The following is a conversation with Yann LeCun,

his second time in the podcast.

He is the chief AI scientist at Meta, formerly Facebook,

professor at NYU, touring 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 it was 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 Yann LeCun.

You cowrote the article,

self supervised learning, the dark matter of intelligence.

Great title, by the way, with Yann 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 pydimes I should say,

are supervised learning and reinforcement learning,

and they are extremely efficient.

Supervised learning requires many samples

for learning anything,

and reinforcement learning requires a ridiculously large

number of trial and errors for a system to run anything.

And that's why we don't have self driving cars.

That's a big leap from one to the other, okay?

So 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 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 the immediate response you get from many people

is, well, 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,

how do babies in the 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 supervision learning is 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 we mean by background knowledge.

Perhaps not even watching mom and dad,

just watching the world go by.

Just having eyes open or having eyes closed

or the very active 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 basics objects and all that kind of stuff.

That's right.

I mean, the basic physics of objects,

you don't even need to 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, we'll 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,

tabular rice reinforcement learning system

that doesn't have a model of the world,

you have to repeat folding 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.

The cake has someone that's 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 scalar,

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 scalar.

So as a consequence, you could not possibly learn

something very complicated without many, many, many trials

where you get many, many feedbacks of this type.

In 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 1,000 categories,

that's a little less than 10 bits of information per sample.

But self supervised learning here is the setting.

You, 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.

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,

of 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.

So...

Filling the gaps.

It sounds dumb, but can you?

It is possible that 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 going to 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 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 is our best shot.

So I think this idea of an intelligent 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 a basic knowledge about how humans are made.

And I don't know if you're gonna,

what word you're 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,

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 the physics.

You have a model of what's possible, 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,

retradicting 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 recorded 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, you know,

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 modern natural language processing

is pre trained in self supervised manner

to fill in the blanks.

So 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, you know,

something that you train 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 necessary

because it's fundamentally more difficult.

Maybe like when we're talking about achieving,

like passing the touring test in the full spirit

of the touring 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?

When 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 we start supervising

is how you represent this uncertainty,

how you represent multiple discrete outcomes,

how you represent a continuum of possible outcomes, et cetera.

And if you are sort of a classical machine learning person,

you say, oh, you just represent a distribution, right?

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 the dictionary.

It's a big list of numbers, maybe 100,000 or so.

And you can turn them into a parallel 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, right?

And if I say the blank is chasing the blank in the savannah,

you also have a bunch of plausible options

for those two words, right?

Because you have kind of a underlying reality

you can refer to to sort of fill in those blanks.

So you cannot say for sure in the savannah

if it's a lion or a cheetah or whatever,

you cannot know if it's a zebra or a do or whatever,

wildebeest, the same thing.

But you can represent the uncertainty

by just a long list of numbers.

Now, if I do the same thing with video

when I ask you to predict a video clip,

it's not a discrete set of potential frames.

You have to have somewhere 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 to 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 do you have any interesting ideas

about how to represent all the reality in a compressed way

such that you can form a distribution over?

That's one of the big questions, you know,

how do you do that?

But I mean, what's kind of, you know,

another thing that really is stupid about,

I shouldn't say stupid,

but like simplistic about current approaches

to self supervision in NLP in text

is that not only do you represent

a giant distribution over words,

but for multiple words that are missing,

those distributions are essentially independent

of each other.

And, you know, you don't pay too much of a price for this.

So you can't.

So, you know, the system, you know,

in the sentence that I gave earlier,

if it gives a certain probability for lion and cheetah,

and then a certain probability for, you know,

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, you know, 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,

when they're, because the number grows exponentially

with the length of the symbols.

And so we have to use tricks for this,

but those techniques can, you know, get around,

like don't even deal with it.

So the big question is, like, would there be

some sort of abstract latent representation of text

that would say that, you know, 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 feeling 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 just statistics.

It's not intelligence.

Oh, 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 are kind of pointless.

So first of all, it's quite possible

that intelligence is just statistics.

It's just statistics of a particular kind.

Yes.

Where this is the philosophical question.

Is it possible that intelligence is just statistics?

