Yoshua Bengio: Deep Learning - Lex Fridman 知识图谱

Yoshua Bengio (03:30.720)

I think a big part of that is probably forgetting. You're only remembering the really important

Yoshua Bengio (03:36.080)

things. It's very efficient forgetting.

Yoshua Bengio (03:40.000)

Yes. So there's a selection of what we remember. And I think there are really cool connection to

Yoshua Bengio (03:46.160)

higher level cognition here regarding consciousness, deciding and emotions,

Lex Fridman (03:52.080)

so deciding what comes to consciousness and what gets stored in memory, which are not trivial either.

Lex Fridman (04:00.720)

So you've been at the forefront there all along, showing some of the amazing things that neural

Yoshua Bengio (04:07.120)

networks, deep neural networks can do in the field of artificial intelligence is just broadly

Yoshua Bengio (04:12.640)

in all kinds of applications. But we can talk about that forever. But what, in your view,

Yoshua Bengio (04:19.120)

because we're thinking towards the future, is the weakest aspect of the way deep neural networks

Lex Fridman (04:23.920)

represent the world? What is that? What is in your view is missing?

Lex Fridman (04:29.200)

So current state of the art neural nets trained on large quantities of images or texts

Yoshua Bengio (04:38.240)

have some level of understanding of, you know, what explains those data sets, but it's very

Yoshua Bengio (04:45.360)

basic, it's it's very low level. And it's not nearly as robust and abstract and general

Yoshua Bengio (04:54.160)

as our understanding. Okay, so that doesn't tell us how to fix things. But I think it encourages

Yoshua Bengio (05:02.400)

us to think about how we can maybe train our neural nets differently, so that they would

Yoshua Bengio (05:14.240)

focus, for example, on causal explanation, something that we don't do currently with neural

Yoshua Bengio (05:20.400)

net training. Also, one thing I'll talk about in my talk this afternoon is the fact that

Yoshua Bengio (05:27.440)

instead of learning separately from images and videos on one hand and from texts on the other

Yoshua Bengio (05:33.680)

hand, we need to do a better job of jointly learning about language and about the world

Yoshua Bengio (05:42.000)

to which it refers. So that, you know, both sides can help each other. We need to have good world

Yoshua Bengio (05:50.160)

models in our neural nets for them to really understand sentences, which talk about what's

Yoshua Bengio (05:57.360)

going on in the world. And I think we need language input to help provide clues about

Lex Fridman (06:06.400)

what high level concepts like semantic concepts should be represented at the top levels of our

Yoshua Bengio (06:13.600)

neural nets. In fact, there is evidence that the purely unsupervised learning of representations

Yoshua Bengio (06:21.920)

doesn't give rise to high level representations that are as powerful as the ones we're getting

Yoshua Bengio (06:28.960)

from supervised learning. And so the clues we're getting just with the labels, not even sentences,

Yoshua Bengio (06:35.680)

is already very, very high level. And I think that's a very important thing to keep in mind.

Lex Fridman (06:42.400)

It's already very powerful. Do you think that's an architecture challenge or is it a data set challenge?

Lex Fridman (06:49.520)

Neither. I'm tempted to just end it there. Can you elaborate slightly?

Yoshua Bengio (07:02.880)

Of course, data sets and architectures are something you want to always play with. But

Yoshua Bengio (07:06.800)

I think the crucial thing is more the training objectives, the training frameworks. For example,

Yoshua Bengio (07:13.040)

going from passive observation of data to more active agents, which

Yoshua Bengio (07:22.320)

learn by intervening in the world, the relationships between causes and effects,

Yoshua Bengio (07:27.280)

the sort of objective functions, which could be important to allow the highest level explanations

Yoshua Bengio (07:36.640)

to rise from the learning, which I don't think we have now, the kinds of objective functions,

Yoshua Bengio (07:43.840)

which could be used to reward exploration, the right kind of exploration. So these kinds of

Yoshua Bengio (07:50.400)

questions are neither in the data set nor in the architecture, but more in how we learn,

Yoshua Bengio (07:57.200)

under what objectives and so on. Yeah, I've heard you mention in several contexts, the idea of sort

Yoshua Bengio (08:04.240)

of the way children learn, they interact with objects in the world. And it seems fascinating

Yoshua Bengio (08:08.880)

because in some sense, except with some cases in reinforcement learning, that idea

Yoshua Bengio (08:15.520)

is not part of the learning process in artificial neural networks. So it's almost like,

Lex Fridman (08:21.360)

do you envision something like an objective function saying, you know what, if you

Yoshua Bengio (08:29.680)

poke this object in this kind of way, it would be really helpful for me to further learn.

Yoshua Bengio (08:36.400)

Right, right.

Lex Fridman (08:37.040)

Sort of almost guiding some aspect of the learning.

