episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
But there's been a lot of success on the supervised learning side.
https://karpathy.ai/lexicap/0004-large.html#00:31:35.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes, yes.
https://karpathy.ai/lexicap/0004-large.html#00:31:39.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And one of the things I'm really passionate about is how humans and robots work together.
https://karpathy.ai/lexicap/0004-large.html#00:31:40.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And in the context of supervised learning, that means the process of annotation. Do you
https://karpathy.ai/lexicap/0004-large.html#00:31:46.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
think about the problem of annotation put in a more interesting way as humans teaching
https://karpathy.ai/lexicap/0004-large.html#00:31:52.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
machines?
https://karpathy.ai/lexicap/0004-large.html#00:32:00.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes.
https://karpathy.ai/lexicap/0004-large.html#00:32:01.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Is there?
https://karpathy.ai/lexicap/0004-large.html#00:32:02.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes. I think it's an important subject. Reducing it to annotation may be useful for somebody
https://karpathy.ai/lexicap/0004-large.html#00:32:03.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
building a system tomorrow. But longer term, the process of teaching, I think, is something
https://karpathy.ai/lexicap/0004-large.html#00:32:09.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
that deserves a lot more attention from the machine learning community. So there are people
https://karpathy.ai/lexicap/0004-large.html#00:32:16.300
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
who have coined the term machine teaching. So what are good strategies for teaching a
https://karpathy.ai/lexicap/0004-large.html#00:32:19.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
learning agent? And can we design and train a system that is going to be a good teacher?
https://karpathy.ai/lexicap/0004-large.html#00:32:24.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So in my group, we have a project called BBI or BBI game, where there is a game or scenario
https://karpathy.ai/lexicap/0004-large.html#00:32:33.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
where there's a learning agent and a teaching agent. Presumably, the teaching agent would
https://karpathy.ai/lexicap/0004-large.html#00:32:42.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
eventually be a human. But we're not there yet. And the role of the teacher is to use
https://karpathy.ai/lexicap/0004-large.html#00:32:48.480
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
its knowledge of the environment, which it can acquire using whatever way brute force
https://karpathy.ai/lexicap/0004-large.html#00:32:57.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to help the learner learn as quickly as possible. So the learner is going to try to learn by
https://karpathy.ai/lexicap/0004-large.html#00:33:04.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
itself, maybe using some exploration and whatever. But the teacher can choose, can have an influence
https://karpathy.ai/lexicap/0004-large.html#00:33:10.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
on the interaction with the learner, so as to guide the learner, maybe teach it the things
https://karpathy.ai/lexicap/0004-large.html#00:33:19.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
that the learner has most trouble with, or just add the boundary between what it knows
https://karpathy.ai/lexicap/0004-large.html#00:33:27.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
and doesn't know, and so on. So there's a tradition of these kind of ideas from other
https://karpathy.ai/lexicap/0004-large.html#00:33:30.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
fields and like tutorial systems, for example, and AI. And of course, people in the humanities
https://karpathy.ai/lexicap/0004-large.html#00:33:36.180
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
have been thinking about these questions. But I think it's time that machine learning
https://karpathy.ai/lexicap/0004-large.html#00:33:45.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
people look at this, because in the future, we'll have more and more human machine interaction
https://karpathy.ai/lexicap/0004-large.html#00:33:48.240
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
with the human in the loop. And I think understanding how to make this work better, all the problems
https://karpathy.ai/lexicap/0004-large.html#00:33:55.440
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
around that are very interesting and not sufficiently addressed. You've done a lot of work with
https://karpathy.ai/lexicap/0004-large.html#00:34:01.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
language, too. What aspect of the traditionally formulated Turing test, a test of natural
https://karpathy.ai/lexicap/0004-large.html#00:34:06.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
language understanding and generation in your eyes is the most difficult of conversation?
https://karpathy.ai/lexicap/0004-large.html#00:34:14.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
What in your eyes is the hardest part of conversation to solve for machines? So I would say it's
https://karpathy.ai/lexicap/0004-large.html#00:34:19.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
everything having to do with the non linguistic knowledge, which implicitly you need in order
https://karpathy.ai/lexicap/0004-large.html#00:34:25.640
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to make sense of sentences, things like the Winograd schema. So these sentences that are
https://karpathy.ai/lexicap/0004-large.html#00:34:32.300
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
semantically ambiguous. In other words, you need to understand enough about the world
https://karpathy.ai/lexicap/0004-large.html#00:34:37.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in order to really interpret properly those sentences. I think these are interesting challenges
https://karpathy.ai/lexicap/0004-large.html#00:34:43.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
for machine learning, because they point in the direction of building systems that both
https://karpathy.ai/lexicap/0004-large.html#00:34:49.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
understand how the world works and this causal relationships in the world and associate that
https://karpathy.ai/lexicap/0004-large.html#00:34:57.300
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
knowledge with how to express it in language, either for reading or writing.