Yeah.

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,

a current machine learning system don't care about causality,

which by the way is wrong, I agree with that.

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 of 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 learned 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 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,

competition on 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 like with machines, we're not even at the level

of even reproducing the learning processes in a cat brain.

You know, the most intelligent of our intelligence 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

like how they can act on the world to get what they want.

They certainly have a fantastic model of intuitive physics,

certainly of 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 like the high level cognition

and kind of long term planning and reasoning that humans

can do until we figure out like,

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 running machines that can also reason.

So whenever I give a talk, I say there are three challenges

in the three main challenges in machine learning.

The first one is, you know,

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, you know, like, you know,

we know how to trend them to learn hierarchical representations

of perception, you know, 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, you know, 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 giving that background knowledge be able to make

hierarchical plans in the world.

Right.

So if you take classical optimal control,

there's something classical optimal control called

model predictive control.

And it's, you know, 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 the system.

So for a rocket to be thrust and, you know,

all the controls you can have.

It gives you the state of the system at time t plus delta t.

Right.

So basically 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 enroll that model in time.

You feel it a hypothesized sequence of actions.

And then you have some objective function that measures

how well at the end of the trajectory of the system

has succeeded or matched what you wanted to do.

You know, is it a robot harm?

Have you grasped the object you want to grasp?

If it's a rocket, you know, 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, you know, 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, you know, 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 going to 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 going to be bad.

So you can sort of imagine different scenarios

and then, you know, 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, you know, all those things have names, you know,

going back, you know, decades.

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 learn predictive models of the world

that deal with uncertainty and deal with the real world

in all this complexity.

So it's not just trajectory of a rocket,

which you can reduce to first principles.

It's not even just a 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 deal 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 is a, I go, you go, I go, you go.

Like Andre Carpathia 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,

the ontology of what defines a car door,

this big philosophical question.

But to me, it was interesting because 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're slowing down is going to signal something.

You're speeding up is going to signal something.

And 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,

because 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,

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're, we've evolved to solve.

This is the kind of problem we've evolved to solve.

And so we're good at it because, you know, 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, as a game is to be challenging.

And if there is something that, you know,

recent progress in chess and Go has made us realize is that humans

are really terrible at those things.

Like really bad.

You know, there was a story, right?

Before AlphaGo that, you know, the best Go player thought there were

maybe two or three stones behind, you know,

an ideal player that they would call God.

In fact, no, there are like nine or 10 stones behind.

I mean, we're just bad.

So we're not good at, and it's because we have limited working memory.

We, you know, we're not very good at like doing this tree exploration

that, you know, computers are much better at doing than we are.

But we are much better at learning differentiable models to the world.

I mean, I said differentiable in the kind of, you know,

I should say not differentiable in the sense that, you know,

we went back far through it, but in the sense that our brain has some

mechanism for estimating gradients of some kind.

And that's what, you know, makes us efficient.

So if you have an agent that consists of a model of the world,

which, you know, 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, you know, 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, you know, the equivalent of what people

in Reference Metronomy 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, you know, 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, you know, your world model, then you can use

gradient based type methods to do planning, to do reasoning,

to do learning, you know, to do all the things that would like

an intelligent agent to do.

And 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.

And 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,

is that, you know, all the learning algorithms we know from

engineering and computer science proceed by optimizing

some objective function.

Yeah.

So one question we may ask is, is does learning in the brain

minimize an objective function?

I mean, it could be, you know, a composite of multiple

objective functions, but it's still an objective function.

Second, if it does optimize an objective function, does it do,

does it do it by some sort of gradient estimation?

You know, it doesn't need to be backprop, but, you know, some

way of estimating the gradient in efficient manner, whose

complexity is on the same order of magnitude as, you know,

actually running the inference.

Because 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 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 quarried.

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 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.

Well, I'm terrible at chess.

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

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 can imagine in advance, like, will we be the result of,

you know, cutting the wood in this particular way?

Are you going to use, you know, screws on nails or whatever?

When you're interacting with someone, you also have a model

in mind to kind of tell the person what you think is useful

to them.

So I think this ability to construct models to 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, has that 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 going, 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.

