces of information and understanding connect to this.
01:06:12 And I see some flaw or some connection and that's where the creativity comes in.
01:06:24 And how I have the impulse of talking about it. And that's just one turn in a discussion. And we go like this. And,
01:06:39 Yoshua: New ideas spring like this. And it's very, very rewarding.
01:06:43 Speaker 2: Is it possible for you not to think?
01:06:47 Yoshua: Well, yes. Yes, it is possible not to think.
01:06:56 It's hard, but if you really relax or you are experiencing something very intensely,
01:07:06 then you're not into your thoughts, you're just into some present-time experience.
01:07:18 Speaker 2: Like it's more emotional rather than rational?
01:07:22 Yoshua: For example, yes, but thinking isn't just rational.
01:07:28 A lot of it is, I don't mean it's irrational,
01:07:31 but a lot of the thinking is something that happens somehow behind the scenes.
01:07:36 It has to do with intuition that has to do with analogies and it's not necessarily a causes b causes c.
01:07:49 It's not that kind of logical thinking that's going on in my mind most of the time.
01:07:54 It's much softer and that's why we need the math in order to filter and fine tune the ideas,
01:08:03 but the raw thinking is very fuzzy. But it's very rich because it's connecting a lot of things together.
01:08:15 And it's discovering the inconsistencies that allow us to move to the next stage and solve problems.
01:08:26 Speaker 2: Are you aware of that you are in that situation when you are thinking?
01:08:34 Yoshua: It happens to me.
01:08:36 I used to spend some time meditating and there you're learning to pay attention to your own thoughts.
01:08:46 So it does happen to me.
01:08:50 It happens to me also that I get so immersed in my thoughts in ordinary,
01:08:54 daily activities that people think that I'm very distracted and not present and they can be offended. [LAUGH]
01:09:02 Yoshua: But it's not always like this, sometimes I'm actually very, very present.
01:09:08 I can be very, very present with somebody talking to me and that's really important for my job, right?
01:09:16 Because if I listen to somebody in a way that's not complete, I can't really understand fully
01:09:29 and participate in a rich exchange.
01:09:34 Speaker 2: I can imagine that when you are focused on a thought.
01:09:39 Or you were having this problem and you're thinking about it, thinking about it.
01:09:40 And then you are in this situation that other people they want something else of you like attention for your
01:09:46 children or whatever. Then there's something in you which decides to keep focused or how does it work with you?
01:09:54 Yoshua: Right.
01:09:54 Speaker 2: You don't want to lose the thought of course.
01:09:57 Yoshua: That's right. So I write, I have some notebooks. I write my ideas.
01:10:05 Often when I wake up or sometimes an idea comes and I want to write it down, like if I was afraid of losing it.
01:10:11 But actually the good ideas, they don't they don't go.
01:10:15 It turns out very often I write them, but I don't even go back to reading them.
01:10:17 It's just that it makes me feel better, and it anchors.
01:10:21 Also, the fact of writing an idea kind of makes it take more room in my mind.
01:10:31 And there's also something to be said about concentration.
01:10:36 So my work now, because I'm immersed with so many people, can be very distractive.
01:10:42 But to really make big progress in science, I also need times when I can be very focused
01:10:54 and where the ideas about a problem and different points of view and all the elements sort of fill my mind.
01:11:02 I'm completely filled with this.
01:11:05 That's when you can be really productive and it might take a long time before you reach that state.
01:11:11 Sometimes it could take years for a student to really go deep into a subject. So that he can be fully immersed in it.
01:11:20 That's when you can really start seeing through things and getting things to stand together solidly.
01:11:28 Now you can extend science, right? Now, when things are solid in your mind, you can move forward.
01:11:36 Speaker 2: Like a base of understanding?
01:11:38 Yoshua: Yeah, yeah, when you need enough concentration on something to really get these moves.
01:11:46 There's the other mode of thinking, which is the brainstorming mode.
01:11:49 Where, out of the blue, I start a discussion, five minutes later something comes up.
01:11:54 So that's more like random and it's also very, it could be very productive as well.
01:12:01 It depends on the stimulation from someone else.
01:12:03 If someone introduces a problem and immediately I get a, something comes up. And we have maybe an exchange.
01:12:13 So that's more superficial, but a lot of good things come out of that exchange because of the brainstorming.
01:12:19 Whereas the other, there's the other mode of thinking which is I'm alone nobody bothers me.
01:12:26 Nobody's asking for my attention. I'm walking.
01:12:30 I'm half asleep, and there I can fully concentrate, eyes closed
01:12:36 or not really paying attention to what's going on in front of me, because I'm completely in my thoughts.
01:12:41 Speaker 2: When do you think?
01:12:46 Yoshua: When?
01:12:47 Speaker 2: During the day. Let's start a day.