https://karpathy.ai/lexicap/0004-large.html#00:35:03.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
You speak French?
https://karpathy.ai/lexicap/0004-large.html#00:35:12.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes, it's my mother tongue.
https://karpathy.ai/lexicap/0004-large.html#00:35:13.080
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
It's one of the romance languages. Do you think passing the Turing test and all the
https://karpathy.ai/lexicap/0004-large.html#00:35:14.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
underlying challenges we just mentioned depend on language? Do you think it might be easier
https://karpathy.ai/lexicap/0004-large.html#00:35:20.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in French than it is in English, or is independent of language?
https://karpathy.ai/lexicap/0004-large.html#00:35:24.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
I think it's independent of language. I would like to build systems that can use the same
https://karpathy.ai/lexicap/0004-large.html#00:35:28.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
principles, the same learning mechanisms to learn from human agents, whatever their language.
https://karpathy.ai/lexicap/0004-large.html#00:35:37.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Well, certainly us humans can talk more beautifully and smoothly in poetry, some Russian originally.
https://karpathy.ai/lexicap/0004-large.html#00:35:46.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
I know poetry in Russian is maybe easier to convey complex ideas than it is in English.
https://karpathy.ai/lexicap/0004-large.html#00:35:53.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
But maybe I'm showing my bias and some people could say that about French. But of course,
https://karpathy.ai/lexicap/0004-large.html#00:36:02.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
the goal ultimately is our human brain is able to utilize any kind of those languages
https://karpathy.ai/lexicap/0004-large.html#00:36:09.480
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
to use them as tools to convey meaning.
https://karpathy.ai/lexicap/0004-large.html#00:36:15.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yeah, of course, there are differences between languages, and maybe some are slightly better
https://karpathy.ai/lexicap/0004-large.html#00:36:18.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
at some things, but in the grand scheme of things, where we're trying to understand how
https://karpathy.ai/lexicap/0004-large.html#00:36:22.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
the brain works and language and so on, I think these differences are minute.
https://karpathy.ai/lexicap/0004-large.html#00:36:26.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So you've lived perhaps through an AI winter of sorts?
https://karpathy.ai/lexicap/0004-large.html#00:36:32.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Yes.
https://karpathy.ai/lexicap/0004-large.html#00:36:38.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
How did you stay warm and continue your research?
https://karpathy.ai/lexicap/0004-large.html#00:36:39.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Stay warm with friends.
https://karpathy.ai/lexicap/0004-large.html#00:36:44.740
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
With friends. Okay, so it's important to have friends. And what have you learned from the
https://karpathy.ai/lexicap/0004-large.html#00:36:45.740
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
experience?
https://karpathy.ai/lexicap/0004-large.html#00:36:51.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Listen to your inner voice. Don't, you know, be trying to just please the crowds and the
https://karpathy.ai/lexicap/0004-large.html#00:36:53.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
fashion. And if you have a strong intuition about something that is not contradicted by
https://karpathy.ai/lexicap/0004-large.html#00:37:02.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
actual evidence, go for it. I mean, it could be contradicted by people.
https://karpathy.ai/lexicap/0004-large.html#00:37:10.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Not your own instinct of based on everything you've learned?
https://karpathy.ai/lexicap/0004-large.html#00:37:17.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Of course, you have to adapt your beliefs when your experiments contradict those beliefs.
https://karpathy.ai/lexicap/0004-large.html#00:37:20.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
But you have to stick to your beliefs. Otherwise, it's what allowed me to go through those years.
https://karpathy.ai/lexicap/0004-large.html#00:37:28.320
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
It's what allowed me to persist in directions that, you know, took time, whatever other
https://karpathy.ai/lexicap/0004-large.html#00:37:35.000
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
people think, took time to mature and bring fruits.
https://karpathy.ai/lexicap/0004-large.html#00:37:42.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So history of AI is marked with these, of course, it's marked with technical breakthroughs,
https://karpathy.ai/lexicap/0004-large.html#00:37:48.040
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
but it's also marked with these seminal events that capture the imagination of the community.