Yoshua Bengio (08:40.320)

Right, right, right. So I was talking to Rebecca Sacks just a few minutes ago,

Lex Fridman (08:43.600)

and she was talking about lots and lots of evidence from infants seem to clearly pick

Lex Fridman (08:52.960)

what interests them in a directed way. And so they're not passive learners, they focus their

Yoshua Bengio (09:03.040)

attention on aspects of the world, which are most interesting, surprising in a non trivial way.

Yoshua Bengio (09:10.480)

That makes them change their theories of the world.

Lex Fridman (09:16.000)

So that's a fascinating view of the future progress. But on a more maybe boring question,

Lex Fridman (09:26.080)

do you think going deeper and larger, so do you think just increasing the size of the things that

Lex Fridman (09:33.760)

have been increasing a lot in the past few years, is going to be a big thing?

Yoshua Bengio (09:38.800)

I think increasing the size of the things that have been increasing a lot in the past few years

Yoshua Bengio (09:44.320)

will also make significant progress. So some of the representational issues that you mentioned,

Yoshua Bengio (09:51.840)

they're kind of shallow, in some sense.

Lex Fridman (09:54.880)

Oh, shallow in the sense of abstraction.

Yoshua Bengio (09:58.400)

In the sense of abstraction, they're not getting some...

Yoshua Bengio (10:00.800)

I don't think that having more depth in the network in the sense of instead of 100 layers,

Lex Fridman (10:06.880)

you're going to have more layers. I don't think so. Is that obvious to you?

Yoshua Bengio (10:11.680)

Yes. What is clear to me is that engineers and companies and labs and grad students will continue

Yoshua Bengio (10:19.200)

to tune architectures and explore all kinds of tweaks to make the current state of the art

Yoshua Bengio (10:25.600)

slightly ever slightly better. But I don't think that's going to be nearly enough. I think we need

Yoshua Bengio (10:31.440)

changes in the way that we're considering learning to achieve the goal that these learners actually

Yoshua Bengio (10:39.920)

understand in a deep way the environment in which they are, you know, observing and acting.

Lex Fridman (10:46.640)

But I guess I was trying to ask a question that's more interesting than just more layers.

Yoshua Bengio (10:53.200)

It's basically, once you figure out a way to learn through interacting, how many parameters

Yoshua Bengio (11:00.800)

it takes to store that information. So I think our brain is quite bigger than most neural networks.

Yoshua Bengio (11:07.760)

Right, right. Oh, I see what you mean. Oh, I'm with you there. So I agree that in order to

Yoshua Bengio (11:14.240)

build neural nets with the kind of broad knowledge of the world that typical adult humans have,

Lex Fridman (11:20.960)

probably the kind of computing power we have now is going to be insufficient.

Lex Fridman (11:25.600)

So the good news is there are hardware companies building neural net chips. And so

Yoshua Bengio (11:30.320)

it's going to get better. However, the good news in a way, which is also a bad news,

Yoshua Bengio (11:37.520)

is that even our state of the art, deep learning methods fail to learn models that understand

Lex Fridman (11:46.960)

even very simple environments, like some grid worlds that we have built.

Yoshua Bengio (11:52.000)

Even these fairly simple environments, I mean, of course, if you train them with enough examples,

Yoshua Bengio (11:56.080)

eventually they get it. But it's just like, instead of what humans might need just

Yoshua Bengio (12:03.440)

dozens of examples, these things will need millions for very, very, very simple tasks.

Lex Fridman (12:10.000)

And so I think there's an opportunity for academics who don't have the kind of computing

Yoshua Bengio (12:16.640)

power that, say, Google has to do really important and exciting research to advance

Yoshua Bengio (12:23.440)

the state of the art in training frameworks, learning models, agent learning in even simple

Yoshua Bengio (12:30.960)

environments that are synthetic, that seem trivial, but yet current machine learning fails on.

Lex Fridman (12:38.240)

We talked about priors and common sense knowledge. It seems like

Yoshua Bengio (12:43.760)

we humans take a lot of knowledge for granted. So what's your view of these priors of forming

Yoshua Bengio (12:52.160)

this broad view of the world, this accumulation of information and how we can teach neural networks

Yoshua Bengio (12:58.880)

or learning systems to pick that knowledge up? So knowledge, for a while, the artificial

Yoshua Bengio (13:05.520)

intelligence was maybe in the 80s, like there's a time where knowledge representation, knowledge,

Yoshua Bengio (13:14.320)

acquisition, expert systems, I mean, the symbolic AI was a view, was an interesting problem set to

Yoshua Bengio (13:22.240)

solve and it was kind of put on hold a little bit, it seems like. Because it doesn't work.

Yoshua Bengio (13:27.680)

It doesn't work. That's right. But that's right. But the goals of that remain important.

Lex Fridman (13:34.960)

Yes. Remain important. And how do you think those goals can be addressed?

Yoshua Bengio (13:39.760)

Right. So first of all, I believe that one reason why the classical expert systems approach failed

Yoshua Bengio (13:48.400)

is because a lot of the knowledge we have, so you talked about common sense intuition,

Yoshua Bengio (13:56.320)

there's a lot of knowledge like this, which is not consciously accessible.

