episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
dozens of examples, these things will need millions for very, very, very simple tasks.
https://karpathy.ai/lexicap/0004-large.html#00:12:03.440
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And so I think there's an opportunity for academics who don't have the kind of computing
https://karpathy.ai/lexicap/0004-large.html#00:12:10.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
power that, say, Google has to do really important and exciting research to advance
https://karpathy.ai/lexicap/0004-large.html#00:12:16.640
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
the state of the art in training frameworks, learning models, agent learning in even simple
https://karpathy.ai/lexicap/0004-large.html#00:12:23.440
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
environments that are synthetic, that seem trivial, but yet current machine learning fails on.
https://karpathy.ai/lexicap/0004-large.html#00:12:30.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
We talked about priors and common sense knowledge. It seems like
https://karpathy.ai/lexicap/0004-large.html#00:12:38.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
we humans take a lot of knowledge for granted. So what's your view of these priors of forming
https://karpathy.ai/lexicap/0004-large.html#00:12:43.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
this broad view of the world, this accumulation of information and how we can teach neural networks
https://karpathy.ai/lexicap/0004-large.html#00:12:52.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
or learning systems to pick that knowledge up? So knowledge, for a while, the artificial
https://karpathy.ai/lexicap/0004-large.html#00:12:58.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
intelligence was maybe in the 80s, like there's a time where knowledge representation, knowledge,
https://karpathy.ai/lexicap/0004-large.html#00:13:05.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
acquisition, expert systems, I mean, the symbolic AI was a view, was an interesting problem set to
https://karpathy.ai/lexicap/0004-large.html#00:13:14.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
solve and it was kind of put on hold a little bit, it seems like. Because it doesn't work.
https://karpathy.ai/lexicap/0004-large.html#00:13:22.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
It doesn't work. That's right. But that's right. But the goals of that remain important.
https://karpathy.ai/lexicap/0004-large.html#00:13:27.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes. Remain important. And how do you think those goals can be addressed?
https://karpathy.ai/lexicap/0004-large.html#00:13:34.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Right. So first of all, I believe that one reason why the classical expert systems approach failed
https://karpathy.ai/lexicap/0004-large.html#00:13:39.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
is because a lot of the knowledge we have, so you talked about common sense intuition,
https://karpathy.ai/lexicap/0004-large.html#00:13:48.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
there's a lot of knowledge like this, which is not consciously accessible.
https://karpathy.ai/lexicap/0004-large.html#00:13:56.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
There are lots of decisions we're taking that we can't really explain, even if sometimes we make
https://karpathy.ai/lexicap/0004-large.html#00:14:01.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
up a story. And that knowledge is also necessary for machines to take good decisions. And that
https://karpathy.ai/lexicap/0004-large.html#00:14:05.440
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
knowledge is hard to codify in expert systems, rule based systems and classical AI formalism.
https://karpathy.ai/lexicap/0004-large.html#00:14:15.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And there are other issues, of course, with the old AI, like not really good ways of handling
https://karpathy.ai/lexicap/0004-large.html#00:14:22.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
uncertainty, I would say something more subtle, which we understand better now, but I think still
https://karpathy.ai/lexicap/0004-large.html#00:14:29.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
isn't enough in the minds of people. There's something really powerful that comes from
https://karpathy.ai/lexicap/0004-large.html#00:14:37.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
distributed representations, the thing that really makes neural nets work so well.
https://karpathy.ai/lexicap/0004-large.html#00:14:43.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And it's hard to replicate that kind of power in a symbolic world. The knowledge in expert systems
https://karpathy.ai/lexicap/0004-large.html#00:14:49.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
and so on is nicely decomposed into like a bunch of rules. Whereas if you think about a neural net,
https://karpathy.ai/lexicap/0004-large.html#00:14:58.640
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
it's the opposite. You have this big blob of parameters which work intensely together to
https://karpathy.ai/lexicap/0004-large.html#00:15:04.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
represent everything the network knows. And it's not sufficiently factorized. It's not
https://karpathy.ai/lexicap/0004-large.html#00:15:10.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
sufficiently factorized. And so I think this is one of the weaknesses of current neural nets,
https://karpathy.ai/lexicap/0004-large.html#00:15:16.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
that we have to take lessons from classical AI in order to bring in another kind of compositionality,
https://karpathy.ai/lexicap/0004-large.html#00:15:24.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
which is common in language, for example, and in these rules, but that isn't so native to neural
https://karpathy.ai/lexicap/0004-large.html#00:15:32.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
nets. And on that line of thinking, disentangled representations. Yes. So let me connect with
https://karpathy.ai/lexicap/0004-large.html#00:15:38.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
disentangled representations, if you might, if you don't mind. So for many years, I've thought,
https://karpathy.ai/lexicap/0004-large.html#00:15:48.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
and I still believe that it's really important that we come up with learning algorithms,
https://karpathy.ai/lexicap/0004-large.html#00:15:55.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
either unsupervised or supervised, but reinforcement, whatever, that build representations
https://karpathy.ai/lexicap/0004-large.html#00:16:00.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in which the important factors, hopefully causal factors are nicely separated and easy to pick up
https://karpathy.ai/lexicap/0004-large.html#00:16:06.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
from the representation. So that's the idea of disentangled representations. It says transform
https://karpathy.ai/lexicap/0004-large.html#00:16:13.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
the data into a space where everything becomes easy. We can maybe just learn with linear models
https://karpathy.ai/lexicap/0004-large.html#00:16:18.480
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
about the things we care about. And I still think this is important, but I think this is missing out
https://karpathy.ai/lexicap/0004-large.html#00:16:25.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
on a very important ingredient, which classical AI systems can remind us of.