01:12:49 Yoshua: So the two times when I spend more on this concentrated thinking, is usually when I wake up, and
01:13:00 when I'm walking back and forth between home and university.
01:13:04 Speaker 2: Just enlarge this moment, what happens?
01:13:10 Yoshua: So I emerge to conciousness like everybody does every morning, and eyes closed
01:13:20 and so on Some thought related to a research question or maybe non-research question comes up
01:13:31 and if I'm interested in it I start like going deeper into it. And.
01:13:37 Speaker 2: Still with your eyes closed?
01:13:39 Yoshua: Still with my eyes closed.
01:13:40 And then it's like If you see a thread dangling and you pull on it, and then, more stuff comes down.
01:13:55 Now, you see more things and you pull more, and there's an avalanche of things coming.
01:14:02 The more you pull on those strings, and the more new things come, or information comes together.
01:14:11 And sometimes it goes nowhere and sometimes that's how new ideas come about.
01:14:15 Speaker 2: And at what stage in this pulling the thread, do you open your eyes?
01:14:21 Yoshua: I could stay like this for an hour.
01:14:24 Speaker 2: Eyes closed.
01:14:25 Yoshua: Yeah.
01:14:26 Speaker 2: Pulling a thread.
01:14:28 Yoshua: Yeah.
01:14:28 Speaker 2: Seeing what's happening.
01:14:30 Yoshua: Yeah.
01:14:31 Often what happens is I see something that I hadn't seen before and I get too excited, so that wakes me up
01:14:37 and I want to write it down. So I have my notebook not far and I write it down.
01:14:42 Or I wanna send an email to somebody saying, I thought about this and it's like six in the morning [LAUGH]
01:14:48 and they wonder if I'm working all the time. [LAUGH]
01:14:53 Speaker 2: So, and then, what happens then? Then you woke up.
01:14:58 Yoshua: Yeah.
01:14:58 Speaker 2: You open your eyes or you wrote it down?
01:15:02 Yoshua: So once I'm writing it down, my eyes are open and it's like, I feel relieved, it's like now I can go
01:15:13 and maybe have breakfast or take a shower, or something.
01:15:16 So having written it down, it might take some time to write it down, also sometimes I write an email
01:15:26 and then it's longer. And now the act of writing it is a different thing.
01:15:33 So there's the initial sort of spark of vision, which is still very fuzzy.
01:15:41 But then, when you have to communicate the idea to someone else. Say, in an email.
01:15:46 You have to really make a different kind of effort, you realize some flaws in your initial ideas
01:15:51 and you have to clean it up and make sure it's understandable. Now it takes a different form.
01:15:58 And sometimes you realize when you do it, that it was nothing really. Yeah, it was just half dream.
01:16:06 Speaker 2: What does your partner think of the ideas, that [INAUDIBLE]
01:16:10 Yoshua: I didn't understand the question.
01:16:13 Speaker 2: What does your partner think of this? That you wake up or you have to write something down?
01:16:24 Yoshua: She's fine with that. I think she's glad to see this kind of thing happen.
01:16:34 And she's happy for me that I live these very rewarding moments.
01:16:41 Speaker 2: But she understands what happens.
01:16:43 Yoshua: Yeah. I tell her often, I just had an idea. I wanna say, i just wanna.
01:16:54 Speaker 2: Does she understand?
01:16:55 Yoshua: What do you mean the science?
01:16:56 Speaker 2: Yes.
01:16:56 Yoshua: No, no but she understands that it's really important for me and this is how I move forward in my work
01:17:07 and also how emotionally fulfilling it is.
01:17:12 Speaker 2: Okay, then at a certain moment you have to go to work.
01:17:21 Yoshua: Yes.
01:17:23 Speaker 2: Let's talk about the walk you do every day.
01:17:23 Yoshua: Yes.
01:17:24 Speaker 2: So what does it mean?
01:17:26 Yoshua: So that walk is you can really think of it as a kind of meditation.
01:17:30 Speaker 2: Tell me about what you were doing if you want to.
01:17:31 Yoshua: So everyday I walk from my house. Yeah, so everyday I walk up the hill from my home to the university.
01:17:43 And it's about half an hour and it's more or less always the same path.
01:17:50 And because I know this path so well, I don't have to really pay much attention to what's going on.
01:17:55 And I can just relax and let thoughts go by, and eventually focus on something, or not.
01:18:05 Sometimes it's just maybe more in the evening where I'm tired maybe just a way to relax and let go.
01:18:17 Speaker 2: Quality thinking time is the problem.
01:18:19 Yoshua: Yes. Absolutely. Because I'm not bombarded by the outside world I can just.
01:18:30 Speaker 2: Normal people are bombarded by every signs, and cars, and sounds.
01:18:33 Yoshua: Yeah.
01:18:34 Speaker 2: And the weather.