The representation does.

Everything, all knowledge, everything, right?

There's less than a billion.

A dog is 2 billion, but a cat is less than 1 billion.

And so multiply that by 1000 and you get the number of synapses.

And I think almost all of it is learned through a sort of

self supervised running.

Although I think a tiny sliver is learned through reinforcement

running and certainly very little through classical

supervised running, although it's not even clear how

supervised running actually works in the biological world.

So I think almost all of it is self supervised running.

But it's driven by the sort of ingrained objective functions

that a cat or human have at the base of their brain,

which kind of drives their behavior.

So, you know, 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 than 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, you know, the Orbezo Ganglia

drive us to do things that are different from, say, an orangutan

or certainly a cat is what makes, you know, human nature

versus orangutan nature versus cat nature.

So for example, you know, our Bezo Ganglia

drives us to seek the company of other humans.

And that's because nature has figured out that we need to be

a social animal 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.

Because for their survival, you know, evolution has figured out

that's the best thing.

I mean, they're occasionally social, of course, for, you know,

reproduction and stuff like that, but they're mostly solitary.

So all of those behaviors are not part of intelligence.

You know, people say, oh, you're never going to have intelligent

machines because, you know, 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.

You know, in a year of their 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.

We think, 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 for, to solve the house cat intelligence problem,

you think you could do it on a desert island.

You could 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, you know, the right set of drives

to kind of, you know, get it to do the thing

and learn the appropriate things, right?

But like, for example, you know, baby humans are driven

to learn to stand up and walk.

Okay, that's kind of, this desire is hardwired.

How to do it precisely is not, that's learned.

But the desire to walk, move around and stand up,

that's sort of hardwired.

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 got to 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

because 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 you get on two feet?

It's really hard.

Like most animals don't get on two feet.

Well, they get on four feet.

You know, many mammals get on four feet.

Yeah, they do.

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

than 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 do

your work is like, how do you solve MNIST with very little example data?

That's right.

The answer to this probably is self supervised running.

Just learn to represent images and then learning to recognize

handwritten digits on top of this will only require a few samples.

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.

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 because 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, a couple of 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 or weakly supervised transfer learning.

People are making very, very fast progress using self supervised

running with this kind of scenario as well.

And my guess is that that's going to be the future.

For self supervised learning, how much cleaning do you think is needed

for filtering malicious signal or what's the 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.

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 hash TBT or anything like that.

Also, they keep a very select set of hashtags.

Is that what you're saying?

Yeah.

Okay.

But 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 a 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 going to 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 Jeff Newton worked on

also in the early 90s.

There was sort of parallel work.

I used to call this same is network.

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 inputs.

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

the lab, 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

symmetrically similar produce similar representations and

things that are different produce different representations.

So we actually came up with this idea for a project of doing

signature verification.

So we would collect signatures from like multiple signatures on

the same person and then train a neural net to produce the same

representation.

And then, you know, 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.

They were interested in storing 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, you know, that we would quantize on bytes so that we

could encode the signature.

And that encoding was then used to compare whether the

signature matches or not.

That's right.

So then you would, you know, 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, you know, financial payment system is

incredibly lax in that respect compared to Europe.

Oh, the signatures.

What's the purpose of signatures anyway?

Nobody looks at them.

Nobody cares.

Yeah.

So that's contrastive learning, right?

So you need positive and negative pairs.

And the problem with that is that, you know, even though I had

the original paper on this, I 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, you know, Jeff Hinton is the senior member there.

It's called SimClear, SIN, CLR.

And, you know, basically a particular way of implementing this

idea of contrastive learning is a 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 Jeff 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, you know, he proposed that idea had, you know,

a couple of papers in the early 90s, 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 Steph Anthony,

who is now a junior professor in Finland at University of

Alto.

We came up with something called, that we call Balu twins.

And it's a particular way of maximizing the information content

of a vector, you know, using some hypothesis.

And we have kind of another version of it that's more recent

now called VICRAG, V I C A 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 not, I'm really, really excited about this.

What kind of data augmentation is useful for that

noncontrasting learning method?