https://karpathy.ai/lexicap/0004-large.html#00:37:54.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Most recent, I would say, AlphaGo beating the world champion human Go player was one
https://karpathy.ai/lexicap/0004-large.html#00:38:00.980
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
of those moments. What do you think the next such moment might be?
https://karpathy.ai/lexicap/0004-large.html#00:38:06.400
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Okay, so first of all, I think that these so called seminal events are overrated. As
https://karpathy.ai/lexicap/0004-large.html#00:38:12.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
I said, science really moves by small steps. Now what happens is you make one more small
https://karpathy.ai/lexicap/0004-large.html#00:38:22.600
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
step and it's like the drop that, you know, that fills the bucket and then you have drastic
https://karpathy.ai/lexicap/0004-large.html#00:38:30.200
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
consequences because now you're able to do something you were not able to do before.
https://karpathy.ai/lexicap/0004-large.html#00:38:39.480
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Or now, say, the cost of building some device or solving a problem becomes cheaper than
https://karpathy.ai/lexicap/0004-large.html#00:38:43.920
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
what existed and you have a new market that opens up, right? So especially in the world
https://karpathy.ai/lexicap/0004-large.html#00:38:49.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
of commerce and applications, the impact of a small scientific progress could be huge.
https://karpathy.ai/lexicap/0004-large.html#00:38:53.900
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
But in the science itself, I think it's very, very gradual.
https://karpathy.ai/lexicap/0004-large.html#00:39:03.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
And where are these steps being taken now? So there's unsupervised learning.
https://karpathy.ai/lexicap/0004-large.html#00:39:07.800
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So if I look at one trend that I like in my community, so for example, at Milan, my institute,
https://karpathy.ai/lexicap/0004-large.html#00:39:13.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
what are the two hardest topics? GANs and reinforcement learning. Even though in Montreal
https://karpathy.ai/lexicap/0004-large.html#00:39:23.380
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in particular, reinforcement learning was something pretty much absent just two or three
https://karpathy.ai/lexicap/0004-large.html#00:39:31.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
years ago. So there's really a big interest from students and there's a big interest from
https://karpathy.ai/lexicap/0004-large.html#00:39:37.020
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
people like me. So I would say this is something where we're going to see more progress, even
https://karpathy.ai/lexicap/0004-large.html#00:39:44.280
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
though it hasn't yet provided much in terms of actual industrial fallout. Like even though
https://karpathy.ai/lexicap/0004-large.html#00:39:51.560
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
there's AlphaGo, there's no, like Google is not making money on this right now. But I
https://karpathy.ai/lexicap/0004-large.html#00:39:58.680
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
think over the long term, this is really, really important for many reasons.
https://karpathy.ai/lexicap/0004-large.html#00:40:03.360
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
So in other words, I would say reinforcement learning may be more generally agent learning
https://karpathy.ai/lexicap/0004-large.html#00:40:08.960
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
because it doesn't have to be with rewards. It could be in all kinds of ways that an agent
https://karpathy.ai/lexicap/0004-large.html#00:40:13.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
is learning about its environment.
https://karpathy.ai/lexicap/0004-large.html#00:40:17.520
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Now reinforcement learning you're excited about, do you think GANs could provide something,
https://karpathy.ai/lexicap/0004-large.html#00:40:20.720
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
at the moment? Well, GANs or other generative models, I believe, will be crucial ingredients
https://karpathy.ai/lexicap/0004-large.html#00:40:28.840
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
in building agents that can understand the world. A lot of the successes in reinforcement
https://karpathy.ai/lexicap/0004-large.html#00:40:38.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
learning in the past has been with policy gradient, where you just learn a policy, you
https://karpathy.ai/lexicap/0004-large.html#00:40:45.480
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
don't actually learn a model of the world. But there are lots of issues with that. And
https://karpathy.ai/lexicap/0004-large.html#00:40:51.160
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
we don't know how to do model based RL right now. But I think this is where we have to
https://karpathy.ai/lexicap/0004-large.html#00:40:55.760
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
go in order to build models that can generalize faster and better like to new distributions
https://karpathy.ai/lexicap/0004-large.html#00:41:00.880
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
that capture to some extent, at least the underlying causal mechanisms in the world.
https://karpathy.ai/lexicap/0004-large.html#00:41:09.340
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
Last question. What made you fall in love with artificial intelligence? If you look
https://karpathy.ai/lexicap/0004-large.html#00:41:16.120
Yoshua Bengio: Deep Learning | Lex Fridman Podcast #4
back, what was the first moment in your life when you were fascinated by either the human
https://karpathy.ai/lexicap/0004-large.html#00:41:21.480