Yoshua Bengio (14:01.680)

There are lots of decisions we're taking that we can't really explain, even if sometimes we make

Yoshua Bengio (14:05.440)

up a story. And that knowledge is also necessary for machines to take good decisions. And that

Yoshua Bengio (14:15.600)

knowledge is hard to codify in expert systems, rule based systems and classical AI formalism.

Lex Fridman (14:22.960)

And there are other issues, of course, with the old AI, like not really good ways of handling

Yoshua Bengio (14:29.520)

uncertainty, I would say something more subtle, which we understand better now, but I think still

Yoshua Bengio (14:37.040)

isn't enough in the minds of people. There's something really powerful that comes from

Yoshua Bengio (14:43.920)

distributed representations, the thing that really makes neural nets work so well.

Lex Fridman (14:49.280)

And it's hard to replicate that kind of power in a symbolic world. The knowledge in expert systems

Lex Fridman (14:58.640)

and so on is nicely decomposed into like a bunch of rules. Whereas if you think about a neural net,

Yoshua Bengio (15:04.960)

it's the opposite. You have this big blob of parameters which work intensely together to

Yoshua Bengio (15:10.960)

represent everything the network knows. And it's not sufficiently factorized. It's not

Yoshua Bengio (15:16.960)

sufficiently factorized. And so I think this is one of the weaknesses of current neural nets,

Yoshua Bengio (15:24.240)

that we have to take lessons from classical AI in order to bring in another kind of compositionality,

Yoshua Bengio (15:32.320)

which is common in language, for example, and in these rules, but that isn't so native to neural

Yoshua Bengio (15:38.800)

nets. And on that line of thinking, disentangled representations. Yes. So let me connect with

Yoshua Bengio (15:48.400)

disentangled representations, if you might, if you don't mind. So for many years, I've thought,

Lex Fridman (15:55.280)

and I still believe that it's really important that we come up with learning algorithms,

Yoshua Bengio (16:00.560)

either unsupervised or supervised, but reinforcement, whatever, that build representations

Yoshua Bengio (16:06.400)

in which the important factors, hopefully causal factors are nicely separated and easy to pick up

Yoshua Bengio (16:13.360)

from the representation. So that's the idea of disentangled representations. It says transform

Yoshua Bengio (16:18.480)

the data into a space where everything becomes easy. We can maybe just learn with linear models

Yoshua Bengio (16:25.120)

about the things we care about. And I still think this is important, but I think this is missing out

Lex Fridman (16:30.960)

on a very important ingredient, which classical AI systems can remind us of.

Lex Fridman (16:38.080)

So let's say we have these disentangled representations. You still need to learn about

Yoshua Bengio (16:43.440)

the relationships between the variables, those high level semantic variables. They're not going

Yoshua Bengio (16:47.200)

to be independent. I mean, this is like too much of an assumption. They're going to have some

Yoshua Bengio (16:52.000)

interesting relationships that allow to predict things in the future, to explain what happened

Yoshua Bengio (16:56.320)

in the past. The kind of knowledge about those relationships in a classical AI system

Yoshua Bengio (17:01.600)

is encoded in the rules. Like a rule is just like a little piece of knowledge that says,

Yoshua Bengio (17:06.000)

oh, I have these two, three, four variables that are linked in this interesting way,

Lex Fridman (17:10.960)

then I can say something about one or two of them given a couple of others, right?

Yoshua Bengio (17:14.800)

In addition to disentangling the elements of the representation, which are like the variables

Yoshua Bengio (17:22.160)

in a rule based system, you also need to disentangle the mechanisms that relate those

Yoshua Bengio (17:31.840)

variables to each other. So like the rules. So the rules are neatly separated. Like each rule is,

Yoshua Bengio (17:37.200)

you know, living on its own. And when I change a rule because I'm learning, it doesn't need to

Yoshua Bengio (17:43.360)

break other rules. Whereas current neural nets, for example, are very sensitive to what's called

Yoshua Bengio (17:48.720)

catastrophic forgetting, where after I've learned some things and then I learn new things,

Yoshua Bengio (17:54.080)

they can destroy the old things that I had learned, right? If the knowledge was better

Lex Fridman (17:59.280)

factorized and separated, disentangled, then you would avoid a lot of that.

Yoshua Bengio (18:06.560)

Now, you can't do this in the sensory domain.

Lex Fridman (18:10.320)

What do you mean by sensory domain?

Yoshua Bengio (18:13.120)

Like in pixel space. But my idea is that when you project the data in the right semantic space,

Yoshua Bengio (18:18.640)

it becomes possible to now represent this extra knowledge beyond the transformation from inputs

Yoshua Bengio (18:25.040)

to representations, which is how representations act on each other and predict the future and so on

Yoshua Bengio (18:31.120)

in a way that can be neatly disentangled. So now it's the rules that are disentangled from each

Yoshua Bengio (18:37.680)

other and not just the variables that are disentangled from each other.

