Yann LeCun: Dark Matter of Intelligence and Self-Supervised Learning
Guests: Yann LeCun
Full transcript · Read the full conversation in order, with its original context.
Contents19 chapters
§01Transcript · 00:00–08:38§02Transcript · 08:40–17:40§03Transcript · 17:41–26:33§04Transcript · 26:35–34:55§05Transcript · 34:56–43:51§06Transcript · 43:53–51:57§07Transcript · 52:00–1:00:42§08Transcript · 1:00:44–1:09:36§09Transcript · 1:09:38–1:18:43§10Transcript · 1:18:44–1:27:44§11Transcript · 1:27:50–1:36:58§12Transcript · 1:36:59–1:45:14§13Transcript · 1:45:16–1:54:38§14Transcript · 1:54:39–2:03:23§15Transcript · 2:03:26–2:11:53§16Transcript · 2:11:56–2:20:43§17Transcript · 2:20:46–2:29:08§18Transcript · 2:29:13–2:37:38§19Transcript · 2:37:41–2:45:04§01Transcript · 00:00–08:38
The following is a conversation with Jan Lekun,
his second time on the podcast.
He is the chief AI scientist at Meta, formerly Facebook,
professor at NYU, Turing Award winner,
one of the seminal figures in the history
of machine learning and artificial intelligence,
and someone who is brilliant and opinionated
in the best kind of way,
and so is always fun to talk to.
This is the Lex Friedman Podcast.
To support it, please check out our sponsors
in the description, and now, here's my conversation
with Jan Lekun.
You co-wrote the article,
"Self-supervised learning, the dark matter of intelligence."
Great title, by the way, with Ishan Misra.
So let me ask, what is self-supervised learning,
and why is it the dark matter of intelligence?
- I'll start by the dark matter part.
There is obviously a kind of learning
that humans and animals are doing
that we currently are not reproducing properly
with machines or with AI, right?
So the most popular approaches to machine learning today
are, or paradigms, I should say,
are supervised learning and reinforcement learning.
And they are extremely inefficient.
Supervised learning requires many samples
for learning anything,
and reinforcement learning requires
a ridiculously large number of trial and errors
for a system to learn anything.
And that's why we don't have self-driving cars.
(Lex laughing)
- That was a big leap from one to the other.
Okay, so that, to solve difficult problems,
you have to have a lot of human annotation
for supervised learning to work,
and to solve those difficult problems
with reinforcement learning, you have to have
some way to maybe simulate that problem
such that you can do that large-scale kind of learning
that reinforcement learning requires.
- Right, so how is it that, you know,
most teenagers can learn to drive a car
in about 20 hours of practice,
whereas even with millions of hours of simulated practice,
a self-driving car can't actually learn
to drive itself properly.
And so obviously we're missing something, right?
And it's quite obvious for a lot of people that,
you know, the immediate response you get from many people
is, well, you know, humans use their background knowledge
to learn faster, and they're right.
Now, how was that background knowledge acquired?
And that's the big question.
So now you have to ask, you know,
how do babies in their first few months of life
learn how the world works?
Mostly by observation,
because they can hardly act in the world.
And they learn an enormous amount
of background knowledge about the world.
That may be the basis of what we call common sense.
This type of learning, it's not learning a task,
it's not being reinforced for anything,
it's just observing the world and figuring out how it works.
Building world models, learning world models.
How do we do this?
And how do we reproduce this in machines?
So self-supervised learning is, you know,
one instance or one attempt
at trying to reproduce this kind of learning.
- Okay, so you're looking at just observation,
so not even the interacting part of a child.
It's just sitting there watching mom and dad walk around,
pick up stuff, all of that.
- That's what you mean by background knowledge.
- Perhaps not even watching mom and dad,
just, you know, watching the world go by.
- Just having eyes open or having eyes closed,
or the very act of opening and closing eyes
that the world appears and disappears,
all that basic information.
And you're saying in order to learn to drive,
like the reason humans are able to learn to drive quickly,
some faster than others,
is because of the background knowledge
they were able to watch cars operate in the world
in the many years leading up to it,
the physics of basic objects, all that kind of stuff.
- That's right.
I mean, the basic physics of objects,
you don't even know, you don't even need to know,
you know, how a car works, right?
Because that you can learn fairly quickly.
I mean, the example I use very often
is you're driving next to a cliff,
and you know in advance,
because of your understanding of intuitive physics,
that if you turn the wheel to the right,
the car will veer to the right,
will run off the cliff, fall off the cliff,
and nothing good will come out of this, right?
But if you are a sort of, you know,
tabula rasa reinforcement learning system
that doesn't have a model of the world,
you have to repeat falling off this cliff
thousands of times before you figure out it's a bad idea.
And then a few more thousand times
before you figure out how to not do it.
And then a few more million times
before you figure out how to not do it
in every situation you ever encounter.
- So self-supervised learning still has to have
some source of truth being told to it by somebody.
- So you have to figure out a way without human assistance
or without significant amount of human assistance
to get that truth from the world.
So the mystery there is how much signal is there,
how much truth is there that the world gives you,
whether it's the human world,
like you watch YouTube or something like that,
or it's the more natural world.
So how much signal is there?
- So here's the trick.
There is way more signal in sort of a self-supervised setting
than there is in either a supervised
or reinforcement setting.
And this is going to my analogy of the cake.
Le cake as someone has called it,
where when you try to figure out how much information
you ask the machine to predict
and how much feedback you give the machine at every trial,
in reinforcement learning,
you give the machine a single scaler,
you tell the machine you did good, you did bad,
and you only tell this to the machine once in a while.
When I say you, it could be the universe
telling the machine, right?
But it's just one scaler.
And so as a consequence,
you cannot possibly learn something very complicated
without many, many, many trials
where you get many, many feedbacks of this type.
Supervised learning, you give a few bits to the machine
at every sample.
Let's say you're training a system on, you know,
recognizing images on ImageNet.
There is 1000 categories,
that's a little less than 10 bits of information per sample.
But self-supervised learning here is a setting.
Ideally, we don't know how to do this yet,
but ideally you would show a machine a segment of video
and then stop the video and ask the machine to predict
what's going to happen next.
And so you let the machine predict,
and then you let time go by
and show the machine what actually happened.
And hope the machine will, you know,
learn to do a better job at predicting next time around.
There's a huge amount of information you give the machine
because it's an entire video clip of, you know,
the future after the video clip you fed it
in the first place.
- So both for language and for vision,
there's a subtle, seemingly trivial construction,
but maybe that's representative
of what is required to create intelligence,
which is filling the gap.
- Filling the gaps.
- Sounds dumb, but can you,
it is possible you could solve all of intelligence
in this way, just for both language,
just give a sentence and continue it,
or give a sentence and there's a gap in it,
some words blanked out,
and you fill in what words go there.
For vision, you give a sequence of images
and predict what's gonna happen next,
or you fill in what happened in between.
Do you think it's possible that formulation alone,
as a signal for self-supervised learning,
can solve intelligence for vision and language?
- I think that's our best shot at the moment.
So whether this will take us all the way to,
you know, human level intelligence or something,
or just cat level intelligence is not clear,
but among all the possible approaches
that people have proposed, I think it's our best shot.
So I think this idea of an intelligence system
filling in the blanks, either predicting the future,
inferring the past, filling in missing information,
I'm currently filling the blank of what is behind your head
and what your head looks like from the back,
because I have basic knowledge about how humans are made.
And I don't know if you're gonna,
what are you gonna say, at which point you're gonna speak,
whether you're gonna move your head this way or that way,
which way you're gonna look.
§02Transcript · 08:40–17:40
But I know you're not gonna just dematerialize
and reappear three meters down the hall,
because I know what's possible and what's impossible,
according to intuitive physics.
- So you have a model of what's possible and what's impossible
and then you'd be very surprised if it happens,
and then you'll have to reconstruct your model.
- Right, so that's the model of the world.
It's what tells you what fills in the blanks.
So given your partial information
about the state of the world, given by your perception,
your model of the world fills in the missing information.
And that includes predicting the future,
retrodicting the past,
filling in things you don't immediately perceive.
- And that doesn't have to be purely generic vision
or visual information or generic language.
You can go to specifics,
like predicting what control decision you make
when you're driving in a lane.
You have a sequence of images from a vehicle,
and then you have information,
if you record it on video, where the car ended up going.
So you can go back in time and predict where the car went
based on the visual information.
That's very specific, domain-specific.
- Right, but the question is whether we can come up
with sort of a generic method for training machines
to do this kind of prediction or filling in the blanks.
So right now, this type of approach
has been unbelievably successful
in the context of natural language processing.
Every model in natural language processing
is pre-trained in self-supervised manner
to fill in the blanks.
You show it a sequence of words, you remove 10% of them,
and then you train some gigantic neural net
to predict the words that are missing.
And once you've pre-trained that network,
you can use the internal representation learned by it
as input to something that you trained, supervised,
or whatever.
That's been incredibly successful.
Not so successful in images, although it's making progress.
And it's based on sort of manual data augmentation.
We can go into this later.
But what has not been successful yet is training from video.
So getting a machine to learn to represent the visual world,
for example, by just watching video.
Nobody has really succeeded in doing this.
- Okay, well, let's kind of give a high-level overview.
What's the difference in kind and in difficulty
between vision and language?
So you said people haven't been able to really kind of crack
the problem of vision open
in terms of self-supervised learning,
but that may not be necessarily
because it's fundamentally more difficult.
Maybe like when we're talking about achieving,
like passing the Turing test in the full spirit
of the Turing test in language might be harder than vision.
That's not obvious.
So in your view, which is harder,
or perhaps are they just the same problem?
The farther we get to solving each,
the more we realize it's all the same thing.
It's all the same cake.
- I think what I'm looking for are methods
that make them look essentially like the same cake,
but currently they're not.
And the main issue with learning world models
or learning predictive models is that
the prediction is never a single thing,
because the world is not entirely predictable.
It may be deterministic or stochastic.
We can get into the philosophical discussion about it,
but even if it's deterministic,
it's not entirely predictable.
And so if I play a short video clip
and then I ask you to predict what's going to happen next,
there's many, many plausible continuations
for that video clip.
And the number of continuation grows
with the interval of time that you're asking the system
to make a prediction for.
And so one big question with self-supervised learning
is how you represent this uncertainty,
how you represent multiple discrete outcomes,
how you represent a sort of continuum
of possible outcomes, et cetera.
And if you are a sort of a classical machine learning person,
you say, "Oh, you just represent a distribution."
And that we know how to do when we're predicting words,
missing words in the text,
because you can have a neural net give a score
for every word in a dictionary.
It's a big list of numbers, maybe 100,000 or so.
And you can turn them into a probability distribution
that tells you when I say a sentence,
the cat is chasing the blank in the kitchen.
There are only a few words that make sense there.
It could be a mouse or it could be a lizard spot
or something like that.
And if I say the blank is chasing the blank in the savanna,
you also have a bunch of plausible options
for those two words, right?
Because you have kind of underlying reality
that you can refer to to sort of fill in those blanks.
So you cannot say for sure in the savanna
if it's a lion or a cheetah or whatever,
you cannot know if it's a zebra or a gnu or whatever,
wildebeest, the same thing.
- Yeah.
- But you can represent the uncertainty
by just a long list of numbers.
Now, if I do the same thing with video
and I ask you to predict a video clip,
it's not a discrete set of potential frames.
You have to have some way of representing
a sort of infinite number of plausible continuations
of multiple frames in a high dimensional continuous space.
And we just have no idea how to do this properly.
- Finite high dimensional.
So like you--
- It's finite high dimensional, yes.
- Just like the words,
they try to get it down to a small finite set
of like under a million, something like that.
- Something like that.
- I mean, it's kind of ridiculous
that we're doing a distribution
of every single possible word for language and it works.
It feels like that's a really dumb way to do it.
Like there seems to be like there should be
some more compressed representation
of the distribution of the words.
- You're right about that.
- And so--
- I agree.
- Do you have any interesting ideas
about how to represent all of reality in a compressed way
such that you can form a distribution over it?
- That's one of the big questions, how do you do that?
But I mean, what's kind of,
another thing that really is stupid about,
I shouldn't say stupid,
but like simplistic about current approaches
to self-supervised learning in NLP and text
is that not only do you represent
a giant distribution of words,
but for multiple words that are missing,
those distributions are essentially independent
of each other.
And you don't pay too much of a price for this.
So you can't, so the system,
in the sentence that I gave earlier,
if it gives a certain probability for a lion and a cheetah,
and then a certain probability for gazelle, wildebeest,
and zebra, those two probabilities
are independent of each other.
And it's not the case that those things are independent.
Lions actually attack like bigger animals than cheetahs.
So, there's a huge independence hypothesis in this process,
which is not actually true.
The reason for this is that we don't know how to represent
properly distributions over combinatorial sequences
of symbols, essentially,
because the number grows exponentially
with the length of the symbols.
And so we have to use tricks for this,
but those techniques can get around,
like don't even deal with it.
So the big question is,
would there be some sort of abstract
latent representation of text that would say that,
when I switch lion for gazelle,
lion for cheetah, I also have to switch zebra for gazelle.
- Yeah, so this independence assumption,
let me throw some criticism at you that I often hear
and see how you respond.
So this kind of filling in the blanks is just statistics.
You're not learning anything,
like the deep underlying concepts.
You're just mimicking stuff from the past.
You're not learning anything new
such that you can use it to generalize about the world.
Or, okay, let me just say the crude version,
which is it's just statistics.
It's not intelligence.
What do you have to say to that?
What do you usually say to that
if you kind of hear this kind of thing?
- I don't get into those discussions
because they're kind of pointless.
So first of all, it's quite possible
that intelligence is just statistics.
It's just statistics of a particular kind.
- But this is the philosophical question.
Is it possible that intelligence is just statistics?
- Yeah.
§03Transcript · 17:41–26:33
But what kind of statistics?
So if you are asking the question,
are the models of the world that we learn,
do they have some notion of causality?
Yes.
So if the criticism comes from people who say,
current machine learning system don't care about causality,
which by the way is wrong, I agree with them.
Your model of the world should have your actions
as one of the inputs,
and that will drive you to learn causal models of the world
where you know what intervention in the world
will cause what result,
or you can do this by observation of other agents
acting in the world and observing the effect,
other humans, for example.
So I think at some level of description,
intelligence is just statistics,
but that doesn't mean you won't have models
that have deep mechanistic explanation for what goes on.
The question is, how do you learn them?
That's the question I'm interested in.
Because a lot of people who actually voice their criticism
say that those mechanistic model
have to come from someplace else.
They have to come from human designers,
they have to come from I don't know what.
And obviously we learn them.
Or if we don't learn them as an individual,
nature learn them for us using evolution.
So regardless of what you think,
those processes have been learned somehow.
- So if you look at the human brain,
just like when we humans introspect
about how the brain works,
it seems like when we think about what is intelligence,
we think about the high level stuff,
like the models we've constructed,
concepts like cognitive science,
like concepts of memory and reasoning module,
almost like these high level modules.
Is this serve as a good analogy?
Like are we ignoring the dark matter,
the basic low level mechanisms,
just like we ignore the way the operating system works,
we're just using the high level software.
We're ignoring that at the low level,
the neural network might be doing something like statistics.
Like me, sorry to use this word
probably incorrectly and crudely,
but doing this kind of fill in the gap kind of learning
and just kind of updating the model constantly
in order to be able to support the raw sensory information,
to predict it and then adjust to the prediction
when it's wrong.
