episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
we can in principle do with artificial neural nets, but it's not very convenient and it's
https://karpathy.ai/lexicap/0004-large.html#00:00:46.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
not biologically plausible. And this mismatch, I think this kind of mismatch
https://karpathy.ai/lexicap/0004-large.html#00:00:50.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
may be an interesting thing to study to, A, understand better how brains might do these
https://karpathy.ai/lexicap/0004-large.html#00:00:55.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
things because we don't have good corresponding theories with artificial neural nets, and B,
https://karpathy.ai/lexicap/0004-large.html#00:01:02.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
maybe provide new ideas that we could explore about things that brain do differently and that
https://karpathy.ai/lexicap/0004-large.html#00:01:09.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
we could incorporate in artificial neural nets. So let's break credit assignment up a little bit.
https://karpathy.ai/lexicap/0004-large.html#00:01:18.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes. So what, it's a beautifully technical term, but it could incorporate so many things. So is it
https://karpathy.ai/lexicap/0004-large.html#00:01:23.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
more on the RNN memory side, that thinking like that, or is it something about knowledge, building
https://karpathy.ai/lexicap/0004-large.html#00:01:30.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
up common sense knowledge over time? Or is it more in the reinforcement learning sense that you're
https://karpathy.ai/lexicap/0004-large.html#00:01:37.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
picking up rewards over time for a particular, to achieve a certain kind of goal? So I was thinking
https://karpathy.ai/lexicap/0004-large.html#00:01:44.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
more about the first two meanings whereby we store all kinds of memories, episodic memories
https://karpathy.ai/lexicap/0004-large.html#00:01:50.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in our brain, which we can access later in order to help us both infer causes of things that we
https://karpathy.ai/lexicap/0004-large.html#00:01:59.440
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
are observing now and assign credit to decisions or interpretations we came up with a while ago
https://karpathy.ai/lexicap/0004-large.html#00:02:10.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
when those memories were stored. And then we can change the way we would have reacted or interpreted
https://karpathy.ai/lexicap/0004-large.html#00:02:20.640
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
things in the past, and now that's credit assignment used for learning.
https://karpathy.ai/lexicap/0004-large.html#00:02:29.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So in which way do you think artificial neural networks, the current LSTM, the current architectures
https://karpathy.ai/lexicap/0004-large.html#00:02:33.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
are not able to capture the, presumably you're thinking of very long term?
https://karpathy.ai/lexicap/0004-large.html#00:02:43.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes. So current, the current nets are doing a fairly good jobs for sequences with dozens or
https://karpathy.ai/lexicap/0004-large.html#00:02:50.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
say hundreds of time steps. And then it gets harder and harder and depending on what you have
https://karpathy.ai/lexicap/0004-large.html#00:02:58.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to remember and so on, as you consider longer durations. Whereas humans seem to be able to
https://karpathy.ai/lexicap/0004-large.html#00:03:04.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
do credit assignment through essentially arbitrary times, like I could remember something I did last
https://karpathy.ai/lexicap/0004-large.html#00:03:12.480
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
year. And then now because I see some new evidence, I'm going to change my mind about the way I was
https://karpathy.ai/lexicap/0004-large.html#00:03:16.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
thinking last year. And hopefully not do the same mistake again.
https://karpathy.ai/lexicap/0004-large.html#00:03:23.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
I think a big part of that is probably forgetting. You're only remembering the really important
https://karpathy.ai/lexicap/0004-large.html#00:03:30.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
things. It's very efficient forgetting.
https://karpathy.ai/lexicap/0004-large.html#00:03:36.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes. So there's a selection of what we remember. And I think there are really cool connection to
https://karpathy.ai/lexicap/0004-large.html#00:03:40.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
higher level cognition here regarding consciousness, deciding and emotions,
https://karpathy.ai/lexicap/0004-large.html#00:03:46.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
so deciding what comes to consciousness and what gets stored in memory, which are not trivial either.
https://karpathy.ai/lexicap/0004-large.html#00:03:52.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So you've been at the forefront there all along, showing some of the amazing things that neural
https://karpathy.ai/lexicap/0004-large.html#00:04:00.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
networks, deep neural networks can do in the field of artificial intelligence is just broadly
https://karpathy.ai/lexicap/0004-large.html#00:04:07.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in all kinds of applications. But we can talk about that forever. But what, in your view,
https://karpathy.ai/lexicap/0004-large.html#00:04:12.640
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
because we're thinking towards the future, is the weakest aspect of the way deep neural networks
https://karpathy.ai/lexicap/0004-large.html#00:04:19.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
represent the world? What is that? What is in your view is missing?