https://karpathy.ai/lexicap/0004-large.html#00:16:30.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So let's say we have these disentangled representations. You still need to learn about
https://karpathy.ai/lexicap/0004-large.html#00:16:38.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
the relationships between the variables, those high level semantic variables. They're not going
https://karpathy.ai/lexicap/0004-large.html#00:16:43.440
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to be independent. I mean, this is like too much of an assumption. They're going to have some
https://karpathy.ai/lexicap/0004-large.html#00:16:47.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
interesting relationships that allow to predict things in the future, to explain what happened
https://karpathy.ai/lexicap/0004-large.html#00:16:52.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in the past. The kind of knowledge about those relationships in a classical AI system
https://karpathy.ai/lexicap/0004-large.html#00:16:56.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
is encoded in the rules. Like a rule is just like a little piece of knowledge that says,
https://karpathy.ai/lexicap/0004-large.html#00:17:01.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
oh, I have these two, three, four variables that are linked in this interesting way,
https://karpathy.ai/lexicap/0004-large.html#00:17:06.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
then I can say something about one or two of them given a couple of others, right?
https://karpathy.ai/lexicap/0004-large.html#00:17:10.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
In addition to disentangling the elements of the representation, which are like the variables
https://karpathy.ai/lexicap/0004-large.html#00:17:14.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in a rule based system, you also need to disentangle the mechanisms that relate those
https://karpathy.ai/lexicap/0004-large.html#00:17:22.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
variables to each other. So like the rules. So the rules are neatly separated. Like each rule is,
https://karpathy.ai/lexicap/0004-large.html#00:17:31.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
you know, living on its own. And when I change a rule because I'm learning, it doesn't need to
https://karpathy.ai/lexicap/0004-large.html#00:17:37.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
break other rules. Whereas current neural nets, for example, are very sensitive to what's called
https://karpathy.ai/lexicap/0004-large.html#00:17:43.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
catastrophic forgetting, where after I've learned some things and then I learn new things,
https://karpathy.ai/lexicap/0004-large.html#00:17:48.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
they can destroy the old things that I had learned, right? If the knowledge was better
https://karpathy.ai/lexicap/0004-large.html#00:17:54.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
factorized and separated, disentangled, then you would avoid a lot of that.
https://karpathy.ai/lexicap/0004-large.html#00:17:59.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Now, you can't do this in the sensory domain.
https://karpathy.ai/lexicap/0004-large.html#00:18:06.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
What do you mean by sensory domain?
https://karpathy.ai/lexicap/0004-large.html#00:18:10.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Like in pixel space. But my idea is that when you project the data in the right semantic space,
https://karpathy.ai/lexicap/0004-large.html#00:18:13.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
it becomes possible to now represent this extra knowledge beyond the transformation from inputs
https://karpathy.ai/lexicap/0004-large.html#00:18:18.640
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to representations, which is how representations act on each other and predict the future and so on
https://karpathy.ai/lexicap/0004-large.html#00:18:25.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in a way that can be neatly disentangled. So now it's the rules that are disentangled from each
https://karpathy.ai/lexicap/0004-large.html#00:18:31.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
other and not just the variables that are disentangled from each other.
https://karpathy.ai/lexicap/0004-large.html#00:18:37.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And you draw a distinction between semantic space and pixel, like does there need to be
https://karpathy.ai/lexicap/0004-large.html#00:18:40.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
an architectural difference?