But like when we look at our brain at the high level,
it feels like we're playing chess,
like we're playing with high level concepts
and we're stitching them together
and we're putting them into long-term memory.
But really what's going underneath
is something we're not able to introspect,
which is this kind of simple, large neural network
that's just filling in the gaps.
- Right, well, okay, so there's a lot of questions
and a lot of answers there.
Okay, so first of all,
there's a whole school of thought in neuroscience,
computational neuroscience in particular,
that likes the idea of predictive coding,
which is really related to the idea I was talking about
in self-supervised learning.
So everything is about prediction,
the essence of intelligence is the ability to predict
and everything the brain does is trying to predict
everything from everything else.
Okay, and that's really sort of the underlying principle,
if you want, that self-supervised learning
is trying to kind of reproduce this idea of prediction
as kind of an essential mechanism
of task-independent learning, if you want.
The next step is what kind of intelligence
are you interested in reproducing?
And of course, we all think about trying to reproduce
high-level cognitive processes in humans,
but with machines, we're not even at the level
of even reproducing the learning processes in a cat brain.
The most intelligent or intelligent systems
don't have as much common sense as a house cat.
So how is it that cats learn?
And cats don't do a whole lot of reasoning.
They certainly have causal models.
They certainly have, because many cats can figure out
how they can act on the world to get what they want.
They certainly have a fantastic model of intuitive physics,
certainly the dynamics of their own bodies,
but also of praise and things like that, right?
So they're pretty smart.
They only do this with about 800 million neurons.
We are not anywhere close
to reproducing this kind of thing.
So to some extent, I could say,
let's not even worry about the high-level cognition
and long-term planning and reasoning that humans can do
until we figure out,
can we even reproduce what cats are doing?
Now, that said, this ability to learn world models,
I think is the key to the possibility
of learning machines that can also reason.
So whenever I give a talk, I say,
there are three main challenges in machine learning.
The first one is getting machines to learn
to represent the world,
and I'm proposing self-supervised learning.
The second is getting machines to reason
in ways that are compatible
with essentially gradient-based learning,
because this is what deep learning is all about, really.
And the third one is something we have no idea how to solve,
at least I have no idea how to solve,
is can we get machines to learn hierarchical representations
of action plans?
We know how to train them
to learn hierarchical representations of perception,
with convolutional nets and things like that,
and transformers, but what about action plans?
Can we get them to spontaneously learn
good hierarchical representations of actions?
- Also gradient-based.
- Yeah, all of that needs to be somewhat differentiable
so that you can apply sort of gradient-based learning,
which is really what deep learning is about.
- So it's background, knowledge, ability to reason
in a way that's differentiable,
that is somehow connected, deeply integrated
with that background knowledge,
or builds on top of that background knowledge,
and then given that background knowledge,
be able to make hierarchical plans in the world.
- So if you take classical optimal control,
there's something in classical optimal control
called model predictive control,
and it's been around since the early '60s.
NASA uses that to compute trajectories of rockets.
And the basic idea is that you have a predictive model
of the rocket, let's say,
or whatever system you intend to control,
which, given the state of the system at time t,
and given an action that you're taking on the system,
so for a rocket to be thrust,
and all the controls you can have,
it gives you the state of the system
at time t plus delta t, right?
So basically a differential equation, something like that.
And if you have this model,
and you have this model in the form
of some sort of neural net,
or some sort of set of formula
that you can back propagate gradient through,
you can do what's called model predictive control,
or gradient-based model predictive control.
So you have, you can unroll that model in time,
you feed it a hypothesized sequence of actions,
and then you have some objective function
that measures how well, at the end of the trajectory,
the system has succeeded or matched what you want it to do.
Is it a robot harm?
Have you grasped the object you want to grasp?
If it's a rocket, are you at the right place
near the space station?
Things like that.
And by back propagation through time,
and again, this was invented in the 1960s
by optimal control theorists,
you can figure out what is the optimal sequence of actions
that will get my system to the best final state.
So that's a form of reasoning.
It's basically planning,
and a lot of planning systems in robotics
are actually based on this.
And you can think of this as a form of reasoning.
So to take the example of the teenager driving a car again,
you have a pretty good dynamical model of the car,
it doesn't need to be very accurate,
but you know, again, that if you turn the wheel
to the right and there is a cliff,
you're gonna run off the cliff, right?
You don't need to have a very accurate model
to predict that.
And you can run this in your mind
and decide not to do it for that reason,
because you can predict in advance
that the result is gonna be bad.
So you can sort of imagine different scenarios
and then employ or take the first step in the scenario
that is most favorable,
and then repeat the process of planning.
That's called receding horizon model predictive control.
So even all those things have names,
going back decades.
§04Transcript · 26:35–34:55
And so if you're not, you know,
in classical optimal control,
the model of the world is not generally learned.
There's, you know, sometimes a few parameters
you have to identify,
that's called systems identification.
But generally the model is mostly deterministic
and mostly built by hand.
So the big question of AI,
I think the big challenge of AI for the next decade
is how do we get machines to run predictive models
of the world that deal with uncertainty
and deal with the real world in all this complexity.
So it's not just the trajectory of a rocket,
which you can reduce to first principles.
It's not even just the trajectory of a robot arm,
which again, you can model by, you know,
careful mathematics, but it's everything else,
everything we observe in the world, you know,
people behavior, you know, physical systems
that involve collective phenomena like water or, you know,
trees and, you know, branches in a tree or something,
or like complex things that, you know,
humans have no trouble developing abstract representations
and predictive model for,
but we still don't know how to do with machines.
- Where do you put in these three,
maybe in the planning stages,
the game theoretic nature of this world,
where your actions not only respond
to the dynamic nature of the world, the environment,
but also affect it.
So if there's other humans involved,
is this point number four,
or is it somehow integrated
into the hierarchical representation of action in your view?
- I think it's integrated.
It's just that now your model of the world has to deal with,
you know, it just makes it more complicated, right?
The fact that humans are complicated
and not easily predictable,
that makes your model of the world much more complicated,
that much more complicated.
- Well, there's a chess, I mean,
I suppose chess is an analogy.
So Monte Carlo tree search.
I mean, there's a, I go, you go, I go, you go.
Like Andrej Karpathy recently gave a talk
at MIT about car doors.
I think there's some machine learning too,
but mostly car doors.
And there's a dynamic nature to the car,
like the person opening the door checking.
I mean, he wasn't talking about that.
He was talking about the perception problem
of what the ontology of what defines a car door,
this big philosophical question.
But to me, it was interesting,
'cause like it's obvious
that the person opening the car doors,
they're trying to get out, like here in New York,
trying to get out of the car.
You slowing down is going to signal something.
You speeding up is gonna signal something.
That's a dance.
It's a asynchronous chess game.
I don't know.
So it feels like it's not just,
I mean, I guess you can integrate all of them
to one giant model,
like the entirety of these little interactions.
'Cause it's not as complicated as chess.
It's just like a little dance.
We do like a little dance together
and then we figure it out.
- Well, in some ways it's way more complicated than chess
because it's continuous,
it's uncertain in a continuous manner.
- It doesn't feel more complicated.
- But it doesn't feel more complicated
because that's what we've evolved to solve.
This is the kind of problem we've evolved to solve.
And so we're good at it
because nature has made us good at it.
Nature has not made us good at chess.
We completely suck at chess.
In fact, that's why we designed it as a game,
is to be challenging.
And if there is something that recent progress
in chess and Go has made us realize
is that humans are really terrible at those things,
like really bad.
There was a story before AlphaGo
that the best Go players thought
there were maybe two or three stones
behind an ideal player that they would call God.
In fact, no, there are like nine or 10 stones behind,
I mean, which is bad.
So we're not good at,
and it's because we have limited working memory.
We're not very good at doing this tree exploration
that computers are much better at doing than we are,
but we are much better
at learning differentiable models of the world.
I mean, I say differentiable in a kind of,
I should say not differentiable in the sense
that we run backprop through it,
but in the sense that our brain has some mechanism
for estimating gradients of some kind.
And that's what makes us efficient.
So if you have an agent that consists of a model
of the world, which in the human brain
is basically the entire front half of your brain,
an objective function, which in humans
is a combination of two things.
There is your sort of intrinsic motivation module,
which is in the basal ganglia,
the base of your brain.
That's the thing that measures pain and hunger
and things like that, like immediate feelings and emotions.
And then there is the equivalent
of what people in reinforcement learning call a critic,
which is a sort of module that predicts ahead
what the outcome of a situation will be.
And so it's not a cost function,
but it's sort of not an objective function,
but it's sort of a trained predictor
of the ultimate objective function.
And that also is differentiable.
And so if all of this is differentiable,
your cost function, your critic, your world model,
then you can use gradient-based type methods
to do planning, to do reasoning, to do learning,
to do all the things that we'd like
an intelligent agent to do.
- And a gradient-based learning,
like what's your intuition?
That's probably at the core of what can solve intelligence.
So you don't need like a logic-based reasoning in your view.
- I don't know how to make logic-based reasoning
compatible with efficient learning.
- Yeah.
- Okay, I mean, there is a big question,
perhaps a philosophical question.
I mean, it's not that philosophical,
but that we can ask is that, you know,
all the learning algorithms we know
from engineering and computer science
proceed by optimizing some objective function.
- Yeah. - Right?
So one question we may ask is,
does learning in the brain minimize an objective function?
I mean, it could be a composite
of multiple objective functions,
but it's still an objective function.
Second, if it does optimize an objective function,
does it do it by some sort of gradient estimation?
You know, it doesn't need to be backprop,
but some way of estimating the gradient in an efficient manner
whose complexity is on the same order of magnitude
as actually running the inference.
'Cause you can't afford to do things like, you know,
perturbing a weight in your brain
to figure out what the effect is,
and then sort of, you know,
you can't do sort of estimating gradient by perturbation.
It's, to me, it seems very implausible
that the brain uses some sort of, you know,
zeroth-order black box gradient-free optimization
because it's so much less efficient
than gradient optimization.
So it has to have a way of estimating gradient.
- Is it possible that some kind of logic-based reasoning
emerges in pockets as a useful, like you said,
if the brain is an objective function,
maybe it's a mechanism for creating objective functions.
It's a mechanism for creating knowledge bases, for example,
that can then be queried.
Like, maybe it's like an efficient representation
of knowledge that's learned in a gradient-based way
or something like that.
- Well, so I think there is a lot of different types
of intelligence.
So first of all, I think the type of logical reasoning
that we think about, that we are, you know,
maybe stemming from, you know, sort of classical AI
of the 1970s and '80s,
I think humans use that relatively rarely
and are not particularly good at it.
- But we judge each other based on our ability
to solve those rare problems.
It's called an IQ test.
- I don't think so.
Like, I'm not very good at chess.
- Yes, I'm judging you this whole time because,
well, we actually-
- With your, you know, heritage,
I'm sure you're good at chess.
- No, stereotypes.
Not all stereotypes are true.
§05Transcript · 34:56–43:51
- Well, I'm terrible at chess.
So, you know, but I think perhaps another type
of intelligence that I have is this, you know,
ability of sort of building models to the world from,
you know, reasoning, obviously, but also data.
And those models generally are more kind of analogical,
right?
So it's reasoning by simulation and by analogy,
where you use one model to apply to a new situation,
even though you've never seen that situation,
you can sort of connect it to a situation
you've encountered before.
And your reasoning is more, you know,
akin to some sort of internal simulation.
So you're kind of simulating what's happening
when you're building, I don't know,
a box out of wood or something, right?
You kind of imagine in advance,
like what would be the result of, you know,
cutting the wood in this particular way?
Are you going to use, you know, screws and nails or whatever?
When you are interacting with someone,
you also have a model of that person
and sort of interact with that person, you know,
having this model in mind to kind of tell the person
what you think is useful to them.
So I think this ability to construct models of the world
is basically the essence of intelligence.
And the ability to use it then to plan actions
that will fulfill a particular criterion,
of course, is necessary as well.
- So I'm going to ask you a series of impossible questions
as we keep asking, as I've been doing.
So if that's the fundamental sort of dark matter
of intelligence, this ability to form a background model,
what's your intuition about how much knowledge is required?
You know, I think dark matter,
you can put a percentage on it
of the composition of the universe
and how much of it is dark matter,
how much of it is dark energy.
How much information do you think is required
to be a house cat?
So you have to be able to, when you see a box,
go in it, when you see a human compute the most evil action,
if there's a thing that's near an edge, you knock it off.
All of that, plus the extra stuff you mentioned,
which is a great self-awareness of the physics
of your own body and the world.
How much knowledge is required, do you think, to solve it?
I don't even know how to measure an answer to that question.
- I'm not sure how to measure it,
but whatever it is, it fits in about 800,000 neurons.
800 million neurons, sorry.
- The representation does.
- Everything, all knowledge, everything, right?
It's less than a billion.
A dog is two billion, but a cat is less than one billion.
And so multiply that by a thousand
and you get the number of synapses.
And I think almost all of it is learned
through a sort of self-supervised learning,
although I think a tiny sliver
is learned through reinforcement learning,
and certainly very little through
classical supervised learning,
although it's not even clear how supervised learning
actually works in the biological world.
So I think almost all of it is self-supervised learning,
but it's driven by the sort of ingrained objective functions
that a cat or a human have at the base of their brain,
which kind of drives their behavior.
So nature tells us you're hungry.
It doesn't tell us how to feed ourselves.
That's something that the rest of our brain
has to figure out, right?
- What's interesting is there might be more
like deeper objective functions
that are allowing the whole thing.
So hunger may be some kind of,
now you go to like neurobiology,
it might be just the brain
trying to maintain homeostasis.
So hunger is just one of the human perceivable symptoms
of the brain being unhappy with the way things are currently.
It could be just like one really dumb
objective function at the core.
- But that's how behavior is driven.
The fact that the basal ganglia
drive us to do things that are different
from say an orangutan or certainly a cat
is what makes human nature
versus orangutan nature versus cat nature.
So for example,
our basal ganglia drives us to
seek the company of other humans.
And that's because nature has figured out
that we need to be social animals
for our species to survive.
And it's true of many primates.
It's not true of orangutans.
Orangutans are solitary animals.
They don't seek the company of others.
In fact, they avoid them.
In fact, they scream at them when they come too close
because they're territorial.
'Cause for their survival,
evolution has figured out that's the best thing.
I mean, they're occasionally social, of course,
for reproduction and stuff like that.
But they're mostly solitary.
So all of those behaviors are not part of intelligence.
People say, "Oh, you're never gonna have
"intelligent machines because human intelligence is social."
But then you look at orangutans, you look at octopus.
Octopus never know their parents.
They barely interact with any other.
And they get to be really smart
in less than a year, in like half a year.
In a year, they're adults.
In two years, they're dead.
So there are things that we think, as humans,
are intimately linked with intelligence,
like social interaction, like language.
I think we give way too much importance to language
as a substrate of intelligence as humans
because we think our reasoning is so linked with language.
- So to solve the house cat intelligence problem,
you think you could do it on a desert island?
You could have-- - Pretty much.
- You could just have a cat sitting there
looking at the waves, at the ocean waves,
and figure a lot of it out.
- It needs to have sort of the right set of drives
to kind of get it to do the thing
and learn the appropriate things, right?
But like, for example, baby humans are driven
to learn to stand up and walk.
That's kind of, this desire is hardwired.