01:18:36 Yoshua: Yeah I kind of ignore that. [LAUGH]
01:18:38 Speaker 2: So you are when there are thoughts around you.
01:18:46 Yoshua: When I was young I used to hit my head [LAUGH] on poles. [LAUGH]
01:18:54 Speaker 2: Because you were thinking [CROSSTALK] yourself..
01:18:57 Yoshua: Yeah, or reading while walking [LAUGH]
01:19:07 Speaker 2: [LAUGH] Now it doesnt happen any more.
01:19:09 Yoshua: No.
01:19:10 Well, actually it does now, because I sometimes, I check my phone [LAUGH] I see lots of people do that,
01:19:16 not being paying attention to what's going on.
01:19:20 Speaker 2: Yeah.
01:19:20 Yoshua: Yeah.
01:19:20 Speaker 2: So, well we will film your walk may be something happen Mm-hm.
01:19:27 [LAUGH] but during this walk, if you do it for such a long time, walking uphill.
01:19:32 Yoshua: Yeah.
01:19:32 Speaker 2: That's kind of a nice metaphor, walking up the hill.
01:19:43 Yoshua: Yeah.
01:19:43 Speaker 2: Are there, on this route situations, or positions, or places
01:19:47 when you had some really good ideas that you can remember?
01:19:51 Yoshua: Well.
01:19:52 Speaker 2: How was it?
01:19:53 I was waiting at the traffic light, or was it- Yeah, I have some memories of specific moments going up.
01:20:02 Yoshua: Thinking about some of the ideas that have been going through my mind over the last year in particular.
01:20:14 I guess these are more recent memories. Can you enlarge one of those moments like you did with waking up?
01:20:22 Right, right, so, as I said earlier, it's like if the rest of the world is in a haze, right.
01:20:32 It's like there's automatic control of the walking and watching for other people and cars, potentially.
01:20:43 But it's like if I had a 3-D projection of my thoughts in front of me, that are taking most of the room.
01:20:52 And my thinking works a lot by visualization. And I think a lot of people are like this.
01:20:59 It's a very nice tool that we have, using our kind of visual analogies to understand things.
01:21:09 Even if it's not a faithful portrait of what's going on, the visual analogies are really helping me, at least,
01:21:18 to make sense of things. So it's like I have pictures in my mind to illustrate what's going on, and it's like I see
01:21:26 Yoshua: What do I see? I see information flow, neural networks.
01:21:41 It's like if I was running a simulation in my mind of what would happen if
01:21:49 Yoshua: Some rule of conduct was followed by in this algorithm in this process.
01:21:58 Speaker 2: And that's when you walk up the hill that's what you see?
01:22:00 Yoshua: Yeah, yeah, so it's like if I was running a computer simulation in my mind.
01:22:07 To try to figure out what would happen if I made such choices or if we consider such equation.
01:22:18 what would it entail what would happen?
01:22:21 Imagine different situations and then of course it's not as detailed as if we did a real computer simulation.
01:22:30 But it provides a lot of insight for what's going on.
01:22:36 Speaker 2: But then you walk up the hill everyday.
01:22:37 Yoshua: Yeah.
01:22:37 Speaker 2: And describe the most defining moment during one of those walks. Where you were? Where you stood?
01:22:48 Which corner?
01:22:49 Yoshua: Well, so I remember a particular moment. I was walking on the north sidewalk of Queen Mary Street.
01:23:01 And I was seeing the big church we have there, which is called the oratoire. It's beautiful.
01:23:15 Yoshua: And then I got this insight about perturbations propagating in brains.
01:23:24 Speaker 2: Maybe you want to do that sooner than that.
01:23:26 Yoshua: Yeah, yeah. From the beginning or just the last sentence?
01:23:29 Speaker 2: The last one. Go on.
01:23:30 Yoshua: And so, then I got this insight, visually of these perturbations happening on neurons.
01:23:40 That propagate to other neurons, that propagate to other neurons.
01:23:43 And like I'm doing with my hands, but it was something visual. Then suddenly I had the thought that this could work.
01:23:56 That this could explain things that I'm always trying to understand.
01:23:59 Speaker 2: How did this feel?
01:24:01 Yoshua: Great, I think of all the good feelings that we can have in life, the feeling we get when something clicks,
01:24:15 the eureka. Is probably, maybe, the strongest and most powerful one that we can seek again and again.
01:24:24 And only brings positive things. Maybe stronger than food and sex and those usual good things we get from experience.
01:24:39 Speaker 2: You mean this moment?
01:24:40 Yoshua: This- These kinds of moments provide pleasure.
01:24:46 Yoshua: It's a different kind of pleasure, just like different pleasures or different sensory pleasure or so on.
01:24:54 But it's really, I think, when your brain realizes something, understands something.
01:25:00 It's like you send yourself some molecules to reward you. Say great, do it again if you can, right?