Lex Fridman (18:40.400)

And you draw a distinction between semantic space and pixel, like does there need to be

Lex Fridman (18:45.200)

an architectural difference?

Yoshua Bengio (18:46.560)

Well, yeah. So there's the sensory space like pixels, which where everything is entangled.

Yoshua Bengio (18:52.080)

The information, like the variables are completely interdependent in very complicated ways.

Lex Fridman (18:58.160)

And also computation, like it's not just the variables, it's also how they are related to

Yoshua Bengio (19:03.520)

each other is all intertwined. But I'm hypothesizing that in the right high level

Yoshua Bengio (19:11.280)

representation space, both the variables and how they relate to each other can be

Yoshua Bengio (19:16.720)

disentangled. And that will provide a lot of generalization power.

Lex Fridman (19:20.800)

Generalization power.

Yoshua Bengio (19:22.240)

Yes.

Yoshua Bengio (19:22.720)

Distribution of the test set is assumed to be the same as the distribution of the training set.

Yoshua Bengio (19:29.280)

Right. This is where current machine learning is too weak. It doesn't tell us anything,

Yoshua Bengio (19:35.600)

is not able to tell us anything about how our neural nets, say, are going to generalize to

Yoshua Bengio (19:40.080)

a new distribution. And, you know, people may think, well, but there's nothing we can say

Yoshua Bengio (19:45.120)

if we don't know what the new distribution will be. The truth is humans are able to generalize

Yoshua Bengio (19:50.880)

to new distributions.

Lex Fridman (19:52.560)

Yeah. How are we able to do that?

Yoshua Bengio (19:54.000)

Yeah. Because there is something, these new distributions, even though they could look

Yoshua Bengio (19:57.920)

very different from the training distributions, they have things in common. So let me give you

Yoshua Bengio (20:02.240)

a concrete example. You read a science fiction novel. The science fiction novel, maybe, you

Yoshua Bengio (20:07.920)

know, brings you in some other planet where things look very different on the surface,

Lex Fridman (20:15.200)

but it's still the same laws of physics. And so you can read the book and you understand

Yoshua Bengio (20:20.000)

what's going on. So the distribution is very different. But because you can transport

Yoshua Bengio (20:27.360)

a lot of the knowledge you had from Earth about the underlying cause and effect relationships

Lex Fridman (20:33.120)

and physical mechanisms and all that, and maybe even social interactions, you can now

Yoshua Bengio (20:38.720)

make sense of what is going on on this planet where, like, visually, for example,

Lex Fridman (20:42.160)

things are totally different.

Yoshua Bengio (20:45.280)

Taking that analogy further and distorting it, let's enter a science fiction world of,

Yoshua Bengio (20:50.800)

say, Space Odyssey, 2001, with Hal. Or maybe, which is probably one of my favorite AI movies.

Yoshua Bengio (20:59.840)

Me too.

Lex Fridman (21:00.480)

And then there's another one that a lot of people love that may be a little bit outside

Yoshua Bengio (21:05.360)

of the AI community is Ex Machina. I don't know if you've seen it.

Lex Fridman (21:10.000)

Yes. Yes.

Lex Fridman (21:11.600)

By the way, what are your views on that movie? Are you able to enjoy it?

Lex Fridman (21:16.000)

Are there things I like and things I hate?

Lex Fridman (21:21.120)

So you could talk about that in the context of a question I want to ask, which is, there's

Yoshua Bengio (21:26.800)

quite a large community of people from different backgrounds, often outside of AI, who are concerned

Yoshua Bengio (21:32.800)

about existential threat of artificial intelligence. You've seen this community

Yoshua Bengio (21:37.600)

develop over time. You've seen you have a perspective. So what do you think is the best

Yoshua Bengio (21:42.160)

way to talk about AI safety, to think about it, to have discourse about it within AI community

Lex Fridman (21:48.320)

and outside and grounded in the fact that Ex Machina is one of the main sources of information

Lex Fridman (21:54.560)

for the general public about AI?

Lex Fridman (21:56.560)

So I think you're putting it right. There's a big difference between the sort of discussion

Yoshua Bengio (22:02.240)

we ought to have within the AI community and the sort of discussion that really matter

Yoshua Bengio (22:07.600)

in the general public. So I think the picture of Terminator and AI loose and killing people

Lex Fridman (22:17.120)

and super intelligence that's going to destroy us, whatever we try, isn't really so useful

Yoshua Bengio (22:24.560)

for the public discussion. Because for the public discussion, the things I believe really

Yoshua Bengio (22:30.000)

matter are the short term and medium term, very likely negative impacts of AI on society,

Yoshua Bengio (22:37.200)

whether it's from security, like, you know, big brother scenarios with face recognition

Yoshua Bengio (22:43.280)

or killer robots, or the impact on the job market, or concentration of power and discrimination,

Yoshua Bengio (22:50.000)

all kinds of social issues, which could actually, some of them could really threaten democracy,

Yoshua Bengio (22:57.760)

for example.