How to do it precisely is not, that's learned.
But the desire to-- - To walk?
- Move around and stand up, that's sort of hardwired.
But it's very simple to hardwire this kind of stuff.
- Oh, like the desire to, well, that's interesting.
You're hardwired to want to walk.
That's not a, there's gotta be a deeper need for walking.
I think it was probably socially imposed by society
that you need to walk, all the other bipedal--
- No, like a lot of simple animals that, you know,
would probably walk without ever watching
any other members of the species.
- It seems like a scary thing to have to do
'cause you suck at bipedal walking at first.
It seems crawling is much safer, much more,
like, why are you in a hurry?
- Well, because you have this thing
that drives you to do it, you know,
which is sort of part of the sort of human development.
- Is that understood, actually, what--
- Not entirely, no.
- What's the reason to get on two feet?
It's really hard.
Like, most animals don't get on two feet.
Why is that? - Well, they get on four feet.
You know, many mammals get on four feet.
- Yeah, they do get-- - Very quickly.
Some of them, extremely quickly.
- But I don't, you know, like, from the last time
I've interacted with a table, that's much more stable
than a thing on two legs.
It's just a really hard problem.
- Yeah, I mean, birds have figured it out with two feet.
- Well, technically, we can go into ontology.
They have four.
I guess they have two feet.
- They have two feet. - Chickens.
- You know, dinosaurs have two feet, many of them.
- Allegedly.
I'm just now learning that T-Rex was eating grass,
not other animals.
T-Rex might have been a friendly pet.
What do you think about, I don't know if you looked at
the test for general intelligence
that Francois Chollet put together.
I don't know if you got a chance to look at
that kind of thing.
What's your intuition about how to solve
like an IQ type of test?
- I don't know.
I think it's so outside of my radar screen
that it's not really relevant, I think, in the short term.
- Well, I guess one way to ask, another way,
perhaps more closer to what, to your work,
is like, how do you solve MNIST
with very little example data?
- That's right, and that's the answer to this probably,
is self-supervised learning.
Just learn to represent images, and then learning
to recognize handwritten digits on top of this
will only require a few samples.
§06Transcript · 43:53–51:57
And we observe this in humans, right?
You show a young child a picture book
with a couple of pictures of an elephant, and that's it.
The child knows what an elephant is.
And we see this today with practical systems
that we train image recognition systems
with enormous amounts of images,
either completely self-supervised
or very weakly supervised.
For example, you can train a neural net
to predict whatever hashtag people type on Instagram, right?
Then you can do this with billions of images
'cause there's billions per day that are showing up.
So the amount of training data there
is essentially unlimited.
And then you take the output representation,
you know, a couple layers down from the output
of what the system learned,
and feed this as input to a classifier
for any object in the world that you want,
and it works pretty well.
So that's transfer learning, okay?
Or weakly supervised transfer learning.
People are making very, very fast progress
using self-supervised learning
for this kind of scenario as well.
And, you know, my guess is that
that's gonna be the future.
- For self-supervised learning,
how much cleaning do you think is needed
for filtering malicious signal,
or what's a better term?
But like a lot of people use hashtags on Instagram
to get like good SEO
that doesn't fully represent the contents of the image.
Like they'll put a picture of a cat
and hashtag it with like science, awesome, fun.
I don't know, all kinds.
Why would you put science?
That's not very good SEO.
- The way my colleagues who worked on this project
at Facebook now Meta, Meta.AI,
a few years ago dealt with this is that
they only selected something like 17,000 tags
that correspond to kind of physical things or situations.
Like, you know, that has some visual content.
So, you know, you wouldn't have like #tbt
or anything like that.
- Oh, so they keep a very select set of hashtags
is what you're saying? - Yeah.
- Okay.
- It's still on the order of, you know, 10 to 20,000.
So it's fairly large.
- Okay.
Can you tell me about data augmentation?
What the heck is data augmentation?
And how is it used, maybe contrast of learning for video?
What are some cool ideas here?
- Right, so data augmentation,
I mean, first data augmentation, you know,
is the idea of artificially increasing the size
of your training set by distorting the images
that you have in ways that don't change
the nature of the image, right?
So you take, you do MNIST,
you can do data augmentation on MNIST.
And people have done this since the 1990s, right?
You take an MNIST digit and you shift it a little bit,
or you change the size or rotate it, skew it,
you know, et cetera.
- Add noise.
- Add noise, et cetera.
And it works better.
If you train a supervised classifier with augmented data,
you're gonna get better results.
Now it's become really interesting
over the last couple of years,
because a lot of self-supervised learning techniques
to pre-train vision systems are based on data augmentation.
And the basic techniques is originally inspired
by techniques that I worked on in the early '90s
and Geoff Hinton worked on also in the early '90s.
They were sort of parallel work.
I used to call this Siamese networks.
So basically you take two identical copies
of the same network, they share the same weights,
and you show two different views of the same object.
Either those two different views may have been obtained
by data augmentation, or maybe it's two different views
of the same scene from a camera that you moved
or at different times or something like that, right?
Or two pictures of the same person, things like that.
And then you train this neural net,
those two identical copies of this neural net
to produce an output representation, a vector,
in such a way that the representation
for those two images are as close to each other as possible,
as identical to each other as possible, right?
Because you want the system to basically learn a function
that will be invariant, that will not change,
whose output will not change when you transform those inputs
in those particular ways, right?
So that's easy to do.
What's complicated is how do you make sure
that when you show two images that are different,
the system will produce different things?
Because if you don't have a specific provision for this,
the system will just ignore the input.
When you train it, it will end up ignoring the input
and just produce a constant vector
that is the same for every input, right?
That's called a collapse.
Now, how do you avoid collapse?
So there's two ideas.
One idea that I proposed in the early '90s
with my colleagues at Bell Labs,
Jane Bromley and a couple other people,
which we now call contrastive learning,
which is to have negative examples, right?
So you have pairs of images that you know are different
and you show them to the network and those two copies,
and then you push the two output vectors
away from each other.
And it will eventually guarantee
that things that are semantically similar
produce similar representations
and things that are different
produce different representations.
We actually came up with this idea
for a project of doing signature verification.
So we would collect signatures
from multiple signatures on the same person
and then train a neural net
to produce the same representation.
And then force the system to produce different representation
from different signatures.
This was actually, the problem was proposed
by people from what was a subsidiary of AT&T at the time
called NCR.
And they were interested in storing a representation
of the signature on the 80 bytes
of the magnetic strip of a credit card.
So we came up with this idea of having a neural net
with 80 outputs that we quantized on bytes
so that we could encode the-
- And that encoding was then used to compare
whether the signature matches or not.
- That's right.
So then you would sign, it would run through the neural net
and then you would compare the output vector
to whatever is stored on your card.
- Did it actually work?
- It worked, but they ended up not using it.
Because nobody cares actually.
I mean, the American financial payment system
is incredibly lax in that respect compared to Europe.
- Oh, with the signatures?
What's the purpose of signatures anyway?
This is very-
- Nobody looks at them, nobody cares.
- It's, yeah.
- Yeah, no.
So that's contrastive learning, right?
So you need positive and negative pairs.
And the problem with that is that,
even though I had the original paper on this,
I'm actually not very positive about it
because it doesn't work in high dimension.
If your representation is high dimensional,
there's just too many ways for two things to be different.
And so you would need lots and lots and lots
of negative pairs.
So there is a particular implementation of this,
which is relatively recent
from actually the Google Toronto group,
where Geoff Hinton is the senior member there.
And it's called SimClear, S-I-M-C-L-A-R.
And it, basically a particular way
of implementing this idea of contrastive learning,
the particular objective function.
Now, what I'm much more enthusiastic about these days
is non-contrastive methods.
So other ways to guarantee that
the representations would be different
for different inputs.
And it's actually based on an idea
that Geoff Hinton proposed in the early '90s
with his student at the time, Sue Becker.
And it's based on the idea
of maximizing the mutual information
between the outputs of the two systems.
You only show positive pairs,
you only show pairs of images
that you know are somewhat similar.
And you train the two networks to be informative,
but also to be as informative of each other as possible.
So basically one representation has to be predictable
from the other, essentially.
And he proposed that idea,
had a couple of papers in the early '90s,
§07Transcript · 52:00–1:00:42
and then nothing was done about it for decades.
And I kind of revived this idea
together with my postdocs at FAIR,
particularly a postdoc called Stefan Duny,
who's now a junior professor in Finland
at the University of Aalto.
We came up with something that we called Barlow Twins,
and it's a particular way
of maximizing the information content of a vector
using some hypotheses.
And we have kind of another version of it
that's more recent now called VICREG, V-I-C-R-E-G,
that means variance, invariance, covariance, regularization.
And it's the thing I'm the most excited about
in machine learning in the last 15 years.
I mean, I'm really, really excited about this.
- What kind of data augmentation is useful for that,
not contrast to the learning method?
Are we talking about, does that not matter that much?
Or it seems like a very important part of the step.
- Yeah.
- How you generate the images that are similar,
but sufficiently different.
- Yeah, that's right.
It's an important step, and it's also an annoying step,
because you need to have that knowledge
of what data augmentation you can do
that do not change the nature of the object.
And so the standard scenario,
which a lot of people working in this area are using,
is you use the type of distortion.
So basically you do a geometric distortion.
So one basically just shifts the image a little bit,
it's called cropping.
Another one kind of changes the scale a little bit.
Another one kind of rotates it.
Another one changes the colors.
You can do a shift in color balance
or something like that.
Saturation.
Another one sort of blurs it.
Another one adds noise.
So you have like a catalog of kind of standard things,
and people try to use the same ones for different algorithms
so that they can compare.
But some algorithms, some self-supervised algorithm
actually can deal with much bigger,
like more aggressive data augmentation, and some don't.
So that kind of makes the whole thing difficult.
But that's the kind of distortions we're talking about.
And so you train with those distortions,
and then you chop off the last layer,
a couple layers of the network,
and you use the representation as input to a classifier.
You train the classifier on ImageNet, let's say,
or whatever, and measure the performance.
And interestingly enough,
the methods that are really good
at eliminating the information that is irrelevant,
which is the distortions between those images,
do a good job at eliminating it.
And as a consequence,
you cannot use the representations in those systems
for things like object detection and localization,
because that information is gone.
So the type of data augmentation you need to do
depends on the task you want eventually the system to solve.
And the type of data augmentation,
standard data augmentation that we use today,
are only appropriate for object recognition
or image classification.
They're not appropriate for things like--
- Can you help me out understand why localization is--
So you're saying it's just not good at the negative,
like at classifying the negative,
so that's why it can't be used for the localization?
- No, it's just that you train the system,
you give it an image,
and then you give it the same image shifted and scaled,
and you tell it that's the same image.
So the system basically is trained
to eliminate the information about position and size.
So now you want to use that--
- Oh, like literally--
- Where an object is and what size it is.
- Like a bounding box, like to be able to actually, okay.
It can still find the object in the image,
it's just not very good at finding
the exact boundaries of that object, interesting.
Interesting, which, you know,
that's an interesting sort of philosophical question.
How important is object localization anyway?
We're like obsessed by measuring like image segmentation,
obsessed by measuring perfectly
knowing the boundaries of objects,
when arguably that's not that essential
to understanding what are the contents of the scene.
- On the other hand, I think evolutionarily,
the first vision systems in animals
were basically all about localization,
very little about recognition.
And in the human brain, you have two separate pathways
for recognizing the nature of a scene or an object,
and localizing objects.
So you use the first pathway called the ventral pathway
for telling what you're looking at.
The other pathway, the dorsal pathway,
is used for navigation, for grasping, for everything else.
And basically a lot of the things you need for survival
are localization and detection.
- Is similarity learning or contrastive learning,
are these non-contrastive methods
the same as understanding something?
Just because you know a distorted cat
is the same as a non-distorted cat,
does that mean you understand what it means to be a cat?
- To some extent.
I mean, it's a superficial understanding, obviously.
- But what is the ceiling of this method, do you think?
Is this just one trick on the path
to doing self-supervised learning,
or can we go really, really far?
- I think we can go really far.
So if we figure out how to use techniques of that type,
perhaps very different, but of the same nature,
to train a system from video,
to do video prediction, essentially,
I think we'll have a path towards,
I wouldn't say unlimited,
but a path towards some level of physical common sense
in machines.
And I also think that that ability to learn
how the world works from a high-throughput channel,
like vision, is a necessary step
towards real artificial intelligence.
In other words, I believe in grounded intelligence.
I don't think we can train a machine
to be intelligent purely from text,
because I think the amount of information
about the world that's contained in text
is tiny compared to what we need to know.
So for example, and people have attempted to do this
for 30 years, right, the PSYCH project,
and things like that, right,
of basically kind of writing down all the facts
that are known and hoping that some sort of common sense
will emerge, I think it's basically hopeless.
But let me take an example.
You take an object, I describe a situation to you.
I take an object, I put it on the table,
and I push the table.
It's completely obvious to you that the object
will be pushed with the table, right,
because it's sitting on it.
There's no text in the world, I believe,
that explains this.
And so if you train a machine as powerful as it could be,
you know, your GPT 5000, or whatever it is,
it's never gonna learn about this.
That information is just not present in any text.
- Well, the question, like with the PSYCH project,
the dream, I think, is to have like 10 million,
say, facts like that, that give you a head start,
like a parent guiding you.
Now, we humans don't need a parent to tell us
that the table will move, sorry,
the smartphone will move with the table.
But we get a lot of guidance in other ways,
so it's possible that we can give it a quick shortcut.
- What about a cat?
A cat knows that.
- No, but they evolved, so--
- No, they learn like us.
- Sorry, the physics of stuff?
- Yeah.
- Well, yeah, so you're saying it's,
so you're putting a lot of intelligence
onto the nurture side, not the nature.
- Yes.
- 'Cause we seem to have, you know,
there's a very inefficient, arguably,
process of evolution that got us from bacteria
to who we are today.
Started at the bottom, now we're here.
So the question is how, okay,
the question is how fundamental is that,
the nature of the whole hardware?
And then, is there any way to shortcut it
if it's fundamental?
If it's not, if it's most of the intelligence,
most of the cool stuff we've been talking about
is mostly nurture, mostly trained,
we figure it out by observing the world,
we can form that big, beautiful, sexy background model
that you're talking about just by sitting there.
Then, okay, then you need to, then like maybe,
it is all supervised learning all the way down.
It's all supervised learning, so.
Whatever it is that makes human intelligence
different from other animals,
§08Transcript · 1:00:44–1:09:36
which a lot of people think is language
and logical reasoning and this kind of stuff,
it cannot be that complicated
because it only popped up in the last million years.
- Yeah.
- And it only involves less than 1% of a genome,
which is the difference between human genome
and chimps or whatever.
So it can't be that complicated.
It can't be that fundamental.
I mean, most of the complicated stuff
already exists in cats and dogs
and certainly primates, non-human primates.
- Yeah, that little thing with humans
might be just something about social interaction
and ability to maintain ideas
across a collective of people.
It sounds very dramatic and very impressive,
but it probably isn't, mechanistically speaking.
- It is, but we're not there yet.
Like, we have, I mean, this is numbers.
- Number 634 in the list of problems we have to solve.
(laughs)
- So basic physics of the world is number one.
What do you, just a quick tangent on data augmentation.
So a lot of it is hard-coded versus learned.
