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