Yoshua Bengio (22:58.800)

Just to clarify, when you said killer robots, you mean autonomous weapon, weapon systems.

Yoshua Bengio (23:04.000)

Yes, I don't mean that's right.

Lex Fridman (23:06.320)

So I think these short and medium term concerns should be important parts of the public debate.

Yoshua Bengio (23:13.040)

Now, existential risk, for me is a very unlikely consideration, but still worth academic investigation

Yoshua Bengio (23:24.640)

in the same way that you could say, should we study what could happen if meteorite, you

Yoshua Bengio (23:30.080)

know, came to earth and destroyed it. So I think it's very unlikely that this is going

Yoshua Bengio (23:33.920)

to happen in or happen in a reasonable future. The sort of scenario of an AI getting loose

Yoshua Bengio (23:43.040)

goes against my understanding of at least current machine learning and current neural

Yoshua Bengio (23:46.560)

nets and so on. It's not plausible to me. But of course, I don't have a crystal ball

Lex Fridman (23:51.120)

and who knows what AI will be in 50 years from now. So I think it is worth that scientists

Yoshua Bengio (23:55.520)

study those problems. It's just not a pressing question as far as I'm concerned.

Lex Fridman (23:59.680)

So before I continue down that line, I have a few questions there. But what do you like

Lex Fridman (24:05.840)

and not like about Ex Machina as a movie? Because I actually watched it for the second

Yoshua Bengio (24:09.840)

time and enjoyed it. I hated it the first time, and I enjoyed it quite a bit more the

Yoshua Bengio (24:15.600)

second time when I sort of learned to accept certain pieces of it, see it as a concept

Lex Fridman (24:23.440)

movie. What was your experience? What were your thoughts?

Lex Fridman (24:26.320)

So the negative is the picture it paints of science is totally wrong. Science in general

Lex Fridman (24:36.080)

and AI in particular. Science is not happening in some hidden place by some, you know, really

Yoshua Bengio (24:44.160)

smart guy, one person. This is totally unrealistic. This is not how it happens. Even a team of

Yoshua Bengio (24:52.160)

people in some isolated place will not make it. Science moves by small steps, thanks to

Yoshua Bengio (24:59.840)

the collaboration and community of a large number of people interacting. And all the

Yoshua Bengio (25:10.480)

scientists who are expert in their field kind of know what is going on, even in the industrial

Yoshua Bengio (25:14.560)

labs. It's information flows and leaks and so on. And the spirit of it is very different

Yoshua Bengio (25:21.920)

from the way science is painted in this movie.

Yoshua Bengio (25:25.600)

Yeah, let me ask on that point. It's been the case to this point that kind of even if

Yoshua Bengio (25:32.400)

the research happens inside Google or Facebook, inside companies, it still kind of comes out,

Yoshua Bengio (25:36.800)

ideas come out. Do you think that will always be the case with AI? Is it possible to bottle

Yoshua Bengio (25:41.680)

ideas to the point where there's a set of breakthroughs that go completely undiscovered

Lex Fridman (25:47.360)

by the general research community? Do you think that's even possible?

Yoshua Bengio (25:52.240)

It's possible, but it's unlikely. It's not how it is done now. It's not how I can foresee

Yoshua Bengio (25:59.520)

it in the foreseeable future. But of course, I don't have a crystal ball and science is

Yoshua Bengio (26:09.520)

a crystal ball. And so who knows? This is science fiction after all.

Lex Fridman (26:14.960)

I think it's ominous that the lights went off during that discussion.

Lex Fridman (26:21.440)

So the problem, again, there's one thing is the movie and you could imagine all kinds

Yoshua Bengio (26:25.320)

of science fiction. The problem for me, maybe similar to the question about existential

Yoshua Bengio (26:30.320)

risk, is that this kind of movie paints such a wrong picture of what is the actual science

Lex Fridman (26:39.440)

and how it's going on that it can have unfortunate effects on people's understanding of current

Yoshua Bengio (26:45.640)

science. And so that's kind of sad.

Yoshua Bengio (26:50.800)

There's an important principle in research, which is diversity. So in other words, research

Yoshua Bengio (26:58.440)

is exploration. Research is exploration in the space of ideas. And different people will

Yoshua Bengio (27:03.720)

focus on different directions. And this is not just good, it's essential. So I'm totally

Yoshua Bengio (27:09.520)

fine with people exploring directions that are contrary to mine or look orthogonal to

Yoshua Bengio (27:16.440)

mine. I am more than fine. I think it's important. I and my friends don't claim we have universal

Yoshua Bengio (27:24.920)

truth about what will, especially about what will happen in the future. Now that being

Yoshua Bengio (27:29.560)

said, we have our intuitions and then we act accordingly according to where we think we

Yoshua Bengio (27:36.560)

can be most useful and where society has the most to gain or to lose. We should have those

Yoshua Bengio (27:42.480)

debates and not end up in a society where there's only one voice and one way of thinking

Lex Fridman (27:49.800)

and research money is spread out.