Do you have any intuition that maybe
there could be some weird data augmentation,
like generative type of data augmentation,
like doing something weird to images,
which then improves the similarity learning process?
So not just kind of dumb, simple distortions,
but by you shaking your head,
just saying that even simple distortions are enough.
- I think, no, I think data augmentation
is a temporary, necessary evil.
So what people are working on now is two things.
One is the type of self-supervised learning,
like trying to translate the type of self-supervised learning
people use in language, translating these two images,
which is basically a denoising autoencoder method, right?
So you take an image, you block, you mask some parts of it,
and then you train some giant neural net
to reconstruct the parts that are missing.
And until very recently,
there was no working methods for that.
All the autoencoder type methods for images
weren't producing very good representation,
but there's a paper now coming out of the FAIR group
at MNLO Park that actually works very well.
So that doesn't require data augmentation,
that requires only masking.
Okay.
- Only masking for images.
Okay.
- Right, so you mask part of the image
and you train a system, which, you know,
in this case is a transformer because you can,
the transformer represents the image
as non-overlapping patches,
so it's easy to mask patches and things like that.
- Okay, then my question transfers to that problem,
the masking, like why should the mask be a square
or a rectangle?
- So it doesn't matter, like, you know,
I think we're gonna come up probably in the future
with sort of, you know, ways to mask that are, you know,
kind of random, essentially.
I mean, they are random already, but-
- No, no, but like something that's challenging,
like optimally challenging.
So like, I mean, maybe it's a metaphor that doesn't apply,
but you're, it seems like there's a data augmentation
or masking, there's an interactive element with it.
Like, you're almost like playing with an image.
- Yeah.
- And like, it's like the way we play
with an image in our minds.
- No, but it's like dropout.
It's like Boson machine training.
You know, every time you see a percept,
you also, you can perturb it in some way.
And then the principle of the training procedure
is to minimize the difference of the output
or the representation between the clean version
and the corrupted version, essentially, right?
And you can do this in real time, right?
So, you know, Boson machine work like this, right?
You show a percept, you tell the machine
that's a good combination of activities
or your input neurons.
And then you either let them go their merry way
without clamping them to values,
or you only do this with a subset.
And what you're doing is you're training the system
so that the stable state of the entire network
is the same regardless of whether it sees the entire input
or whether it sees only part of it.
You know, denoising autoencoder method
is basically the same thing, right?
You're training a system to reproduce the input,
the complete input and filling the blanks,
regardless of which parts are missing.
And that's really the underlying principle.
And you could imagine sort of even in the brain,
some sort of neural principle where, you know,
neurons kind of oscillate, right?
So they take their activity and then temporarily
they kind of shut off to, you know,
force the rest of the system to basically reconstruct
the input without their help, you know?
And I mean, you could imagine, you know,
more or less biologically possible processes.
- Something like that.
And I guess with this denoising autoencoder
and masking and data augmentation,
you don't have to worry about being super efficient.
You can just do as much as you want
and get better over time.
'Cause I was thinking like you might wanna be clever
about the way you do all of these procedures, you know,
but that's only, it's somehow costly to do every iteration,
but it's not really.
- Not really, maybe.
And then there is, you know,
data augmentation without explicit data augmentation.
Is data augmentation by weighting,
which is, you know, the sort of video prediction.
You're observing a video clip,
observing the, you know, the continuation of that video clip
and you try to learn a representation
using the joint embedding architectures
in such a way that the representation of the future clip
is easily predictable from the representation
of the observed clip.
- Do you think YouTube has enough raw data
from which to learn how to be a cat?
- I think so.
- So the amount of data is not the constraint.
- No, it would require some selection, I think.
- Some selection?
- Some selection of, you know,
maybe the right type of data.
- Don't go down the rabbit hole of just cat videos.
I might, you might need to watch some lectures or something.
- No.
- How meta would that be?
If it like watches lectures about intelligence
and then learns, watches your lectures on NYU
and learns from that how to be intelligent.
- I don't think there'd be enough.
- What's your, do you find multimodal learning interesting?
We've been talking about visual language,
like combining those together,
maybe audio, all those kinds of things.
- There's a lot of things that I find interesting
in the short term, but are not addressing
the important problem that I think are really
kind of the big challenges.
So I think, you know, things like multitask learning,
continual learning, you know, adversarial issues.
I mean, those have, you know, great practical interests
in the relatively short term possibly,
but I don't think they're fundamental, you know,
active learning, even to some extent reinforcement learning.
I think those things will become either obsolete
or useless or easy once we figured out
how to do self-supervised representation learning
or learning predictive world models.
And so I think that's what, you know,
the entire community should be focusing on.
At least people who are interested
in sort of fundamental questions or, you know,
really kind of pushing the envelope of AI
towards the next stage.
But of course there's like a huge amount of, you know,
very interesting work to do in sort of practical questions
that have, you know, short term impact.
- Well, you know, it's difficult to talk about
the temporal scale because all of human civilization
will eventually be destroyed because the sun will die out.
And even if Elon Musk is successful
in multi-planetary colonization across the galaxy,
eventually the entirety of it
will just become giant black holes.
And that's gonna keep the universe.
- That's gonna take a while, though.
- So, but what I'm saying is then that logic
can be used to say it's all meaningless.
I'm saying all that to say that multitask learning
might be, you're calling it practical or pragmatic
or whatever, that might be the thing
that achieves something very akin to intelligence
while we're trying to solve the more general problem
of self-supervised learning of background knowledge.
So the reason I bring that up,
maybe one way to ask that question.
I've been very impressed by what
Tesla Autopilot team is doing.
I don't know if you've gotten a chance to glance
§09Transcript · 1:09:38–1:18:43
at this particular one example of multitask learning
where they're literally taking the problem,
like, I don't know, Charles Darwin studying animals.
They're studying the problem of driving and asking,
okay, what are all the things you have to perceive?
And the way they're solving it is, one,
there's an ontology where you're bringing that to the table.
So you're formulating a bunch of different tasks.
It's like over 100 tasks or something like that
that they're involved in driving.
And then they're deploying it
and then getting data back from people that run into trouble
and they're trying to figure out, do we add tasks?
Do we, like, we focus on each individual task separately.
- Sure. - In fact, half,
so I would say, I'll classify Andrej Karpathy's talk
in two ways.
So one was about doors and the other one
about how much ImageNet sucks.
He kept going back and forth on those two topics,
which ImageNet sucks,
meaning you can't just use a single benchmark.
There's so, like, you have to have, like,
a giant suite of benchmarks to understand
how well your system actually works.
- Oh, I agree with him.
I mean, he's a very sensible guy.
Now, okay, it's very clear that if you're faced
with an engineering problem that you need to solve
in a relatively short time,
particularly if you have Elon Musk breathing down your neck,
you're going to have to take shortcuts, right?
You might think about the fact that the right thing to do
and the long-term solution involves, you know,
some fancy self-supervisioning,
but you have Elon Musk breathing down your neck,
and this involves human lives,
and so you have to basically just do
the systematic engineering and fine-tuning and refinements
and trial and error and all that stuff.
There's nothing wrong with that.
That's called engineering.
That's called putting technology out
in the world, and you have to kind of ironclad it
before you do this, you know,
so much for, you know, grand ideas and principles.
But, you know, I'm placing myself sort of, you know,
some, you know, upstream of this,
quite a bit upstream of this.
- Your Plato, think about platonic forms.
You're- - It's not platonic,
because eventually I want that stuff to get used,
but it's okay if it takes five or 10 years
for the community to realize this is the right thing to do.
I've done this before.
It's been the case before that, you know,
I've made that case.
I mean, if you look back in the mid-2000s, for example,
and you ask yourself the question,
okay, I want to recognize cars or faces or whatever,
you know, I can use convolutional nets,
so I can use sort of more conventional
kind of computer vision techniques, you know,
using interest point detectors or SIFT,
dense SIFT features and, you know,
sticking an SVM on top.
At that time, the datasets were so small
that those methods that use more hand engineering
worked better than conv nets.
There was just not enough data for conv nets,
and conv nets were a little slow
with the kind of hardware that was available at the time.
And there was a sea change when, basically when, you know,
datasets became bigger and GPUs became available.
That's what, you know, two of the main factors
that basically made people change their mind.
And you can look at the history of,
like all sub branches of AI or pattern recognition,
and there's a similar trajectory followed by techniques
where people start by, you know,
engineering the hell out of it.
You know, be it optical character recognition,
speech recognition, computer vision,
like image recognition in general,
natural language understanding, like, you know,
translation, things like that, right?
You start to engineer the hell out of it.
You start to acquire all the knowledge,
the prior knowledge you know about image formation,
about, you know, the shape of characters,
about, you know, morphological operations,
about like feature extraction, Fourier transforms,
you know, Wernicke moments, you know, whatever, right?
People have come up with thousands of ways
of representing images so that they could be
easily classified afterwards.
Same for speech recognition, right?
There is, you know, it took decades for people
to figure out a good front end to pre-process
a speech signal so that, you know,
the information about what is being said is preserved,
but most of the information about the identity
of the speaker is gone.
You know, Kestrel coefficients or whatever, right?
And same for text, right?
You do name entity recognition and you parse
and you do tagging of the parts of speech.
And, you know, you do this sort of tree representation
of clauses and all that stuff, right,
before you can do anything.
So that's how it starts, right?
Just engineer the hell out of it.
And then you start having data
and maybe you have more powerful computers,
maybe you know something about statistical learning.
So you start using machine learning
and it's usually a small sliver on top of your
kind of handcrafted system where, you know,
you extract features by hand.
Okay, and now, you know, nowadays,
the standard way of doing this is that
you train the entire thing end to end
with a deep learning system and it learns its own features
and, you know, speech recognition systems nowadays,
OCR systems, are completely end to end.
It's, you know, it's some giant neural net
that takes raw waveforms and produces a sequence
of characters coming out.
And it's just a huge neural net, right?
There's no, in a Markov model,
there's no language model that is explicit
other than, you know, something that's ingrained
in the sort of neural language model, if you want.
Same for translation, same for all kinds of stuff.
So you see this continuous evolution
from, you know, less and less handcrafting
and more and more learning.
And I think, I mean, it's true in biology as well.
- So, I mean, we might disagree about this.
Maybe not.
In this one little piece at the end,
you mentioned active learning.
It feels like active learning,
which is the selection of data and also the interactivity,
needs to be part of this giant neural network.
You cannot just be an observer
to do self-supervised learning.
You have to, well, I don't,
self-supervised learning is just a word,
but I would, whatever this giant stack
of a neural network that's automatically learning,
it feels, my intuition is that you have to have a system,
whether it's a physical robot or a digital robot
that's interacting with the world
and doing so in a flawed way and improving over time
in order to form the self-supervised learning well.
You can't just give it a giant sea of data.
- Okay, I agree and I disagree.
I agree in the sense that I think, I agree in two ways.
The first way I agree is that if you want,
and you certainly need a causal model of the world
that allows you to predict the consequences of your actions,
to train that model, you need to take actions.
You need to be able to act in a world and see the effect
for you to learn causal models of the world.
- So, that's not obvious because you can observe others.
- You can observe others.
- And you can infer that they're similar to you
and then you can learn from that.
- Yeah, but then you have to kind of hardware that part,
right, and then you don't mirror neurons
and all that stuff, right?
So, and it's not clear to me
how you would do this in a machine.
So, I think the action part would be necessary
for having causal models of the world.
The second reason it may be necessary
or at least more efficient is that active learning
basically goes for the jiggler of what you don't know, right?
There's obvious areas of uncertainty about your world
and about how the world behaves.
And you can resolve this uncertainty
by systematic exploration of that part that you don't know.
And if you know that you don't know,
then it makes you curious.
You kind of look into situations that...
And across the animal world,
different species are different levels of curiosity, right?
Depending on how they're built, right?
So, cats and rats are incredibly curious,
dogs not so much, I mean, less.
- Yeah, so it could be useful
to have that kind of curiosity.
- So, it'd be useful,
but curiosity just makes the process faster.
It doesn't make the process exist.
So, what process, what learning process is it
that active learning makes more efficient?
And I'm asking that first question.
We haven't answered that question yet.
§10Transcript · 1:18:44–1:27:44
So, I worry about active learning once this question is...
- So, it's the more fundamental question to ask.
And if active learning or interaction
increases the efficiency of the learning,
see, sometimes it becomes very different
if the increase is several orders of magnitude, right?
- That's true.
- But fundamentally, it's still the same thing
in building up the intuition about how to,
in a self-supervised way,
to construct background models,
efficient or inefficient, is the core problem.
What do you think about Yoshua Bengio's
talking about consciousness
and all of these kinds of concepts?
- Okay, I don't know what consciousness is, but...
- It's a good opener.
- And to some extent, a lot of the things
that are said about consciousness remind me
of the questions people were asking themselves
in the 18th century or 17th century
when they discovered that, you know,
how the eye works and the fact that the image
at the back of the eye was upside down, right?
Because you have a lens.
And so, on your retina, the image that forms
is an image of the world, but it's upside down.
How is it that you see right side up?
And, you know, with what we know today in science,
you know, we realize this question doesn't make any sense
or is kind of ridiculous in some way, right?
So, I think a lot of what is said about consciousness
is of that nature.
Now, that said, there's a lot of really smart people
that, for whom I have a lot of respect,
who are talking about this topic,
people like David Chalmers, who is a colleague of mine
at NYU.
I have kind of an unorthodox, folk,
speculative hypothesis about consciousness.
So, we're talking about the study of a world model.
And I think, you know, our entire prefrontal cortex
basically is the engine for our world model.
But when we are attending at a particular situation,
we're focused on that situation.
We basically cannot attend to anything else.
And that seems to suggest that we basically have only one
world model engine in our prefrontal cortex.
That engine is configurable to the situation at hand.
So, we are building a box out of wood,
or we are, you know, driving down the highway,
playing chess.
We basically have a single model of the world
that we're configuring to the situation at hand,
which is why we can only attend to one task at a time.
Now, if there is a task that we do repeatedly,
it goes from the sort of deliberate reasoning
using model of the world and prediction,
and perhaps something like model predictive control,
which I was talking about earlier,
to something that is more subconscious
that becomes automatic.
So, I don't know if you've ever played against
a chess grandmaster.
You know, I get wiped out in 10 plies, right?
And, you know, I have to think about my move
for, you know, like 15 minutes.
And the person in front of me, the grandmaster,
you know, would just like react within seconds, right?
You know, he doesn't need to think about it.
That's become part of the subconscious,
because, you know, it's basically just pattern recognition
at this point.
Same, you know, the first few hours you drive a car,
you're really attentive.
You can't do anything else.
And then after 20, 30 hours of practice, 50 hours,
you know, it's subconscious.
You can talk to the person next to you,
you know, things like that, right?
Unless the situation becomes unpredictable,
and then you have to stop talking.
So, that suggests you only have one model in your head.
And it might suggest the idea that consciousness
basically is the module that configures
this world model of yours.
You know, you need to have some sort of executive
kind of overseer that configures your world model
for the situation at hand.
And that leads to kind of the really curious concept
that consciousness is not a consequence of the power
of our minds, but of the limitation of our brains.
But because we have only one world model,
we have to be conscious.
If we had as many world models as there are situations
we encounter, then we could do all of them simultaneously,
and we wouldn't need this sort of executive control
that we call consciousness.
- Yeah, interesting.
And somehow maybe that executive controller,
I mean, the hard problem of consciousness,
there's some kind of chemicals in biology
that's creating a feeling, like it feels
to experience some of these things.