https://karpathy.ai/lexicap/0004-large.html#00:04:23.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So current state of the art neural nets trained on large quantities of images or texts
https://karpathy.ai/lexicap/0004-large.html#00:04:29.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
have some level of understanding of, you know, what explains those data sets, but it's very
https://karpathy.ai/lexicap/0004-large.html#00:04:38.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
basic, it's it's very low level. And it's not nearly as robust and abstract and general
https://karpathy.ai/lexicap/0004-large.html#00:04:45.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
as our understanding. Okay, so that doesn't tell us how to fix things. But I think it encourages
https://karpathy.ai/lexicap/0004-large.html#00:04:54.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
us to think about how we can maybe train our neural nets differently, so that they would
https://karpathy.ai/lexicap/0004-large.html#00:05:02.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
focus, for example, on causal explanation, something that we don't do currently with neural
https://karpathy.ai/lexicap/0004-large.html#00:05:14.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
net training. Also, one thing I'll talk about in my talk this afternoon is the fact that
https://karpathy.ai/lexicap/0004-large.html#00:05:20.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
instead of learning separately from images and videos on one hand and from texts on the other
https://karpathy.ai/lexicap/0004-large.html#00:05:27.440
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
hand, we need to do a better job of jointly learning about language and about the world
https://karpathy.ai/lexicap/0004-large.html#00:05:33.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to which it refers. So that, you know, both sides can help each other. We need to have good world
https://karpathy.ai/lexicap/0004-large.html#00:05:42.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
models in our neural nets for them to really understand sentences, which talk about what's
https://karpathy.ai/lexicap/0004-large.html#00:05:50.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
going on in the world. And I think we need language input to help provide clues about
https://karpathy.ai/lexicap/0004-large.html#00:05:57.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
what high level concepts like semantic concepts should be represented at the top levels of our
https://karpathy.ai/lexicap/0004-large.html#00:06:06.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
neural nets. In fact, there is evidence that the purely unsupervised learning of representations
https://karpathy.ai/lexicap/0004-large.html#00:06:13.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
doesn't give rise to high level representations that are as powerful as the ones we're getting
https://karpathy.ai/lexicap/0004-large.html#00:06:21.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
from supervised learning. And so the clues we're getting just with the labels, not even sentences,
https://karpathy.ai/lexicap/0004-large.html#00:06:28.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
is already very, very high level. And I think that's a very important thing to keep in mind.
https://karpathy.ai/lexicap/0004-large.html#00:06:35.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
It's already very powerful. Do you think that's an architecture challenge or is it a data set challenge?
https://karpathy.ai/lexicap/0004-large.html#00:06:42.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Neither. I'm tempted to just end it there. Can you elaborate slightly?
https://karpathy.ai/lexicap/0004-large.html#00:06:49.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Of course, data sets and architectures are something you want to always play with. But
https://karpathy.ai/lexicap/0004-large.html#00:07:02.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
I think the crucial thing is more the training objectives, the training frameworks. For example,
https://karpathy.ai/lexicap/0004-large.html#00:07:06.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
going from passive observation of data to more active agents, which
https://karpathy.ai/lexicap/0004-large.html#00:07:13.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
learn by intervening in the world, the relationships between causes and effects,
https://karpathy.ai/lexicap/0004-large.html#00:07:22.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
the sort of objective functions, which could be important to allow the highest level explanations
https://karpathy.ai/lexicap/0004-large.html#00:07:27.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to rise from the learning, which I don't think we have now, the kinds of objective functions,
https://karpathy.ai/lexicap/0004-large.html#00:07:36.640
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
which could be used to reward exploration, the right kind of exploration. So these kinds of
https://karpathy.ai/lexicap/0004-large.html#00:07:43.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
questions are neither in the data set nor in the architecture, but more in how we learn,
https://karpathy.ai/lexicap/0004-large.html#00:07:50.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
under what objectives and so on. Yeah, I've heard you mention in several contexts, the idea of sort
https://karpathy.ai/lexicap/0004-large.html#00:07:57.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
of the way children learn, they interact with objects in the world. And it seems fascinating
https://karpathy.ai/lexicap/0004-large.html#00:08:04.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
because in some sense, except with some cases in reinforcement learning, that idea
https://karpathy.ai/lexicap/0004-large.html#00:08:08.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
is not part of the learning process in artificial neural networks. So it's almost like,
https://karpathy.ai/lexicap/0004-large.html#00:08:15.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
do you envision something like an objective function saying, you know what, if you
https://karpathy.ai/lexicap/0004-large.html#00:08:21.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
poke this object in this kind of way, it would be really helpful for me to further learn.
https://karpathy.ai/lexicap/0004-large.html#00:08:29.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Right, right.