Lex Fridman (27:53.520)

So disagreement is a sign of good research, good science.

Yoshua Bengio (27:59.040)

Yes.

Yoshua Bengio (28:00.040)

The idea of bias in the human sense of bias. How do you think about instilling in machine

Yoshua Bengio (28:08.600)

learning something that's aligned with human values in terms of bias? We intuitively as

Yoshua Bengio (28:15.240)

human beings have a concept of what bias means, of what fundamental respect for other human

Lex Fridman (28:21.160)

beings means. But how do we instill that into machine learning systems, do you think?

Lex Fridman (28:26.760)

So I think there are short term things that are already happening and then there are long

Yoshua Bengio (28:32.360)

term things that we need to do. In the short term, there are techniques that have been

Yoshua Bengio (28:38.360)

proposed and I think will continue to be improved and maybe alternatives will come up to take

Yoshua Bengio (28:44.200)

data sets in which we know there is bias, we can measure it. Pretty much any data set

Yoshua Bengio (28:50.120)

where humans are being observed taking decisions will have some sort of bias, discrimination

Yoshua Bengio (28:55.520)

against particular groups and so on.

Lex Fridman (28:59.000)

And we can use machine learning techniques to try to build predictors, classifiers that

Yoshua Bengio (29:04.240)

are going to be less biased. We can do it, for example, using adversarial methods to

Yoshua Bengio (29:11.600)

make our systems less sensitive to these variables we should not be sensitive to.

Lex Fridman (29:18.360)

So these are clear, well defined ways of trying to address the problem. Maybe they have weaknesses

Lex Fridman (29:23.520)

and more research is needed and so on. But I think in fact they are sufficiently mature

Yoshua Bengio (29:28.840)

that governments should start regulating companies where it matters, say like insurance companies,

Lex Fridman (29:35.240)

so that they use those techniques. Because those techniques will probably reduce the

Yoshua Bengio (29:40.480)

bias but at a cost. For example, maybe their predictions will be less accurate and so companies

Lex Fridman (29:46.440)

will not do it until you force them.

Yoshua Bengio (29:48.560)

All right, so this is short term. Long term, I'm really interested in thinking how we can

Yoshua Bengio (29:56.040)

instill moral values into computers. Obviously, this is not something we'll achieve in the

Yoshua Bengio (30:01.560)

next five or 10 years. How can we, you know, there's already work in detecting emotions,

Yoshua Bengio (30:08.120)

for example, in images, in sounds, in texts, and also studying how different agents interacting

Yoshua Bengio (30:19.880)

in different ways may correspond to patterns of, say, injustice, which could trigger anger.

Lex Fridman (30:28.200)

So these are things we can do in the medium term and eventually train computers to model,

Yoshua Bengio (30:37.840)

for example, how humans react emotionally. I would say the simplest thing is unfair situations

Yoshua Bengio (30:46.960)

which trigger anger. This is one of the most basic emotions that we share with other animals.

Yoshua Bengio (30:52.680)

I think it's quite feasible within the next few years that we can build systems that can

Yoshua Bengio (30:57.160)

detect these kinds of things to the extent, unfortunately, that they understand enough

Yoshua Bengio (31:01.980)

about the world around us, which is a long time away. But maybe we can initially do this

Yoshua Bengio (31:08.240)

in virtual environments. So you can imagine a video game where agents interact in some

Yoshua Bengio (31:14.840)

ways and then some situations trigger an emotion. I think we could train machines to detect

Yoshua Bengio (31:21.640)

those situations and predict that the particular emotion will likely be felt if a human was

Yoshua Bengio (31:27.400)

playing one of the characters.

Yoshua Bengio (31:29.460)

You have shown excitement and done a lot of excellent work with unsupervised learning.

Lex Fridman (31:35.720)

But there's been a lot of success on the supervised learning side.

Lex Fridman (31:39.840)

Yes, yes.

Lex Fridman (31:40.840)

And one of the things I'm really passionate about is how humans and robots work together.

Lex Fridman (31:46.680)

And in the context of supervised learning, that means the process of annotation. Do you

Yoshua Bengio (31:52.800)

think about the problem of annotation put in a more interesting way as humans teaching

Lex Fridman (32:00.080)

machines?

Yoshua Bengio (32:01.080)

Yes.

Lex Fridman (32:02.080)

Is there?

Yoshua Bengio (32:03.080)

Yes. I think it's an important subject. Reducing it to annotation may be useful for somebody

Yoshua Bengio (32:09.560)

building a system tomorrow. But longer term, the process of teaching, I think, is something

Yoshua Bengio (32:16.300)

that deserves a lot more attention from the machine learning community. So there are people

Yoshua Bengio (32:19.960)

who have coined the term machine teaching. So what are good strategies for teaching a

Lex Fridman (32:24.560)

learning agent? And can we design and train a system that is going to be a good teacher?