That's kind of like the hard question is,
what the heck is that, and why is that useful?
Maybe the more pragmatic question,
why is it useful to feel like this is really you
experiencing this versus just like information
being processed?
- It could be just a very nice side effect
of the way we evolved that it's just very useful
to feel a sense of ownership to the decisions you make,
to the perceptions you make, to the model
you're trying to maintain.
Like you own this thing, and it's the only one you got,
and if you lose it, it's gonna really suck.
And so you should really send the brain
some signals about it.
- What ideas do you believe might be true
that most or at least many people disagree with you with,
let's say in the space of machine learning?
- Well, it depends who you talk about.
But I think, so certainly there is a bunch of people
who are nativist, right, who think that a lot
of the basic things about the world are kind of hardwired
in our minds.
Things like the world is three-dimensional, for example.
Is that hardwired?
Things like object permanence, is it something
that we learn before the age of three months or so,
or are we born with it?
And there are wide disagreements among
the cognitive scientists for this.
I think those things are actually very simple to learn.
Is it the case that the oriented edge detectors in V1
are learned, or are they hardwired?
I think they are learned.
They might be learned before birth,
because it's really easy to generate signals
from the retina that actually will train edge detectors.
So, and again, those are things that can be learned
within minutes of opening your eyes, right?
I mean, since the 1990s, we have algorithms
that can learn oriented edge detectors
completely unsupervised with the equivalent
of a few minutes of real time.
So those things have to be learned.
There's also those MIT experiments where you kind of plug
the optical nerve on the auditory cortex of a baby ferret,
right, and that auditory cortex
becomes a visual cortex, essentially.
So, you know, clearly there's learning taking place there.
So, you know, I think a lot of what people think
are so basic that they need to be hardwired,
I think a lot of those things are learned
because they are easy to learn.
- So you put a lot of value in the power of learning.
What kind of things do you suspect might not be learned?
Is there something that could not be learned?
- So your intrinsic drives are not learned.
There are the things that, you know, make humans human
or make, you know, cats different from dogs, right?
It's the basic drives that are kind of hardwired
in our basal ganglia.
I mean, there are people who are working
on this kind of stuff.
It's called intrinsic motivation
in the context of reinforcement learning.
So these are objective functions
where the reward doesn't come from the external world.
It's computed by your own brain.
Your own brain computes whether you're happy or not, right?
It measures your degree of comfort or incomfort.
And because it's your brain computing this,
presumably it knows also how to estimate
gradients of this, right?
So it's easier to learn when your objective is intrinsic.
So that has to be hardwired.
The critic that makes long-term prediction of the outcome,
which is the eventual result of this, that's learned.
And perception is learned,
and your model of the world is learned.
But let me take an example of, you know, why the critic,
I mean, an example of how the critic may be learned, right?
If I come to you, you know, I reach across the table
and I pinch your arm, right?
Complete surprise for you.
You would not have expected this from me.
- I was expecting that the whole time, but yes, right.
Let's say for the sake of the story, yes.
- Okay, your basal ganglia is gonna light up
'cause it's gonna hurt, right?
And now your model of the world includes the fact that
I may pinch you if I approach my-
- Don't trust humans.
- Right, my hand to your arm.
So if I try again, you're gonna recoil.
And that's your critic, your predictive,
you know, your predictor of your ultimate pain system
§11Transcript · 1:27:50–1:36:58
that predicts that something bad is gonna happen
and you recoil to avoid it.
- So even that can be learned.
- That is learned, definitely.
This is what allows you also to, you know,
define sub goals, right?
So the fact that, you know, you're a school child,
you wake up in the morning and you go to school and,
you know, it's not because you necessarily like waking up
early and going to school,
but you know that there is a long-term objective
you're trying to optimize.
- So Ernest Becker, I'm not sure if you're familiar
with him, the philosopher, he wrote the book
"Denial of Death" and his idea is that
one of the core motivations of human beings
are terror of death, are fear of death.
That's what makes us unique from cats.
Cats are just surviving.
They do not have a deep, like a cognizance,
introspection that over the horizon is the end.
And he says that, I mean,
there's a terror management theory
that just all these psychological experiments that show,
basically this idea that all of human civilization,
everything we create is kind of trying to forget
if even for a brief moment that we're going to die.
When do you think humans understand
that they're going to die?
Is it learned early on also?
Like?
- I don't know at what point, I mean, it's a question,
like, you know, at what point do you realize that,
you know, what death really is?
And I think most people don't actually realize
what death is, right?
I mean, most people believe that you go to heaven
or something, right?
- So to push back on that, what Ernest Becker says
and Sheldon Solomon, all of those folks,
and I find those ideas a little bit compelling
is that there is moments in life, early in life,
a lot of this fun happens early in life
when you are, when you do deeply experience the terror
of this realization and all the things you think about,
about religion, all those kinds of things
that we kind of think about more like teenage years
and later, we're talking about way earlier.
- No, it was like seven or eight years,
something like that, yeah.
- You realize, holy crap, this is,
like the mystery, the terror, like,
it's almost like you're a little prey,
a little baby deer sitting in the darkness
of the jungle, the woods, looking all around you,
the darkness full of terror.
I mean, that realization says, okay,
I'm going to go back in the comfort of my mind
where there is a deep meaning,
where there is a, maybe like, pretend I'm immortal
in however way, however kind of idea I can construct
to help me understand that I'm immortal.
Religion helps with that.
You can delude yourself in all kinds of ways,
like lose yourself in the busyness of each day,
have little goals in mind, all those kinds of things
to think that it's gonna go on forever,
and you kind of know you're gonna die, yeah,
and it's gonna be sad, but you don't really understand
that you're going to die.
And so that's their idea.
And I find that compelling because it does seem
to be a core unique aspect of human nature
that we're able to think that we're going,
we're able to really understand that this life is finite.
That seems important.
- There's a bunch of different things there.
So first of all, I don't think there is
a qualitative difference between us and cats in the term.
I think the difference is that we just have
a better long-term ability to predict in the long term,
and so we have a better understanding of how the world works,
so we have a better understanding of finiteness of life
and things like that.
- So we have a better planning engine than cats?
- Yeah.
- Okay.
- But what's the motivation for planning that far?
- Well, I think it's just a side effect
of the fact that we have just a better planning engine
because it makes us, as I said,
the essence of intelligence is the ability to predict.
And so because we're smarter, as a side effect,
we also have this ability to kind of make predictions
about our own future existence or lack thereof.
- Okay.
- You say religion helps with that.
I think religion hurts, actually.
It makes people worry about what's gonna happen
after their death, et cetera.
If you believe that you just don't exist after death,
it solves completely the problem, at least.
- You're saying if you don't believe in God,
you don't worry about what happens after death?
- Yeah.
- I don't know.
- You only worry about this life
because that's the only one you have.
- I think it's, well, I don't know.
If I were to say what Ernest Becker says,
and I would say I agree with him more than not,
is you do deeply worry.
If you believe there's no God,
there's still a deep worry of the mystery of it all.
Like, how does that make any sense that it just ends?
I don't think we can truly understand that this right,
I mean, so much of our life, the consciousness, the ego,
is invested in this being.
- Science keeps bringing humanity down from its pedestal.
And that's just another example of it.
- That's wonderful, but for us individual humans,
we don't like to be brought down from a pedestal.
You're saying like-- - I'm fine with it.
- But see, you're fine with it because, well,
so what Ernest Becker would say is you're fine with it
because that's just a more peaceful existence for you,
but you're not really fine.
You're hiding from, in fact, some of the people
that experience the deepest trauma earlier in life,
they often, before they seek extensive therapy,
will say that I'm fine.
It's like when you talk to people who are truly angry,
how are you doing, I'm fine.
The question is what's going on.
- Now, I had a near-death experience.
I had a very bad motorbike accident when I was 17.
So, but that didn't have any impact
on my reflection on that topic.
- So I'm basically just playing a bit of a devil's advocate,
pushing back on wondering is it truly possible
to accept death?
And the flip side that's more interesting, I think,
for AI and robotics is how important is it to have this
as one of the suite of motivations,
is to not just avoid falling off the roof or something
like that, but ponder the end of the ride.
If you listen to the Stoics, it's a great motivator.
It adds a sense of urgency.
So maybe to truly fear death or be cognizant of it
might give a deeper meaning and urgency
to the moment, to live fully.
- Maybe I don't disagree with that.
I mean, I think what motivates me here is,
you know, knowing more about human nature.
I mean, I think human nature and human intelligence
is a big mystery.
It's a scientific mystery, in addition to, you know,
philosophical and et cetera,
but, you know, I'm a true believer in science.
So, and I do have kind of a belief that for complex systems
like the brain and the mind, the way to understand it
is to try to reproduce it with, you know,
artifacts that you build, because you know
what's essential to it when you try to build it.
You know, the same way I've used this analogy before
with you, I believe, the same way we only started
to understand aerodynamics when we started
building airplanes, and that helped us understand
how birds fly.
So I think there's kind of a similar process here
where we don't have a theory, a full theory of intelligence,
but building, you know, intelligent artifacts
will help us perhaps develop some, you know,
underlying theory that encompasses not just
artificial implements, but also human
and biological intelligence in general.
- So you're an interesting person to ask this question
about sort of all kinds of different other
intelligent entities or intelligences.
What are your thoughts about kind of like the Turing
or the Chinese room question?
If we create an AI system that exhibits a lot of properties
of intelligence and consciousness,
how comfortable are you thinking of that entity
as intelligent or conscious?
So you're trying to build now systems that have intelligence
and there's metrics about their performance,
but that metric is external.
- Okay.
- So how are you, are you okay calling a thing intelligent
or are you going to be like most humans
and be once again unhappy to be brought down
from a pedestal of consciousness/intelligence?
- No, I'll be very happy to understand
more about human nature, human mind, and human intelligence
through the construction of machines
that have similar abilities.
And if a consequence of this is to bring down humanity
one notch down from its already low pedestal,
I'm just fine with it.
§12Transcript · 1:36:59–1:45:14
That's just the reality of life.
So I'm fine with that.
Now, you were asking me about things that,
opinions I have that a lot of people may disagree with.
I think if we think about the design
of an autonomous intelligence system,
so assuming that we are somewhat successful at some level
of getting machines to learn models of the world,
predictive models of the world,
we build intrinsic motivation objective functions
to drive the behavior of that system.
The system also has perception modules
that allows it to estimate the state of the world
and then have some way of figuring out a sequence of actions
that, you know, to optimize a particular objective.
If it has a critic of the type that I was describing before,
the thing that makes you recoil your arm
the second time I try to pinch you,
intelligent autonomous machine will have emotions.
I think emotions are an integral part
of autonomous intelligence.
If you have an intelligent system
that is driven by intrinsic motivation, by objectives,
if it has a critic that allows it to predict in advance
whether the outcome of a situation
is going to be good or bad, it's going to have emotions.
It's going to have fear.
- Yes.
- When it predicts that the outcome is going to be bad
and something to avoid, it's going to have elation
when it predicts it's going to be good.
If it has drives to relate with humans,
you know, in some ways, the way humans have,
you know, it's going to be social, right?
And so it's going to have emotions about attachment
and things of that type.
So I think, you know, the sort of sci-fi thing
where, you know, you see commander data
like having an emotion chip that you can turn off, right?
I think that's ridiculous.
- So, I mean, here's the difficult
philosophical social question.
Do you think there will be a time,
like a civil rights movement for robots where,
okay, forget the movement, but a discussion,
like the Supreme Court, that particular kinds of robots,
you know, particular kinds of systems
deserve the same rights as humans
because they can suffer just as humans can,
all those kinds of things?
- Well, perhaps, perhaps not.
Like imagine that humans were,
that you could, you know, die and be restored.
Like, you know, you could be sort of, you know,
be 3D reprinted and, you know,
your brain could be reconstructed in its finest details.
Our ideas of rights will change in that case.
If you can always just,
there's always a backup, you could always restore.
Maybe like the importance of murder
will go down one notch.
- That's right.
But also the, your, you know,
desire to do dangerous things like, you know,
doing skydiving or, you know,
or, you know, race car driving, you know,
car racing or that kind of stuff, you know,
would probably increase.
Or, you know, airplane aerobatics
or that kind of stuff, right?
It would be fine to do a lot of those things
or explore, you know, dangerous areas and things like that.
It would kind of change your relationship.
So now it's very likely that robots would be like that
because, you know, they'll be based on perhaps technology
that is somewhat similar to today's technology.
And you can always have a backup.
- So it's possible.
I don't know if you like video games,
but there's a game called Diablo and-
- Oh, my sons are huge fans of this.
- Yes.
- And in fact, they made a game that's inspired by it.
- Awesome.
Like built a game?
- My three sons have a game design studio between them.
Yeah. - That's awesome.
- They came out with a game.
- Like it just came out last year?
- No, this was last year, early last year,
about a year ago.
- That's awesome.
But so in Diablo, there's something called hardcore mode,
which if you die, there's no, you're gone.
- Right.
- That's it.
So it's possible with AI systems,
for them to be able to operate successfully
and for us to treat them in a certain way,
'cause they have to be integrated in human society,
they have to be able to die, no copies allowed.
In fact, copying is illegal.
It's possible with humans as well,
like cloning will be illegal, even when it's possible.
- But cloning is not copying, right?
I mean, you don't reproduce the mind of the person
and experience.
- Right.
- It's just a delayed twin, so.
- But then it's, but we were talking about with computers
that you will be able to copy.
- Right.
- You will be able to perfectly save,
pickle the mind state.
And it's possible that that will be illegal
because that goes against,
that will destroy the motivation of the system.
- Okay, so let's say you have a domestic robot.
- Yes.
- Okay, sometime in the future.
- Yes.
- And the domestic robot, you know,
comes to you kind of somewhat pre-trained,
you know, it can do a bunch of things.
- Yes.
- But it has a particular personality
that makes it slightly different from the other robots
because that makes them more interesting.
And then because it's, you know,
it's lived with you for five years,
you've grown some attachment to it and vice versa.
And it's learned a lot about you.
Or maybe it's not a household robot.
Maybe it's a virtual assistant that lives in your,
you know, augmented reality glasses or whatever, right?
You know, the horror movie type thing, right?
And that system, to some extent,
the intelligence in that system is a bit like your child
or maybe your PhD student in the sense that
there's a lot of you in that machine now, right?
- Yeah.
- And so if it were a living thing,
you would do this for free if you want, right?
If it's your child, your child can, you know,
then live his or her own life.
And, you know, the fact that they learn stuff from you
doesn't mean that you have any ownership of it, right?
- Yeah.
- But if it's a robot that you've trained,
perhaps you have some intellectual property claim about-
- Intellectual property?
Oh, I thought you meant like a permanence value
in the sense that's part of you is in-
- Well, there is permanence value, right?
So you would lose a lot if that robot were to be destroyed
and you had no backup, you would lose a lot.
You would lose a lot of investment, you know,
kind of like a person dying, you know,
that a friend of yours dying or a coworker
or something like that.
- But also you have like intellectual property rights
in the sense that that system is fine-tuned
to your particular existence.
So that's now a very unique instantiation
of that original background model,
whatever it was that arrived.
- And then there are issues of privacy, right?
Because now imagine that that robot has its own
kind of volition and decides to work for someone else
or kind of, you know, thinks life with you
is sort of untenable or whatever.
- Right.