https://karpathy.ai/lexicap/0004-large.html#00:08:36.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Sort of almost guiding some aspect of the learning.
https://karpathy.ai/lexicap/0004-large.html#00:08:37.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Right, right, right. So I was talking to Rebecca Sacks just a few minutes ago,
https://karpathy.ai/lexicap/0004-large.html#00:08:40.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
and she was talking about lots and lots of evidence from infants seem to clearly pick
https://karpathy.ai/lexicap/0004-large.html#00:08:43.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
what interests them in a directed way. And so they're not passive learners, they focus their
https://karpathy.ai/lexicap/0004-large.html#00:08:52.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
attention on aspects of the world, which are most interesting, surprising in a non trivial way.
https://karpathy.ai/lexicap/0004-large.html#00:09:03.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
That makes them change their theories of the world.
https://karpathy.ai/lexicap/0004-large.html#00:09:10.480
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So that's a fascinating view of the future progress. But on a more maybe boring question,
https://karpathy.ai/lexicap/0004-large.html#00:09:16.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
do you think going deeper and larger, so do you think just increasing the size of the things that
https://karpathy.ai/lexicap/0004-large.html#00:09:26.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
have been increasing a lot in the past few years, is going to be a big thing?
https://karpathy.ai/lexicap/0004-large.html#00:09:33.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
I think increasing the size of the things that have been increasing a lot in the past few years
https://karpathy.ai/lexicap/0004-large.html#00:09:38.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
will also make significant progress. So some of the representational issues that you mentioned,
https://karpathy.ai/lexicap/0004-large.html#00:09:44.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
they're kind of shallow, in some sense.
https://karpathy.ai/lexicap/0004-large.html#00:09:51.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Oh, shallow in the sense of abstraction.
https://karpathy.ai/lexicap/0004-large.html#00:09:54.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
In the sense of abstraction, they're not getting some...
https://karpathy.ai/lexicap/0004-large.html#00:09:58.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
I don't think that having more depth in the network in the sense of instead of 100 layers,
https://karpathy.ai/lexicap/0004-large.html#00:10:00.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
you're going to have more layers. I don't think so. Is that obvious to you?
https://karpathy.ai/lexicap/0004-large.html#00:10:06.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes. What is clear to me is that engineers and companies and labs and grad students will continue
https://karpathy.ai/lexicap/0004-large.html#00:10:11.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to tune architectures and explore all kinds of tweaks to make the current state of the art
https://karpathy.ai/lexicap/0004-large.html#00:10:19.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
slightly ever slightly better. But I don't think that's going to be nearly enough. I think we need
https://karpathy.ai/lexicap/0004-large.html#00:10:25.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
changes in the way that we're considering learning to achieve the goal that these learners actually
https://karpathy.ai/lexicap/0004-large.html#00:10:31.440
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
understand in a deep way the environment in which they are, you know, observing and acting.
https://karpathy.ai/lexicap/0004-large.html#00:10:39.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
But I guess I was trying to ask a question that's more interesting than just more layers.
https://karpathy.ai/lexicap/0004-large.html#00:10:46.640
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
It's basically, once you figure out a way to learn through interacting, how many parameters
https://karpathy.ai/lexicap/0004-large.html#00:10:53.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
it takes to store that information. So I think our brain is quite bigger than most neural networks.
https://karpathy.ai/lexicap/0004-large.html#00:11:00.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Right, right. Oh, I see what you mean. Oh, I'm with you there. So I agree that in order to
https://karpathy.ai/lexicap/0004-large.html#00:11:07.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
build neural nets with the kind of broad knowledge of the world that typical adult humans have,
https://karpathy.ai/lexicap/0004-large.html#00:11:14.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
probably the kind of computing power we have now is going to be insufficient.
https://karpathy.ai/lexicap/0004-large.html#00:11:20.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So the good news is there are hardware companies building neural net chips. And so
https://karpathy.ai/lexicap/0004-large.html#00:11:25.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
it's going to get better. However, the good news in a way, which is also a bad news,
https://karpathy.ai/lexicap/0004-large.html#00:11:30.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
is that even our state of the art, deep learning methods fail to learn models that understand
https://karpathy.ai/lexicap/0004-large.html#00:11:37.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
even very simple environments, like some grid worlds that we have built.
https://karpathy.ai/lexicap/0004-large.html#00:11:46.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Even these fairly simple environments, I mean, of course, if you train them with enough examples,
https://karpathy.ai/lexicap/0004-large.html#00:11:52.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
eventually they get it. But it's just like, instead of what humans might need just
https://karpathy.ai/lexicap/0004-large.html#00:11:56.080