Lex Fridman (32:33.160)

So in my group, we have a project called BBI or BBI game, where there is a game or scenario

Yoshua Bengio (32:42.200)

where there's a learning agent and a teaching agent. Presumably, the teaching agent would

Yoshua Bengio (32:48.480)

eventually be a human. But we're not there yet. And the role of the teacher is to use

Yoshua Bengio (32:57.960)

its knowledge of the environment, which it can acquire using whatever way brute force

Yoshua Bengio (33:04.840)

to help the learner learn as quickly as possible. So the learner is going to try to learn by

Yoshua Bengio (33:10.760)

itself, maybe using some exploration and whatever. But the teacher can choose, can have an influence

Yoshua Bengio (33:19.920)

on the interaction with the learner, so as to guide the learner, maybe teach it the things

Yoshua Bengio (33:27.160)

that the learner has most trouble with, or just add the boundary between what it knows

Lex Fridman (33:30.840)

and doesn't know, and so on. So there's a tradition of these kind of ideas from other

Yoshua Bengio (33:36.180)

fields and like tutorial systems, for example, and AI. And of course, people in the humanities

Yoshua Bengio (33:45.320)

have been thinking about these questions. But I think it's time that machine learning

Yoshua Bengio (33:48.240)

people look at this, because in the future, we'll have more and more human machine interaction

Yoshua Bengio (33:55.440)

with the human in the loop. And I think understanding how to make this work better, all the problems

Yoshua Bengio (34:01.040)

around that are very interesting and not sufficiently addressed. You've done a lot of work with

Yoshua Bengio (34:06.160)

language, too. What aspect of the traditionally formulated Turing test, a test of natural

Lex Fridman (34:14.000)

language understanding and generation in your eyes is the most difficult of conversation?

Lex Fridman (34:19.520)

What in your eyes is the hardest part of conversation to solve for machines? So I would say it's

Yoshua Bengio (34:25.640)

everything having to do with the non linguistic knowledge, which implicitly you need in order

Yoshua Bengio (34:32.300)

to make sense of sentences, things like the Winograd schema. So these sentences that are

Yoshua Bengio (34:37.680)

semantically ambiguous. In other words, you need to understand enough about the world

Yoshua Bengio (34:43.720)

in order to really interpret properly those sentences. I think these are interesting challenges

Yoshua Bengio (34:49.280)

for machine learning, because they point in the direction of building systems that both

Yoshua Bengio (34:57.300)

understand how the world works and this causal relationships in the world and associate that

Yoshua Bengio (35:03.760)

knowledge with how to express it in language, either for reading or writing.

Lex Fridman (35:12.080)

You speak French?

Yoshua Bengio (35:13.080)

Yes, it's my mother tongue.

Yoshua Bengio (35:14.760)

It's one of the romance languages. Do you think passing the Turing test and all the

Yoshua Bengio (35:20.400)

underlying challenges we just mentioned depend on language? Do you think it might be easier

Lex Fridman (35:24.320)

in French than it is in English, or is independent of language?

Yoshua Bengio (35:28.920)

I think it's independent of language. I would like to build systems that can use the same

Yoshua Bengio (35:37.600)

principles, the same learning mechanisms to learn from human agents, whatever their language.

Yoshua Bengio (35:46.720)

Well, certainly us humans can talk more beautifully and smoothly in poetry, some Russian originally.

Yoshua Bengio (35:53.560)

I know poetry in Russian is maybe easier to convey complex ideas than it is in English.

Lex Fridman (36:02.600)

But maybe I'm showing my bias and some people could say that about French. But of course,

Yoshua Bengio (36:09.480)

the goal ultimately is our human brain is able to utilize any kind of those languages

Yoshua Bengio (36:15.880)

to use them as tools to convey meaning.

Yoshua Bengio (36:18.280)

Yeah, of course, there are differences between languages, and maybe some are slightly better

Yoshua Bengio (36:22.040)

at some things, but in the grand scheme of things, where we're trying to understand how

Lex Fridman (36:26.120)

the brain works and language and so on, I think these differences are minute.

Lex Fridman (36:32.040)

So you've lived perhaps through an AI winter of sorts?

Lex Fridman (36:38.880)

Yes.

Lex Fridman (36:39.920)

How did you stay warm and continue your research?

Lex Fridman (36:44.740)

Stay warm with friends.

Yoshua Bengio (36:45.740)

With friends. Okay, so it's important to have friends. And what have you learned from the

Lex Fridman (36:51.160)

experience?

Yoshua Bengio (36:53.600)

Listen to your inner voice. Don't, you know, be trying to just please the crowds and the

Yoshua Bengio (37:02.040)

fashion. And if you have a strong intuition about something that is not contradicted by

Yoshua Bengio (37:10.320)

actual evidence, go for it. I mean, it could be contradicted by people.