01:25:08 Speaker 2: Did you do it again?
01:25:10 Yoshua: Yeah, yeah, that's my job.
01:25:12 Speaker 2: So this is one moment at the church. Was it a coincidence that it was at a church?
01:25:18 Yoshua: No.
01:25:18 Speaker 2: That has nothing to do with it.
01:25:19 Yoshua: I don't believe in God.
01:25:20 Speaker 2: But, when, I don't believe in God either but if you think of God as someone who created us as is,
01:25:35 and he is our example.
01:25:37 Yoshua: Yes.
01:25:38 Speaker 2: Trying to understand what's happening in your head or your brain.
01:25:44 Yoshua: Yes.
01:25:45 Speaker 2: Isn't that what other people call God?
01:25:49 Speaker 2: Or looking for?
01:25:51 Yoshua: I'm not sure I understand your question.
01:25:58 Speaker 2: How can I rephrase that one?
01:26:11 Speaker 2: When you understand how a brain works-
01:26:18 Yoshua: Yes.
01:26:19 Speaker 2: Maybe then you understand who God is.
01:26:22 Yoshua: When we understand how our brains work we understand who we are to some extent,
01:26:29 I mean a very important part of us. That's one of my motivations.
01:26:33 And the process of doing it is something that defines us individually but also as a collective, as a group,
01:26:46 as a society. So there may be some connections to religion which are about connecting us to some extent.
01:26:54 Speaker 2: That's one of those layers you were talking about. Religion is one of them.
01:27:00 Yoshua: Mm-hm. Yep.
01:27:02 Speaker 2: So but doing this show [NOISE] this half an hour, then you were almost here so-
01:27:09 Yoshua: Sometimes I think it's too short. But then, I have things to do, so.
01:27:16 Speaker 2: Let's continue this metaphor. It's uphill, when you are uphill, what do you feel?
01:27:24 Yoshua: I feel, so I'm going uphill, my body's working hard.
01:27:30 I mean, I'm not running, but I'm walking and I can feel the muscles.
01:27:35 Warming up, and my whole body becoming more full with energy. And I think that helps the brain as well.
01:27:46 That's how it feels, anyway.
01:27:49 Speaker 2: But I mean, when you. Moses went up to the mountain and he saw the Promised Land. [LAUGH]
01:27:55 Speaker 2: When you go uphill what do you see?
01:27:58 Yoshua: When I go uphill [LAUGH] I see the university, but there is something that's related to your question.
01:28:10 Which is, each time I have these insights, these Eureka moments, it's like seeing the Promised Land.
01:28:16 It's very much like that. It's like you have a glimpse of something you had never seen before and it looks great.
01:28:27 And you feel like you now see a path to go there.
01:28:31 So I think it's very, very close to this idea of seeing the Promised Land.
01:28:37 But of course it's not just one Promised Land.
01:28:39 It's one step to the next valley and the next valley, and that's how we climb, really, big mountains.
01:28:46 Speaker 2: So is there anything you want to add to this yourself? Because I think we are ready now to go uphill.
01:28:58 Yoshua: No, I'm fine.
01:29:00 Speaker 2: Maybe just a few questions about Friday, so what you're going to do. What are you going to do on Friday?
01:29:12 Yoshua: So Friday I'm going to make a presentation to the rest of the researchers in the lab in the institute about one
01:29:25 of the topics I'm most excited about these days.
01:29:30 Which is trying to bridge the gap between what we do in machine learning, what has to do with AI
01:29:38 and building intelligent machines and the brain. I'm not really a brain expert.
01:29:44 I'm more a machine learning person, but I talk to neuroscientists and so on.
01:29:48 And I try, I really care about the big question of how is the brain doing the really complex things that it does.
01:29:57 And so the work I'm going to tell about Friday is one small step in that direction that we've achieved in the last few
01:30:08 months.
01:30:10 Speaker 2: On your path to the Promised Land?
01:30:12 Yoshua: Yes, exactly, that's right.
01:30:14 And I've been making those small steps on this particular topic for about a year and a half.
01:30:21 So it's not like just something happens and you're there, right?
01:30:27 It's a lot of insights that make you move and get understanding. And science makes progress by steps.
01:30:41 Most of those steps are small, some are slightly bigger.
01:30:44 Seen from the outside, sometimes people have the impression that, there's this big breakthrough, breakthrough.
01:30:49 And journalists like to talk about breakthrough, breakthrough, breakthrough.
01:30:52 But actually science is very, very progressive because we gradually understand better the world.
TV Get inspired and watch tv episodes of The Mind of the Universe, made by Dutch public broadcaster VPRO
  • Browse through over 30 hours of interviews
  • Download the interviews, including subtitles
  • Remix, re-use and edit under CC-BY-SA license
  • Start exploring