- Now, all the things that that system learned from you,
you know, can you like, you know,
delete all the personal information
that that system knows about you?
- Yeah.
- I mean, that would be kind of an ethical question.
Like, you know, can you erase the mind of a intelligent
robot to protect your privacy?
- Yeah.
- You can't do this with humans.
You can ask them to shut up,
but that you don't have complete power over them.
- Can't erase humans.
Yeah, it's the problem with the relationships, you know,
that you break up, you can't erase the other human.
With robots, I think it will have to be the same thing
with robots, that risk, that there has to be some risk
to our interactions to truly experience them deeply,
it feels like.
So you have to be able to lose your robot friend
and that robot friend to go tweeting
about how much of an asshole you are.
- But then are you allowed to, you know,
murder the robot to protect your private information?
- Yeah, probably not.
- If the robot decides to leave?
- I have this intuition that for robots with certain,
like, it's almost like a regulation.
§13Transcript · 1:45:16–1:54:38
If you declare your robot to be,
let's call it sentient or something like that,
like this robot is designed for human interaction,
then you're not allowed to murder these robots.
It's the same as murdering other humans.
- Well, but what about you do a backup of the robot
that you preserve on a hard drive
with the equivalent in the future?
- That might be illegal.
It's like piracy is illegal.
- No, but it's your own robot, right?
- But you can't, you don't--
- But then you can wipe out his brain,
so this robot doesn't know anything about you anymore,
but you still have, technically it's still in existence
because you backed it up.
- And then there'll be these great speeches
at the Supreme Court by saying,
oh, sure, you can erase the mind of the robot
just like you can erase the mind of a human.
We both can suffer.
There'll be some epic, like,
Obama-type character with a speech that we,
like, the robots and the humans are the same.
We can both suffer, we can both hope,
we can both, all of those kinds of things,
raise families, all that kind of stuff.
It's interesting for these, just like you said,
emotion seems to be a fascinatingly powerful aspect
of human-to-human interaction, human-robot interaction,
and if they're able to exhibit emotions,
at the end of the day, that's probably going to
have us deeply consider human rights,
like what we value in humans,
what we value in other animals.
That's why robots and AI is great.
It makes us ask really good questions.
- The hard questions, yeah.
You asked about the Chinese room-type argument.
Is it real, if it looks real?
I think the Chinese room argument is a ridiculous one.
- So for people who don't know, Chinese room is,
I don't even know how to formulate it well,
but basically, you can mimic the behavior
of an intelligent system by just following
a giant algorithm codebook that tells you
exactly how to respond in exactly each case,
but is that really intelligent?
It's like a giant lookup table.
When this person says this, you answer this.
When this person says this, you answer this.
And if you understand how that works,
you have this giant, nearly infinite lookup table.
Is that really intelligence?
'Cause intelligence seems to be a mechanism
that's much more interesting and complex
than this lookup table.
- I don't think so.
So the real question comes down to,
do you think you can mechanize intelligence in some way,
even if that involves learning?
And the answer is, of course, yes.
There's no question.
There's a second question then,
which is, assuming you can reproduce intelligence
in sort of different hardware than biological hardware,
you know, like computers,
can you match human intelligence in all the domains
in which humans are intelligent?
Is it possible, right?
So that's the hypothesis of strong AI.
The answer to this, in my opinion, is an unqualified yes.
This will happen at some point.
There's no question that machines, at some point,
will become more intelligent than humans
in all domains where humans are intelligent.
This is not for tomorrow.
It's gonna take a long time,
regardless of what Elon and others have claimed or believed.
This is a lot harder than many of those guys think it is.
And many of those guys who thought it was simpler than that
years, you know, five years ago,
now think it's hard because it's been five years
and they realize it's gonna take a lot longer.
That includes a bunch of people at DeepMind, for example.
- Oh, interesting.
I haven't actually touched base with the DeepMind folks,
but some of it, Elon or Demis Hassabis,
I mean, sometimes in your role,
you have to kind of create deadlines
that are nearer than farther away
to kind of create an urgency.
'Cause you have to believe the impossible is possible
in order to accomplish it.
And there's, of course, a flip side to that coin,
but it's a weird, you can't be too cynical
if you wanna get something done.
- Absolutely, I agree with that.
But I mean, you have to inspire people, right,
to work on sort of ambitious things.
So, you know, it's certainly a lot harder than we believe,
but there's no question in my mind that this will happen.
And now, you know, people are kind of worried about
what does that mean for humans?
They are gonna be brought down from their pedestal,
you know, a bunch of notches with that.
And, you know, is that gonna be good or bad?
I mean, it's just gonna give more power, right?
It's an amplifier for human intelligence, really.
- So speaking of doing cool, ambitious things,
FAIR, the Facebook AI Research Group,
has recently celebrated its eighth birthday,
or maybe you can correct me on that.
Looking back, what has been the successes,
the failures, the lessons learned
from the eight years of FAIR?
And maybe you can also give context of
where does the newly minted meta AI fit into,
how does it relate to FAIR?
- Right, so let me tell you a little bit
about the organization of all this.
Yeah, FAIR was created almost exactly eight years ago.
It wasn't called FAIR yet.
It took that name a few months later.
And at the time I joined Facebook,
there was a group called the AI Group
that had about 12 engineers and a few scientists,
like 10 engineers and two scientists
or something like that.
I ran it for three and a half years as a director,
hired the first few scientists
and kind of set up the culture and organized it,
explained to the Facebook leadership
what fundamental research was about
and how it can work within industry
and how it needs to be open and everything.
And I think it's been an unqualified success
in the sense that FAIR has simultaneously produced
top-level research and advanced the science
and the technology, provided tools,
open source tools like PyTorch and many others.
But at the same time has had a direct
or mostly indirect impact on Facebook at the time,
now meta, in the sense that a lot of systems
that meta is built around now are based on research projects
that started at FAIR.
And so if you were to take out deep learning
out of Facebook services now and meta more generally,
I mean, the company would literally crumble.
I mean, it's completely built around AI these days
and it's really essential to the operations.
So what happened after three and a half years
is that I changed role, I became chief scientist.
So I'm not doing day-to-day management of FAIR anymore.
I'm more of a kind of, you know,
think about strategy and things like that.
And I carry my, I conduct my own research.
I have, you know, my own kind of research group
working on self-supervised learning and things like this,
which I didn't have time to do when I was director.
So now FAIR is run by Joël Pinot and Antoine Borde.
Together, because FAIR is kind of split in two now,
there's something called FAIR Labs,
which is sort of bottom-up, science-driven research
and FAIR Excel, which is slightly more organized
for bigger projects that require a little more kind of focus
and more engineering support and things like that.
So Joël needs FAIR Lab and Antoine Borde needs FAIR Excel.
- Where are they located?
- It's delocalized all over.
So there's no question that the leadership of the company
believes that this was a very worthwhile investment.
And what that means is that it's there for the long run.
Right, so there is, if you want to talk in these terms,
which I don't like, there's a business model, if you want,
where FAIR, despite being a very fundamental research lab,
brings a lot of value to the company,
mostly indirectly through other groups.
Now, what happened three and a half years ago
when I stepped down was also the creation of Facebook AI,
which was basically a larger organization
that covers FAIR, so FAIR is included in it,
but also has other organizations that are focused
on applied research or advanced development of AI technology
that is more focused on the products of the company.
- So less emphasis on fundamental research.
- Less fundamental, but it's still research.
I mean, there's a lot of papers coming out
of those organizations and people are awesome
and wonderful to interact with,
but it serves as kind of a way to kind of scale up,
if you want, sort of AI technology,
which may be very experimental and sort of lab prototypes
into things that are usable.
- So FAIR is a subset of meta AI.
Is FAIR become like KFC?
It'll just keep the F, nobody cares what the F stands for?
- We'll know soon enough, probably by the end of 2021.
- I guess it's not a giant change, Mare, FAIR.
- Well, Mare doesn't sound too good,
§14Transcript · 1:54:39–2:03:23
but the brand people are kind of deciding on this
and they've been hesitating for a while now
and they tell us they're gonna come up with an answer
as to whether FAIR is gonna change name
or whether we're gonna change just the meaning of the F.
- Oh, that's a good call.
I would keep FAIR and change the meaning of the F.
- That would be my preference.
- I would turn the F into fundamental.
- Oh, that's good. - Fundamental AI research.
- Oh, that's really good, yeah, yeah.
- Within meta AI.
So this would be sort of meta FAIR,
but people would call it FAIR, right?
- Yeah, exactly.
I like it.
- And now meta AI is part of the reality lab.
So, you know, meta now, the new Facebook, right,
it's called meta and it's kind of divided
into, you know, Facebook, Instagram, WhatsApp,
and reality lab.
And reality lab is about, you know, AR, VR,
you know, telepresence, communication,
technology and stuff like that.
It's kind of the, you can think of it as the sort of,
a combination of sort of new products
and technology part of meta.
- Is that where the touch sensing for robots,
I saw that you were posting about, that's-
- But touch sensing for robot is part of FAIR, actually.
That's a FAIR product. - Oh, it is, okay, cool.
- This is also the, no, but there is the other way,
the haptic glove, right? - Yes.
- That has like- - That's more reality lab.
- That's reality lab research.
- Reality lab research.
But by the way, the touch sensors is super interesting.
Like integrating that modality
into the whole sensing suite is very interesting.
So what do you think about the metaverse?
What do you think about this whole kind of expansion
of the view of the role of Facebook and meta in the world?
- Well, metaverse really should be thought of
as the next step in the internet, right?
Sort of trying to kind of, you know,
make the experience more compelling of, you know,
being connected either with other people or with content.
And, you know, we are evolved and trained to evolve in,
you know, 3D environments where, you know,
we can see other people, we can talk to them when near them,
or, you know, and other people are far away can't hear us,
you know, things like that, right?
So there's a lot of social conventions that exist
in the real world that we can try to transpose.
Now, what is gonna be eventually the,
how compelling is it gonna be?
Like, you know, is it gonna be the case
that people are gonna be willing to do this
if they have to wear, you know,
a huge pair of goggles all day?
Maybe not.
- But then again, if the experience
is sufficiently compelling, maybe so.
- Or if the device that you have to wear
is just basically a pair of glasses, you know,
and technology makes sufficient progress for that.
You know, AR is a much easier concept to grasp
that you're gonna have, you know,
augmented reality glasses that basically contain
some sort of, you know, virtual assistant
that can help you in your daily lives.
- But at the same time with the AR,
you have to contend with reality.
With VR, you can completely detach yourself from reality,
so it gives you freedom.
It might be easier to design worlds in VR.
- Yeah, but you can imagine, you know,
the metaverse being-
- A mix.
- A mix, right.
Or like you can have objects that exist in the metaverse
that, you know, pop up on top of the real world
or only exist in virtual reality.
- Okay, let me ask the hard question.
- Oh, because all of this was easy.
- This was easy.
The Facebook, now Meta, the social network
has been painted by the media as net negative for society,
even destructive and evil at times.
You've pushed back against this, defending Facebook.
Can you explain your defense?
- Yeah, so the description,
the company that is being described in some media
is not the company we know when we work inside.
And, you know, it could be claimed that
a lot of employees are uninformed
about what really goes on in a company,
but, you know, I'm a vice president.
I mean, I have a pretty good vision of what goes on.
You know, I don't know everything, obviously.
I'm not involved in everything,
but certainly not in decision about like, you know,
content moderation or anything like this,
but I have some decent vision of what goes on.
And this evil that is being described, I just don't see it.
And then, you know, I think there is an easy story to buy,
which is that, you know, all the bad things in the world,
and, you know, the reason your friend believe crazy stuff,
you know, there's an easy scapegoat, right,
in social media in general, Facebook in particular.
We have to look at the data.
Like, is it the case that Facebook, for example,
polarizes people politically?
Are there academic studies that show this?
Is it the case that, you know, teenagers
think of themselves less if they use Instagram more?
Is it the case that, you know,
people get more riled up against, you know,
opposite sides in a debate or political opinion
if they are more on Facebook or if they are less?
And study after study show that none of this is true.
This is independent studies by academic.
They're not funded by Facebook or Meta, you know,
studied by Stanford, by some of my colleagues at NYU,
actually, with whom I have no connection.
You know, there's a study recently, they paid people,
I think it was in the former Yugoslavia.
I'm not exactly sure in what part,
but they paid people to not use Facebook for a while
in the period before the anniversary of the Srebrenica
massacres, right?
So, you know, people get riled up, like, should, you know,
should we have a celebration?
I mean, a memorial kind of celebration for it or not.
So they paid a bunch of people to not use Facebook
for a few weeks.
And it turns out that those people ended up being
more polarized than they were at the beginning.
And the people who were more on Facebook
were less polarized.
There's a study, you know, from Stanford of,
economists at Stanford that try to identify the causes
of increasing polarization in the US.
And it's been going on for 40 years before, you know,
Mark Zuckerberg was born.
- Yeah.
- Continuously.
And so if there is a cause,
it's not Facebook or social media.
So you could say if social media just accelerated,
but no, I mean, it's basically a continuous evolution
by some measure of polarization in the US.
And then you compare this with other countries
like the West half of Germany,
because you can go 40 years in the East side
or Denmark or other countries.
And they use Facebook just as much.
And they're not getting more polarized,
they're getting less polarized.
So if you want to look for, you know,
a causal relationship there,
you can find a scapegoat, but you can't find a cause.
Now, if you want to fix the problem,
you have to find the right cause.
And what riles me up is that people now are,
people now are accusing Facebook of bad deeds
that are done by others.
And those others are, we're not doing anything about them.
And by the way, those others include
the owner of the Wall Street Journal
in which all of those papers were published.
- So I should mention that I'm talking to Shrepp,
Mike Shrepp on this podcast and also Mark Zuckerberg,
and probably these are conversations you can have with them.
'Cause it's very interesting to me,
even if Facebook has some measurable negative effect,
you can't just consider that in isolation.
You have to consider about all the positive ways
that it connects us.
So like every technology--
- It connects people, it's a question.
- You can't just say like, there's an increase in division.
Yes, probably Google search engine
has created increase in division,
but you have to consider about how much information
are brought to the world.
Like I'm sure Wikipedia created more division
if you just look at the division,
but you have to look at the full context of the world
and did it make a better world?
- I mean, the printing press has created more division.
- Exactly.
- So when the printing press was invented,
the first books that were printed
were things like the Bible,
and that allowed people to read the Bible by themselves,
not get the message uniquely from priests in Europe.
And that created the Protestant movement
and 200 years of religious persecution and wars.
So that's a bad side effect of the printing press.
§15Transcript · 2:03:26–2:11:53
Social networks aren't being nearly as bad
as the printing press,
but nobody would say the printing press was a bad idea.
- Yeah, a lot of it is perception
and there's a lot of different incentives operating here.
Maybe a quick comment,
since you're one of the top leaders at Facebook
and at Meta, sorry, that's in the tech space,
I'm sure Facebook involves
a lot of incredible technological challenges
that need to be solved.
A lot of it probably is on the computer infrastructure,
the hardware, I mean, it's just a huge amount.
Maybe can you give me context
about how much of Schrepp's life is AI
and how much of it is low-level compute?
How much of it is flying all around doing business stuff
in the same way Zuckerberg, Mark Zuckerberg?
- They really focus on AI.
I mean, certainly in the run-up of the creation affair
and for at least a year after that, if not more,
Mark was very, very much focused on AI
and was spending quite a lot of effort on it.