Are we talking about, does that not matter that much?

Or it seems like a very important part of the step.

Yeah.

Do 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 the documentation you can do that do not

change the nature of the, of the object.

And so the standard scenario, which, you know, a lot of people

working in this area are using is you use the type of distortion.

So, 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 know, you can do a shift in color balance or something like

that.

Saturation.

Another one sort of blurs it.

Another one has 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 surface 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

tasks 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 the localization?

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 know, 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, and now you want to use that.

Oh, like where an object is and what size.

It's like a bounding box.

They'd 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 a ventral pathway for,

you know, telling what you're looking at.

The other path with the dorsal pathway is used for navigation,

for grasping, for everything else.

And, you know, basically a lot of the things you need for

survival are localization and detection.

Is similarity learning or contrastive learning or these

noncontrastive 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 like what is the ceiling of this method, do you think?

Is this just one trick on the path to doing self

supervised learning?

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, you know, of the same nature to train a

system from video to do video prediction, essentially, I think

we'll have a path, you know, towards, you know, I wouldn't

say unlimited, but a path towards some level of, you know,

physical common sense in machines.

And I also think that that ability to learn how the world

works from a sort of high throughput channel like vision

is a necessary step towards sort of 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, let's, and I know people have attempted to do

this for 30 years, right?

The site 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 going to learn about this.

That information is just not present in any text.

Well, the question, like with the site 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 cat? The cat knows that.

No, but they evolved.

No, they learned like us.

Sorry, the physics of stuff.

Well, yeah, so you're saying it's,

you're putting a lot of intelligence onto the nurture side, not the nature.

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.

So 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 intelligence, most of the cool stuff we've been talking about

is mostly nurture, mostly trained.

We figured 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.

Self supervised learning, say.

Whatever it is that makes, you know, human intelligence different from other animals,

which, you know, 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.

And, you know, it only involves, you know, less than 1% of a genome, right?

Which is the difference between human genome and chimps or whatever.

So it can be that complicated, you know, it can be that fundamental.

I mean, the most of the so complicated stuff already existing cats and dogs

and, you know, certainly primates, non human primates.

Yeah, that little thing with humans might be just something about social interaction

and ability to maintain ideas across like 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, you know, we have, I mean, this is number 634, you know, in the list of problems we have to solve.

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

using language, translating these two images, which is basically a denosing auto encoder method.

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 auto encoder type methods for images weren't producing very good representation.

But there's a paper now coming out of the Fair Group in Menlo Park that actually works very well.

So that doesn't require the documentation, that requires only masking.

Only masking for images, okay.

Right, so you mask part of the image and you train a system,

which in this case is a transformer because 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, then masking.

Why should the mask be a square rectangle?

So it doesn't matter.

I think we're going to come up probably in the future with ways to mask that are kind of random, essentially.

I mean, they are random already.

No, but something that's challenging, optimally challenging.

So maybe it's a metaphor that doesn't apply, but it seems like there's an data augmentation or masking.

There's an interactive element with it.

You're almost playing with an image and it's like the way we play with an image in our minds.

No, but it's like dropout.

It's like machine training.

Every time you see a percept, 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 it was a 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, denoting 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 can 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.

Because I was thinking like you might want to be clever about the way you do all these procedures, you know?

But that's only, it's somehow costly to do every iteration, but it's not really.

Not really.

And then there is, you know, data augmentation without explicit data augmentation.

It's 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.

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 of, you know, maybe the right type of data.

You need some...

Don't go down the rabbit hole of just cat videos.

You might need to watch some lectures or something.

No, you...

How meta would that be if it like watches lectures about intelligence and then learns, watches your lectures at NYU and learns from that how to be intelligent?

I don't think that would 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 figure out how to do self supervised representation learning or learning predictive world models.

So I think that's what, you know, the entire community should be focusing on.

At least people 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, multi planetary colonization across the galaxy, eventually the entirety of it will just become giant black holes.

And that's going to 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 your song, you're calling it practical or pragmatic or whatever.

That might be the thing that achieves something very kinder 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 may be 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 at this particular one example of multitask learning where they're literally taking the problem.