Lex Fridman (37:17.280)

Not your own instinct of based on everything you've learned?

Yoshua Bengio (37:20.600)

Of course, you have to adapt your beliefs when your experiments contradict those beliefs.

Lex Fridman (37:28.320)

But you have to stick to your beliefs. Otherwise, it's what allowed me to go through those years.

Yoshua Bengio (37:35.000)

It's what allowed me to persist in directions that, you know, took time, whatever other

Lex Fridman (37:42.040)

people think, took time to mature and bring fruits.

Lex Fridman (37:48.040)

So history of AI is marked with these, of course, it's marked with technical breakthroughs,

Lex Fridman (37:54.520)

but it's also marked with these seminal events that capture the imagination of the community.

Yoshua Bengio (38:00.980)

Most recent, I would say, AlphaGo beating the world champion human Go player was one

Lex Fridman (38:06.400)

of those moments. What do you think the next such moment might be?

Yoshua Bengio (38:12.360)

Okay, so first of all, I think that these so called seminal events are overrated. As

Yoshua Bengio (38:22.600)

I said, science really moves by small steps. Now what happens is you make one more small

Yoshua Bengio (38:30.200)

step and it's like the drop that, you know, that fills the bucket and then you have drastic

Yoshua Bengio (38:39.480)

consequences because now you're able to do something you were not able to do before.

Yoshua Bengio (38:43.920)

Or now, say, the cost of building some device or solving a problem becomes cheaper than

Lex Fridman (38:49.720)

what existed and you have a new market that opens up, right? So especially in the world

Yoshua Bengio (38:53.900)

of commerce and applications, the impact of a small scientific progress could be huge.

Lex Fridman (39:03.760)

But in the science itself, I think it's very, very gradual.

Lex Fridman (39:07.800)

And where are these steps being taken now? So there's unsupervised learning.

Lex Fridman (39:13.160)

So if I look at one trend that I like in my community, so for example, at Milan, my institute,

Lex Fridman (39:23.380)

what are the two hardest topics? GANs and reinforcement learning. Even though in Montreal

Yoshua Bengio (39:31.840)

in particular, reinforcement learning was something pretty much absent just two or three

Yoshua Bengio (39:37.020)

years ago. So there's really a big interest from students and there's a big interest from

Yoshua Bengio (39:44.280)

people like me. So I would say this is something where we're going to see more progress, even

Yoshua Bengio (39:51.560)

though it hasn't yet provided much in terms of actual industrial fallout. Like even though

Yoshua Bengio (39:58.680)

there's AlphaGo, there's no, like Google is not making money on this right now. But I

Yoshua Bengio (40:03.360)

think over the long term, this is really, really important for many reasons.

Lex Fridman (40:08.960)

So in other words, I would say reinforcement learning may be more generally agent learning

Yoshua Bengio (40:13.840)

because it doesn't have to be with rewards. It could be in all kinds of ways that an agent

Lex Fridman (40:17.520)

is learning about its environment.

Yoshua Bengio (40:20.720)

Now reinforcement learning you're excited about, do you think GANs could provide something,

Yoshua Bengio (40:28.840)

at the moment? Well, GANs or other generative models, I believe, will be crucial ingredients

Yoshua Bengio (40:38.880)

in building agents that can understand the world. A lot of the successes in reinforcement

Yoshua Bengio (40:45.480)

learning in the past has been with policy gradient, where you just learn a policy, you

Yoshua Bengio (40:51.160)

don't actually learn a model of the world. But there are lots of issues with that. And

Yoshua Bengio (40:55.760)

we don't know how to do model based RL right now. But I think this is where we have to

Yoshua Bengio (41:00.880)

go in order to build models that can generalize faster and better like to new distributions

Yoshua Bengio (41:09.340)

that capture to some extent, at least the underlying causal mechanisms in the world.

Yoshua Bengio (41:16.120)

Last question. What made you fall in love with artificial intelligence? If you look

Yoshua Bengio (41:21.480)

back, what was the first moment in your life when you were fascinated by either the human

Lex Fridman (41:28.880)

mind or the artificial mind?

Yoshua Bengio (41:31.360)

You know, when I was an adolescent, I was reading a lot. And then I started reading

Yoshua Bengio (41:35.520)

science fiction.

Lex Fridman (41:36.520)

There you go.

Yoshua Bengio (41:37.520)

That's it. That's where I got hooked. And then, you know, I had one of the first personal

Lex Fridman (41:46.520)

computers and I got hooked in programming. And so it just, you know,

Yoshua Bengio (41:52.680)

Start with fiction and then make it a reality.

Lex Fridman (41:54.800)

That's right.

Yoshua Bengio (41:55.800)

Yoshua, thank you so much for talking to me.

Lex Fridman (41:57.560)

My pleasure.