And that's his style.
When he gets interested in something,
he reads everything about it.
He read some of my papers, for example, before I joined.
And so he learned a lot about it.
- He said he liked notes.
- Right.
And Schrepp was really into it also.
I mean, Schrepp is really kind of,
has something I've tried to preserve also
despite my not so young age,
which is a sense of wonder about science and technology.
And he certainly has that.
He's also a wonderful person.
I mean, in terms of like as a manager,
like dealing with people and everything,
Mark also actually.
So, I mean, they're very like, you know, very human people.
In the case of Mark, it's shockingly human,
you know, given his trajectory.
I mean, the personality of him
that he's painting in the press is just completely wrong.
- Yeah.
But you have to know how to play the press.
So that's, I put some of that responsibility on him too.
You have to, it's like, you know, like the director,
the conductor of an orchestra,
you have to play the press and the public
in a certain kind of way
where you convey your true self to them.
If there's a depth of kindness to it.
- It's hard.
And it's probably not the best at it.
So, yeah.
You have to learn.
And it's sad to see, I'll talk to him about it,
but Shrepp is slowly stepping down.
It's always sad to see folks sort of be there
for a long time and slowly, I guess time is sad.
- I think he's done the thing he set out to do.
And, you know, he's got, you know,
family priorities and stuff like that.
And I understand, you know, after 13 years or something.
- It's been a good run.
- Which in Silicon Valley is basically a lifetime.
- Yeah.
- You know, cause you know, it's dog years.
- So in Europe, the conference just wrapped up.
Let me just go back to something else.
You posted that the paper you coauthored
was rejected from Europe.
As you said, proudly in quotes, rejected.
- I'm a joke.
- Yeah, I know.
Can you describe this paper and like,
what was the idea in it?
And also maybe this is a good opportunity
to ask what are the pros and cons,
what works and what doesn't about the review process?
- Yeah, let me talk about the paper first.
I'll talk about the review process afterwards.
The paper is called VICREG.
So this is, I mentioned that before,
variance, invariance, covariance, regularization.
And it's a technique, a non-contrastive learning technique
for what I call joint embedding architecture.
So Siamese nets are an example
of joint embedding architecture.
So joint embedding architecture is,
let me back up a little bit, right?
So if you want to do self-supervised learning,
you can do it by prediction.
So let's say you want to train a system to predict video,
right?
You show it a video clip
and you train the system to predict the next,
the continuation of that video clip.
Now, because you need to handle uncertainty,
because there are many, you know,
many continuations that are plausible,
you need to have, you need to handle this in some way.
You need to have a way for the system
to be able to produce multiple predictions.
And the way, the only way I know to do this
is through what's called a latent variable.
So you have some sort of hidden vector
of a variable that you can vary over a set
or draw from a distribution.
And as you vary this vector over a set,
the output, the prediction,
varies over a set of plausible predictions.
Okay, so that's called,
I call this a generative latent variable model.
- Got it.
- Okay, now there is an alternative to this,
to handle uncertainty.
And instead of directly predicting
the next frames of the clip,
you also run those through another neural net.
So you now have two neural nets,
one that looks at the, you know,
the initial segment of the video clip,
and another one that looks at the continuation
during training, right?
And what you're trying to do is learn a representation
of those two video clips
that is maximally informative
about the video clips themselves,
but is such that you can predict the representation
of the second video clip
from the representation of the first one, easily.
Okay, and you can sort of formalize this
in terms of maximizing mutual information
and some stuff like that, but it doesn't matter.
What you want is informative representations
of the two video clips that are mutually predictable.
What that means is that there's a lot of details
in the second video clips that are irrelevant.
You know, let's say a video clip consists in,
you know, a camera panning the scene.
There's going to be a piece of that room
that is going to be revealed,
and I can somewhat predict
what that room is going to look like,
but I may not be able to predict the details
of the texture of the ground
and where the tiles are ending and stuff like that, right?
So those are irrelevant details
that perhaps my representation will eliminate.
And so what I need is to train this second neural net
in such a way that whenever the continuation
video clip varies over all the plausible continuations,
the representation doesn't change.
- Got it.
So it's the, yeah, yeah, got it.
Over the space of representations,
doing the same kind of thing
as you're doing with similarity learning.
- Right.
So these are two ways to handle
multimodality in a prediction, right?
In the first way, you parametrize the prediction
with a latent variable,
but you predict pixels essentially, right?
In the second one, you don't predict pixels,
you predict an abstract representation of pixels
and you guarantee that this abstract representation
has as much information as possible about the input,
but sort of, you know,
drops all the stuff that you really can't predict,
essentially.
I used to be a big fan of the first approach.
And in fact, in this paper with Ishan Mishra,
this blog post, the dark matter intelligence,
I was kind of advocating for this.
And in the last year and a half,
I've completely changed my mind.
I'm now a big fan of the second one.
And it's because of a small collection of algorithms
that have been proposed over the last year and a half
or so, two years to do this, including V-Craig,
its predecessor called Barlow-Twins,
which I mentioned, a method from our friends
at DeepMind called BYOL.
And there's a bunch of others now that kind of work similarly.
So they're all based on this idea of joint embedding.
Some of them have an explicit criterion
that is an approximation of mutual information.
Some others at BYOL work, but we don't really know why.
And there's been like lots of theoretical papers
about why BYOL works.
No, it's not that because we take it out
and it still works and blah, blah, blah.
I mean, so there's like a big debate,
but the important point is that we now have a collection
of non-contrastive joint embedding methods,
which I think is the best thing since sliced bread.
§16Transcript · 2:11:56–2:20:43
So I'm super excited about this
because I think it's our best shot for techniques
that would allow us to kind of build predictive world models
and at the same time,
learn hierarchical representations of the world
where what matters about the world is preserved
and what is irrelevant is eliminated.
- And by the way, the representations of before and after
is in the space in a sequence of images,
or is it for single images?
- It would be either for a single image, for a sequence.
It doesn't have to be images.
This could be applied to text.
This could be applied to just about any signal.
I'm looking for methods that are generally applicable
that are not specific to one particular modality.
It could be audio or whatever.
- Got it.
So what's the story behind this paper?
This paper is what, is describing one such method?
- This is this Vick-Reich method.
So this is co-authored,
the first author is a student called Adrien Barne,
who is a resident PhD student at Fer Paris,
who is co-advised by me and Jean Ponce,
who's a professor at Ecole Normale Supérieure,
also a research director at INRIA.
So this is a wonderful program in France
where PhD students can basically do their PhD in industry.
And that's kind of what's happening here.
And this paper is a follow-up on this Barlow-Twin paper
by my former postdoc, now Stéphane Denis,
with Li Jing and Joris Bontart
and a bunch of other people from Fer.
And one of the main criticism from reviewers
is that Vick-Reich is not different enough from Barlow-Twins.
But my impression is that it's Barlow-Twins
with a few bugs fixed essentially.
And in the end, this is what people will use.
- Right.
- But I'm used to stuff that I submit being rejected for a while.
- So it might be rejected and actually exceptionally well cited
'cause people use it.
- Well, it's already cited like a bunch of times.
- So, I mean, the question is then to the deeper question
about peer review and conferences.
I mean, computer science as a field is kind of unique
that the conference is highly prized.
That's one.
- Right.
- And it's interesting because the peer review process
there is similar, I suppose, to journals,
but it's accelerated significantly.
Well, not significantly, but it goes fast.
And it's a nice way to get stuff out quickly,
to peer review it quickly,
go to present it quickly to the community.
So not quickly, but quicker.
- Yeah.
- But nevertheless, it has many of the same flaws
of peer review 'cause it's a limited number
of people look at it.
There's bias and the following,
like that if you wanna do new ideas,
you're gonna get pushed back.
There's self-interested people that kind of can infer
who submitted it and kind of be cranky about it,
all that kind of stuff.
- Yeah, I mean, there's a lot of social phenomenon there.
There's one social phenomenon, which is that
because the field has been growing exponentially,
the vast majority of people in the field
are extremely junior.
- Yeah.
- So as a consequence, and that's just a consequence
of the field growing, right?
So as the number of, as the size of the field
kind of starts saturating, you will have less
of that problem of reviewers being very inexperienced.
A consequence of this is that young reviewers,
I mean, there's a phenomenon which is that
reviewers try to make their life easy
and to make their life easy when reviewing a paper
is very simple, you just have to find a flaw in the paper.
Right?
So basically they see their task as finding flaws in papers
and most papers have flaws, even the good ones.
- Yeah.
- And so it's easy to do that.
Your job is easier as a reviewer if you just focus on this.
But what's important is like, is there a new idea
in that paper that is likely to influence?
It doesn't matter if the experiments are not that great,
if the protocol is, you know, so-so, you know,
things like that, as long as there is a worthy idea in it
that will influence the way people think about the problem,
even if they make it better, you know, eventually,
I think that's really what makes a paper useful.
And so this combination of social phenomena
creates a disease that has plagued, you know,
other fields in the past, like speech recognition,
where basically, you know, people chase numbers
on benchmarks and it's much easier to get a paper accepted
if it brings an incremental improvement
on a sort of mainstream, well-accepted method or problem.
And those are, to me, boring papers.
I mean, they're not useless, right?
Because industry, you know, strives on those kind of progress
but they're not the one that I'm interested in
in terms of like new concepts and new ideas.
So papers that are really trying to strike
kind of new advances generally don't make it.
Now, thankfully, we have Archive.
- Archive, exactly.
And then there's open review type of situations where you,
and then, I mean, Twitter is a kind of open review.
I'm a huge believer that review should be done
by thousands of people, not two people.
- I agree.
- And so Archive, like do you see a future
where a lot of really strong papers,
it's already the present, but a growing future
where it'll just be Archive
and you're presenting an ongoing, continuous conference
called Twitter/the Internet/Archive Sanity.
Andre just released a new version.
So just not, you know, not being so elitist
about this particular gating.
- It's not a question of being elitist or not.
It's a question of being basically recommendation
and seal of approvals for people who don't see themselves
as having the ability to do so by themselves, right?
And so it saves time, right?
If you rely on other people's opinion
and you trust those people or those groups
to evaluate a paper for you,
that saves you time.
'Cause you don't have to like scrutinize the paper as much.
It is brought to your attention.
I mean, there's the whole idea
of sort of collective recommender system, right?
So I actually thought about this a lot,
you know, about 10, 15 years ago,
'cause there were discussions at NIPS
and we're about to create iClear with Yoshua Bengio.
And so I wrote a document kind of describing
a reviewing system, which basically was,
you post your paper on some repository,
let's say archive or now could be open review.
And then you can form a reviewing entity,
which is equivalent to a reviewing board,
you know, of a journal or a program committee
of a conference.
You have to list the members
and then that group, reviewing entity,
can choose to review a particular paper,
spontaneously or not.
There is no exclusive relationship anymore
between a paper and a venue or reviewing entity.
Any reviewing entity can review any paper
or may choose not to.
And then, you know, give an evaluation.
It's not published, not published,
it's just an evaluation and a comment,
which would be public, signed by the reviewing entity.
And if it's signed by a reviewing entity,
you know, it's one of the members of reviewing entity.
So if the reviewing entity is, you know,
Lex Friedman's, you know, preferred papers, right?
You know, it's Lex Friedman writing the review.
- Yes.
So for me, that's a beautiful system, I think.
But what's in addition to that,
it feels like there should be a reputation system
for the reviewers.
- For the reviewing entities,
not the reviewers individually.
- The reviewing entities, sure.
But even within that, the reviewers too,
because there's another thing here.
It's not just the reputation,
it's an incentive for an individual person to do great.
Right now, in the academic setting,
the incentive is kind of internal,
just wanting to do a good job.
But honestly, that's not a strong enough incentive
to do a really good job in reading a paper,
in finding the beautiful amidst the mistakes and the flaws
and all that kind of stuff.
Like, if you're the person
that first discovered a powerful paper,
and you get to be proud of that discovery,
then that gives a huge incentive to you.
- That's a big part of my proposal, actually,
where I describe that as, you know,
if your evaluation of papers
is predictive of future success,
then your reputation should go up as a reviewing entity.
So, yeah, exactly.
I mean, I even had a master's student
§17Transcript · 2:20:46–2:29:08
who was a master's student in library science
and computer science,
actually kind of work out exactly
how that should work with formulas and everything.
- So, in terms of implementation,
do you think that's something that's doable?
- I mean, I've been sort of talking about this
to sort of various people like, you know,
Andrew McCallum, who started Open Review.
And the reason why we picked Open Review
for iClear initially,
even though it was very early for them,
is because my hope was that iClear,
it was eventually going to kind of
inaugurate this type of system.
So iClear kept the idea of open reviews.
So where the reviews are, you know,
published with a paper, which I think is very useful.
But in many ways, that's kind of reverted
to kind of more of a conventional type conferences
for everything else.
And that, I mean, I don't run iClear,
I'm just the president of the foundation,
but, you know, people who run it
should make decisions about how to run it.
And I'm not going to tell them
because they are volunteers
and I'm really thankful that they do that.
So, but I'm saddened by the fact that
we're not being innovative enough.
- Yeah, me too.
I hope that changes.
Yeah, 'cause the communication of science broadly,
but the communication of computer science ideas
is how you make those ideas have impact, I think.
- Yeah, and I think, you know, a lot of this is
because people have in their mind kind of an objective,
which is, you know, fairness for authors
and the ability to count points basically
and give credits accurately.
But that comes at the expense of the progress of science.
So to some extent,
we're slowing down the progress of science.
- And are we actually achieving fairness?
- And we're not achieving fairness, you know,
we still have biases, you know,
we're doing, you know, a double-blind review,
but, you know, the biases are still there.
There are different kinds of biases.
- You write that the phenomenon of emergence,
collective behavior exhibited by a large collection
of simple elements in interaction
is one of the things that got you
into neural nets in the first place.
I love cellular automata,
I love simple interacting elements
and the things that emerge from them.
Do you think we understand how complex systems
can emerge from such simple components
that interact simply?
- No, we don't.
It's a big mystery.
Also, it's a mystery for physicists,
it's a mystery for biologists.
You know, how is it that the universe around us
seems to be increasing in complexity and not decreasing?
I mean, that is a kind of curious property of physics
that despite the second law of thermodynamics,
we seem to be, you know, evolution and learning
and et cetera seems to be kind of at least locally
to increase complexity, not decrease it.
So perhaps the ultimate purpose of the universe
is to just get more complex.
- Have these, I mean, small pockets of beautiful complexity.
Does that, does cellular automata,
do these kinds of emergence of complex systems
give you some intuition or guide your understanding
of machine learning systems and neural networks and so on?
Or are these for you right now, disparate concepts?
- Well, it got me into it.
You know, I discovered the existence of the perceptron
when I was a college student, you know,
by reading a book on, it was a debate between Chomsky
and Piaget and Seymour Papert from MIT
was kind of singing the praise of the perceptron
in that book.
And I, the first time I heard about the learning machine,
right, so I started digging the literature
and I found those books,
which were basically transcription of, you know,
workshops or conferences from the 50s and 60s
about self-organizing systems.
So there were, there was a series of conferences
on self-organizing systems and these books on this.
Some of them are, you can actually get them
at the internet archive, you know, the digital version.
And there are like fascinating articles in there by,
there's a guy whose name has been largely forgotten,
Heinz von Foerster.
So it was a German physicist who immigrated to the US
and worked on self-organizing systems in the 50s.