Like, I don't know, Charles Darwin start 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 formulate a bunch of different tasks.

It's like over a hundred 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 then trying to figure out, do we add tasks? Do we, like we focus on each individual tasks separately?

In fact, half.

So the, I would say I'll classify Andre Capati's talking 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 it almost 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 software provisioning, but you have, you know, almost breathing down your neck.

And, you know, this involves, you know, human lives.

And so you have to basically just do the systematic engineering and, you know, fine tuning and refinements and try an error and all that stuff.

There's nothing wrong with that. That's, that's called engineering.

That's called, you know, putting technology out in the, in the world.

And you have to kind of ironclad it before, 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, you know, quite a bit upstream of this.

Your play, I don't think about platonic forms. Your platonic because eventually I want the 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 2000, for example, and you ask yourself the question, okay, I want to recognize cars or faces or whatever.

You know, I can use convolutional net, so I can use a more conventional kind of computer vision techniques, you know, using interest point detectors or a sift,

density features and, you know, sticking an SVM on top.

At that time, the data sets were so small that those methods that use more hand engineering work better than com nets.

There was just not enough data for com nets, and com nets were, were a little, 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, data sets became bigger and GPUs became available.

That's what, you know, the two of the main factors that basically made people change their, change their mind.

And you can, 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, two 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, casserole coefficient, so whatever, right.

And same for text, right. You do name identity recognition and then 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 Markov model, there's no language model that is explicit other than, you know,

something that's ingrained in the, 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 hand crafting and more and more learning.

And I think 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, 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 do 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 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, right?

You need to be able to act in a world and see the effect for you to learn causal models of the world.

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 and mirror neurons and all that stuff.

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 juggler 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 at different levels of curiosity, right?

Depending on how they're built, right?

So, you know, cats and rats are incredibly curious.

Dogs know 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.

You know, we haven't answered that question yet.

So, you know, 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 Yosha Benjos 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 orthodox folk speculative hypothesis about consciousness.

So we're talking about the study of a world model.

And I think, you know, our entire prefrontal context basically is the engine for a 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 context.

That engine is configurable to the situation at hand.

So we are burning 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 configure into 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, you know, 10, 10 plays, 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 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 needs to kind of the really curious concept that consciousness is not a consequence of the power of our mind,

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'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 going to 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 nativists 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 very 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 both because it's really easy to generate signals from the retina that actually will train edge detectors.

And again, those are things that can be learned within minutes of opening your eyes.

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.

And 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 clearly, there's running taking place there.

So 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 make humans human or make cats different from dogs, right?

It's the basic drives that are kind of hardwired in our bezoganglia.

I mean, there are people who are working on this kind of stuff that's called intrinsic motivation in the context of reinforcement learning.

So these are objective functions.

Whether 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 in comfort.

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 why the critic, I mean, an example of how the critic might be learned, right?

If I come to you, I reach across the table and I pinch your arm, right?

Complete surprise for you.

You would not have expected this from me.

Yes, right, let's say for the sake of the story.

Okay, your bezelganglia is going to light up because it's going to 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.

If I try again, you're going to recoil and that's your critic, your predictor of your ultimate pain system that predicts that something bad is going to happen and you recoil to avoid it.

So even that can be learned.

That is learned, definitely.

This is what allows you also to define some goals, right?

So the fact that you're a school child who wake up in the morning and you go to school and 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 the philosopher who wrote the book Denial of Death and his idea is that one of the core motivations of human beings is our terror of death, our fear of death.

That's what makes us unique from cats.

Cats are just surviving.

They do not have a deep, like 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.

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?

I don't know at what point.

I mean, it's a question like, at what point do you realize that 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 would kind of think about more like teenage years and later.

We're talking about way earlier.

No, it's like seven or eight years or something like that.

You realize, holy crap, this is like the mystery, the terror.

It's almost like you're a little prey, a little baby deer sitting in the darkness of the jungle of the woods, looking all around you.

There's darkness full of terror.

And 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 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 going to go on forever.

You kind of know you're going to die and it's going to 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 were able to really understand that this life is finite.

That seems important.

There's a bunch of different things there.