https://karpathy.ai/lexicap/0004-large.html#00:18:45.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Well, yeah. So there's the sensory space like pixels, which where everything is entangled.
https://karpathy.ai/lexicap/0004-large.html#00:18:46.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
The information, like the variables are completely interdependent in very complicated ways.
https://karpathy.ai/lexicap/0004-large.html#00:18:52.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And also computation, like it's not just the variables, it's also how they are related to
https://karpathy.ai/lexicap/0004-large.html#00:18:58.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
each other is all intertwined. But I'm hypothesizing that in the right high level
https://karpathy.ai/lexicap/0004-large.html#00:19:03.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
representation space, both the variables and how they relate to each other can be
https://karpathy.ai/lexicap/0004-large.html#00:19:11.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
disentangled. And that will provide a lot of generalization power.
https://karpathy.ai/lexicap/0004-large.html#00:19:16.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Generalization power.
https://karpathy.ai/lexicap/0004-large.html#00:19:20.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes.
https://karpathy.ai/lexicap/0004-large.html#00:19:22.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Distribution of the test set is assumed to be the same as the distribution of the training set.
https://karpathy.ai/lexicap/0004-large.html#00:19:22.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Right. This is where current machine learning is too weak. It doesn't tell us anything,
https://karpathy.ai/lexicap/0004-large.html#00:19:29.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
is not able to tell us anything about how our neural nets, say, are going to generalize to
https://karpathy.ai/lexicap/0004-large.html#00:19:35.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
a new distribution. And, you know, people may think, well, but there's nothing we can say
https://karpathy.ai/lexicap/0004-large.html#00:19:40.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
if we don't know what the new distribution will be. The truth is humans are able to generalize
https://karpathy.ai/lexicap/0004-large.html#00:19:45.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to new distributions.
https://karpathy.ai/lexicap/0004-large.html#00:19:50.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yeah. How are we able to do that?
https://karpathy.ai/lexicap/0004-large.html#00:19:52.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yeah. Because there is something, these new distributions, even though they could look
https://karpathy.ai/lexicap/0004-large.html#00:19:54.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
very different from the training distributions, they have things in common. So let me give you
https://karpathy.ai/lexicap/0004-large.html#00:19:57.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
a concrete example. You read a science fiction novel. The science fiction novel, maybe, you
https://karpathy.ai/lexicap/0004-large.html#00:20:02.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
know, brings you in some other planet where things look very different on the surface,
https://karpathy.ai/lexicap/0004-large.html#00:20:07.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
but it's still the same laws of physics. And so you can read the book and you understand
https://karpathy.ai/lexicap/0004-large.html#00:20:15.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
what's going on. So the distribution is very different. But because you can transport
https://karpathy.ai/lexicap/0004-large.html#00:20:20.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
a lot of the knowledge you had from Earth about the underlying cause and effect relationships
https://karpathy.ai/lexicap/0004-large.html#00:20:27.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
and physical mechanisms and all that, and maybe even social interactions, you can now
https://karpathy.ai/lexicap/0004-large.html#00:20:33.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
make sense of what is going on on this planet where, like, visually, for example,
https://karpathy.ai/lexicap/0004-large.html#00:20:38.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
things are totally different.
https://karpathy.ai/lexicap/0004-large.html#00:20:42.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Taking that analogy further and distorting it, let's enter a science fiction world of,
https://karpathy.ai/lexicap/0004-large.html#00:20:45.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
say, Space Odyssey, 2001, with Hal. Or maybe, which is probably one of my favorite AI movies.
https://karpathy.ai/lexicap/0004-large.html#00:20:50.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Me too.
https://karpathy.ai/lexicap/0004-large.html#00:20:59.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And then there's another one that a lot of people love that may be a little bit outside
https://karpathy.ai/lexicap/0004-large.html#00:21:00.480
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
of the AI community is Ex Machina. I don't know if you've seen it.
https://karpathy.ai/lexicap/0004-large.html#00:21:05.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes. Yes.
https://karpathy.ai/lexicap/0004-large.html#00:21:10.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
By the way, what are your views on that movie? Are you able to enjoy it?
https://karpathy.ai/lexicap/0004-large.html#00:21:11.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Are there things I like and things I hate?
https://karpathy.ai/lexicap/0004-large.html#00:21:16.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So you could talk about that in the context of a question I want to ask, which is, there's
https://karpathy.ai/lexicap/0004-large.html#00:21:21.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
quite a large community of people from different backgrounds, often outside of AI, who are concerned
https://karpathy.ai/lexicap/0004-large.html#00:21:26.800