And in the 60s, he created,
at University of Illinois Urbana-Champaign,
he created the biological computer laboratory, BCL,
which was, you know, all about neural nets.
Unfortunately, that was kind of towards the end
of the popularity of neural nets.
So that lab never kind of strived very much,
but he wrote a bunch of papers about self-organization
and about the mystery of self-organization.
An example he has is, you take,
imagine you are in space, there's no gravity,
you have a big box with magnets in it, okay?
You know, kind of rectangular magnets
with North Pole on one end, South Pole on the other end.
You shake the box gently and the magnets
will kind of stick to themselves
and probably form a complex structure, you know,
spontaneously, you know,
that could be an example of self-organization.
But, you know, you have lots of examples,
neural nets are an example of self-organization too,
you know, in many respect.
And it's a bit of a mystery, you know,
how, like what is possible with this, you know,
pattern formation in physical systems,
in chaotic system and things like that, you know,
the emergence of life, you know, things like that.
So, you know, how does that happen?
So it's a big puzzle for physicists as well.
- It feels like understanding this,
the mathematics of emergence in some constrained situations
might help us create intelligence.
Like help us add a little spice to the systems
because you seem to be able to,
in complex systems with emergence,
to be able to get a lot from little.
And so that seems like a shortcut
to get big leaps in performance.
But-
- But there's a missing theoretical concept
that we don't have.
- Yeah.
- And it's something also I've been fascinated by
since my undergrad days.
And it's how you measure complexity, right?
So we don't actually have good ways of measuring,
or at least we don't have good ways of interpreting
the measures that we have at our disposal.
Like how do you measure the complexity of something, right?
So there's all those things, you know,
like, you know, Kolmogorov, Chaitin, Solomonov complexity
of, you know, the length of the shortest program
that would generate a bit string
can be thought of as the complexity of that bit string.
- Mm-hmm.
- I've been fascinated by that concept.
The problem with that is that that complexity
is defined up to a constant, which can be very large.
- Right.
- There are similar concepts that are derived from,
you know, Bayesian probability theory,
where, you know, the complexity of something
is the negative log of its probability, essentially, right?
And you have a complete equivalence between the two things.
And there you would think, you know,
the probability is something
that's well-defined mathematically,
which means complexity is well-defined.
But it's not true.
You need to have a model of the distribution.
You may need to have a prior
if you're doing Bayesian inference.
And the prior plays the same role
as the choice of the computer
with which you measure Kolmogorov complexity.
And so every measure of complexity we have
has some arbitrariness in it.
You know, an additive constant,
which can be arbitrarily large.
And so, you know, how can we come up with a good theory
of how things become more complex
if we don't have a good measure of complexity?
- Yeah, which we need for,
there's one way that people study this
in the space of biology,
the people that study the origin of life
or try to recreate life in the laboratory.
And the more interesting one is the alien one,
is when we go to other planets,
how would we recognize this life?
'Cause, you know, complexity, we associate complexity,
maybe some level of mobility with life.
You know, we have to be able to like have concrete algorithms
for like measuring the level of complexity we see
in order to know the difference between life and non-life.
- And the problem is that complexity
is in the eye of the beholder.
So let me give you an example.
If I give you an image of the MNIST digits, right?
§18Transcript · 2:29:13–2:37:38
And I flip through MNIST digits,
there is some, obviously some structure to it
because local structure, you know,
neighboring pixels are correlated
across the entire dataset.
I imagine that I apply a random permutation
to all the pixels, a fixed random permutation.
I show you those images, they will look, you know,
really disorganized to you, more complex.
In fact, they're not more complex in absolute terms,
they're exactly the same as originally, right?
And if you knew what the permutation was, you know,
you could undo the permutation.
Now, imagine I give you special glasses
that undo that permutation.
Now, all of a sudden, what looked complicated becomes simple.
- Right.
- So if you have two, if you have, you know,
humans on one end and then another race of aliens
that sees the universe with permutation glasses.
- Yeah, with the permutation glasses.
(Lex laughing)
- What we perceive as simple to them is hardly complicated,
it's probably heat.
- Yeah, heat, yeah.
- Okay, and what they perceive as simple to us
is random fluctuation, it's heat.
- Yeah.
- So-
- Truly in the eye of the beholder,
depends what kind of glasses you're wearing.
- Right.
- Depends what kind of algorithm you're running
in your perception system.
- So I don't think we'll have a theory of intelligence,
self-organization, evolution, things like this
until we have a good handle on a notion of complexity,
which we know is in the eye of the beholder.
- Yeah, it's sad to think that we might not be able
to detect or interact with alien species
because we're wearing different glasses.
- Because the notion of locality
might be different from ours.
- Yeah, exactly.
- This actually connects with fascinating questions
in physics at the moment, like modern physics,
quantum physics, like, you know, questions about,
like, you know, can we recover the information
that's lost in a black hole and things like this, right?
And that relies on notions of complexity,
which, you know, I find this fascinating.
- Can you describe your personal quest
to build an expressive electronic wind instrument, EWI?
What is it?
What does it take to build it?
- Well, I'm a tinkerer.
I like building things.
I like building things with combinations of electronics
and, you know, mechanical stuff.
You know, I have a bunch of different hobbies,
but, you know, probably my first one was little,
was building model airplanes and stuff like that.
And I still do that to some extent,
but also electronics.
I taught myself electronics before I studied it.
And the reason I taught myself electronics
is because of music.
My cousin was an aspiring electronic musician
and he had an analog synthesizer.
And I was, you know, basically modifying it for him
and building sequencers and stuff like that, right, for him.
I was in high school when I was doing this.
- How's the interest in like progressive rock, like '80s?
Like what's the greatest band of all time,
according to Yann LeCun?
- Oh, man, there's too many of them.
But, you know, it's a combination of, you know,
Mahavishnu Orchestra, Weather Report,
Yes, Genesis, you know,
- Yes, Genesis.
- Pre-Peter Gabriel, Gentle Giant, you know,
things like that.
- Great.
Okay, so this love of electronics
and this love of music combined together.
- Right, so I was actually trained to play
Baroque and Renaissance music.
And I played in an orchestra when I was in high school
and first year of college.
And I played the recorder, cram horn,
a little bit of oboe, you know, things like that.
So I'm a wind instrument player.
But I always wanted to play improvised music,
even though I don't know anything about it.
And the only way I figured, you know,
short of like learning to play a saxophone
was to play electronic wind instruments.
So they behave, you know,
the fingering is similar to a saxophone,
but, you know, you have wide variety of sound
because you control the synthesizer with it.
So I had a bunch of those, you know,
going back to the late 80s from either Yamaha or Akai.
They're both kind of the main manufacturers of those.
So they were classically, you know,
going back several decades.
But I've never been completely satisfied with them
because of lack of expressivity.
And, you know, those things, you know,
are somewhat expressive.
I mean, they measure the breath pressure,
they measure the lip pressure,
and, you know, you have various parameters.
You can vary it with fingers,
but they're not really as expressive
as an acoustic instrument, right?
You hear John Coltrane play two notes
and you know it's John Coltrane,
you know, it's got a unique sound.
Or Miles Davis, right?
You can hear it's Miles Davis playing the trumpet
because the sound reflects their, you know,
physiognomy, basically.
The shape of the vocal track kind of shapes the sound.
So how do you do this with an electronic instrument?
And I was, many years ago I met a guy called David Wessel.
He was a professor at Berkeley
and created the Center for like, you know,
music technology there.
And he was interested in that question.
And so I kept kind of thinking about this for many years.
And finally, because of COVID, you know, I was at home.
I was in my workshop.
My workshop serves also as my kind of Zoom room
and home office.
- This is in New Jersey?
- In New Jersey.
And I started really being serious about, you know,
building my own EWI instrument.
- What else is going on in that New Jersey workshop?
Is there some crazy stuff you've built?
Like just, or like left on the workshop floor, left behind?
- A lot of crazy stuff is, you know,
electronics built with microcontrollers of various kinds
and, you know, weird flying contraptions.
- So you still love flying?
- It's a family disease.
My dad got me into it when I was a kid
and he was building model airplanes when he was a kid.
And he was a mechanical engineer.
He taught himself electronics also.
So he built his early radio control systems
in the late sixties, early seventies.
And so that's what got me into, I mean,
he got me into kind of, you know,
engineering and science and technology.
- Do you also have an interest in appreciation of flight
in other forms, like with drones, quadropters, or do you,
is it model airplane, the thing that's-
- You know, before drones were, you know,
kind of a consumer product, you know,
I built my own, you know,
with also building a microcontroller
with a gyroscopes and accelerometers for stabilization,
writing the firmware for it, you know.
And then when it became kind of a standard thing
you could buy, it was boring, you know,
I stopped doing it, it was not fun anymore.
- Yeah, you were doing it before it was cool.
- Yeah.
- What advice would you give to a young person today
in high school and college that dreams of doing
something big like Yann LeCun,
like let's talk in the space of intelligence,
dreams of having a chance to solve
some fundamental problem in space of intelligence,
both for their career and just in life,
being somebody who was a part of creating something special?
- So try to get interested by big questions,
things like, you know, what is intelligence?
What is the universe made of?
What's life all about?
Things like that.
Like even like crazy big questions, like what's time?
Like nobody knows what time is.
And then learn basic things, like basic methods,
either from math, from physics or from engineering.
Things that have a long shelf life.
Like if you have a choice between like, you know,
learning, you know, mobile programming on iPhone
or quantum mechanics, take quantum mechanics.
Because you're gonna learn things
that you have no idea exist.
And you may not, you may never be a quantum physicist,
but you'll learn about path integrals
and path integrals are used everywhere.
It's the same formula that you use for, you know,
Bayesian integration and stuff like that.
- So the ideas, the little ideas within quantum mechanics,
within some of these kind of more solidified fields
§19Transcript · 2:37:41–2:45:04
will have a longer shelf life.
They'll somehow use indirectly in your work.
- Learn classical mechanics,
like you learn about Lagrangians, for example.
Which is like a hugely useful concept, you know,
for all kinds of different things.
Learn statistical physics, because all the math
that comes out of, you know, for machine learning,
basically comes out of what's figured out
by statistical physicists in the, you know,
late 19th, early 20th century, right?
So, and for some of them actually more recently,
for by people like Giorgio Parisi,
who just got the Nobel prize for the replica method,
among other things, it's used for a lot of different things.
You know, variational inference,
that math comes from statistical physics.
So, a lot of those kind of, you know, basic courses,
you know, if you do electrical engineering,
you take signal processing,
you'll learn about Fourier transforms.
Again, something super useful is at the basis
of things like graph neural nets,
which is an entirely new sub area of, you know,
AI machine learning, deep learning,
which I think is super promising
for all kinds of applications.
Something very promising,
if you're more interested in applications,
is the applications of AI machine learning
and deep learning to science.
Or to science that can help solve big problems in the world.
I have colleagues at Meta, at FAIR,
who started this project called Open Catalyst,
and it's an open project collaborative.
And the idea is to use deep learning
to help design new chemical compounds or materials
that would facilitate the separation
of hydrogen from oxygen.
If you can efficiently separate oxygen from hydrogen
with electricity, you solve climate change.
It's as simple as that.
Because you cover, you know,
some random desert with solar panels,
and you have them work all day, produce hydrogen,
and then you shoot the hydrogen wherever it's needed.
You don't need anything else.
You know, you have controllable power
that's, you know, can be transported anywhere.
So if we have a large-scale, efficient
energy storage technology, like producing hydrogen,
we solve climate change.
Here's another way to solve climate change,
is figuring out how to make fusion work.
Now, the problem with fusion
is that you make a super-hot plasma,
and the plasma is unstable, and you can't control it.
Maybe with deep learning, you can find controllers
that will stabilize plasma
and make, you know, practical fusion reactors.
I mean, that's very speculative,
but, you know, it's worth trying,
because, you know, the payoff is huge.
There's a group at Google working on this,
led by John Platt.
- So, control, convert as many problems
in science and physics and biology and chemistry
into a learnable problem,
and see if a machine can learn it.
- Right, I mean, there's properties of, you know,
complex materials that we don't understand
from first principle, for example, right?
So, you know, if we could design new, you know,
new materials, we could make more efficient batteries.
You know, we could make maybe faster electronics.
We could, I mean, there's a lot of things we can imagine
doing, or, you know, lighter materials
for cars or airplanes and things like that.
Maybe better fuel cells.
I mean, there's all kinds of stuff we can imagine.
If we had good fuel cells, hydrogen fuel cells,
we could use them to power airplanes,
and, you know, transportation wouldn't be, or cars,
we wouldn't have emission problem,
CO2 emission problems for air transportation anymore.
So, there's a lot of those things,
I think, where AI, you know, can be used.
And this is not even talking about all the sort of
medicine, biology, and everything like that, right?
You know, like protein folding, you know,
figuring out, like, how can you design your proteins
that it sticks to another protein at a particular site,
because that's how you design drugs in the end.
So, you know, deep learning would be useful,
all of this, and those are kind of, you know,
would be sort of enormous progress
if we could use it for that.
Here's an example.
If you take, this is like from recent material physics,
you take a monoatomic layer of graphene, right?
So, it's just carbon on an hexagonal mesh,
and you make this single atom thick.
You put another one on top,
you twist them by some magic number of degrees,
three degrees or something, it becomes superconductor.
Nobody has any idea why.
(both laughing)
- I wanna know how that was discovered,
but that's the kind of thing that machine learning
can actually discover, these kinds of things.
- Maybe not, but there is a hint, perhaps,
that with machine learning, we would train a system
to basically be a phenomenological model
of some complex emergent phenomenon,
which superconductivity is one of those,
where this collective phenomenon is too difficult
to describe from first principles
with the usual sort of reductionist type method.
But we could have deep learning systems
that predict the properties of a system
from a description of it,
after being trained with sufficiently many samples.
This guy, Pascal Fouad, at EPFL,
he has a startup company
where he basically trained a convolutional net,
essentially, to predict the aerodynamic properties of solids.
And you can generate as much data as you want
by just running computational free dynamics, right?
So, you give a wing airfoil or something,
a shape of some kind,
and you run computational free dynamics,
you get, as a result, the drag and lift
and all that stuff, right?
And you can generate lots of data,
train a neural net to make those predictions,
and now what you have is a differentiable model
of, let's say, drag and lift,
as a function of the shape of that solid.
And so you can do background and design,
you can optimize the shape,
so you get the properties you want.
- Yeah, that's incredible.
That's incredible.
And on top of all that,
probably you should read a little bit of literature
and a little bit of history for inspiration and for wisdom,
'cause after all, all of these technologies
will have to work in the human world.
- Yes.
- And the human world is complicated.
- It is, certainly.
- Jan, this is an amazing conversation.
I'm really honored that you would talk with me today.
Thank you for all the amazing work
you're doing at FAIR, at META,
and thank you for being so passionate
after all these years about everything that's going on.
You're a beacon of hope for the machine learning community.
And thank you so much
for spending your valuable time with me today.
That was awesome.
- Thanks for having me on.
That was a pleasure.
- Thanks for listening to this conversation with Jan Lekun.
To support this podcast,
please check out our sponsors in the description.
And now, let me leave you with some words
from Isaac Asimov.
"Your assumptions are your windows on the world.
"Scrub them off every once in a while,
"or the light won't come in."
Thank you for listening,
and hope to see you next time.
(upbeat music)
(upbeat music)