The following is a conversation with
Yann LeCun, 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.
And now a quick few second mention of
each sponsor. Check them out in the
description. It's the best way to
support this podcast. First is Public
Goods, an online shop I use for
household products. Second is Indeed, a
hiring website. Third is Roka, my
favorite sunglasses and prescription
glasses. Fourth is NetSuite, business
software for managing HR, financials,
and other details. And fifth is Magic
Spoon, low-carb, keto-friendly cereal.
So, the choice is business, health, or
style.
Choose wisely, my friends.
And now onto the full ads. As always, no
ads in the middle. I try to make this
interesting, but if you skip them,
please still check out our sponsors. I
enjoy their stuff, maybe you will, too.
This show is brought to you by Public
Goods, the one-stop shop for affordable,
sustainable, healthy household products.
I use their hand soap, toothpaste, and
toothbrush. I think I use a bunch of
other stuff, too, but that's that's what
comes to mind.
The products often have this minimalist
black and white design that I just
absolutely find beautiful. I love it. I
love minimalism in design. It doesn't go
over the top. It doesn't doesn't have
all these extra things and features that
you don't need, just the essentials.
I think it's hard to explain.
But there's something about
the
absence
of things that that take up your
attention
that allows you to truly be attentive to
what matters in life.
Anyway, go to publicgoods.com/lex
or use code Lex at checkout to get $15
off your first order, plus you will
receive your choice of either free pack
of bamboo straws
or reusable food storage wraps.
Visit publicgoods.com/lex
or use code Lex at checkout.
This show is also brought to you by
Indeed, a hiring website. I've used them
as part of many hiring efforts I've done
for the teams I've led in the past. Most
of those for was for engineering, for
research efforts.
They have tools like Indeed instant
match giving you quality candidates
whose resumes on Indeed fit your job
description immediately. For the past
few months I've been going through this
process of um building up a team of uh
folks to help me. I've been doing quite
a bit of hiring.
It's a
treacherous
and an exciting process because you get
to meet some friends. So, it's a it's a
beautiful process, but I think it's one
of the most important processes in life.
It's selecting the group of people with
whom you spend your days with.
And so, you got to use the best tools
for the job. Indeed, I think is an
excellent tool.
Right now, you can get a free $75
sponsored job credit to upgrade your job
post at indeed.com/lex.
Terms and conditions apply. Go to
indeed.com/lex.
This show is also brought to you by
Roka, the makers of glasses and
sunglasses that I love wearing for their
design, feel, and
innovation on material, optics, and
grip. Roka was started by two
all-American swimmers from Stanford, and
it was born out of an obsession with
performance. I like the way they feel. I
like the way they look.
Whether I'm uh running like a fast-paced
run, we're talking about 8-minute mile
or faster, or if we're doing a slow pace
run, 9 10-minute mile along the river in
the heat or in the cold, or if I'm just
wearing my suit
out on the town, however that expression
goes, I'm not sure, but they look classy
with a suit, they look badass in running
gear.
It's just uh my go-to sunglasses.
Check them out for both prescription
glasses and sunglasses at roka.com and
enter code Lex to save 20% on your first
order. That's roka.com and enter code
Lex.
This show is also brought to you by
NetSuite. NetSuite allows you to manage
financials,
human resources, inventory, e-commerce,
and uh many more business-related
details all in one place. I'm not sure
why I was doing upspeak on that
sentence.
Maybe because I'm very excited about
NetSuite. Anyway, there's a lot of messy
things you have to get right when
running a company.
If you're a business owner, if you're an
entrepreneur,
if you're a
founder of a startup, this is something
you have to think about. You have to use
the best tools for the job to make sure
all the messy things required to run a
business are taken care of for you so
you can focus on the things that you're
best at, that where your brilliance
shines. If you are starting a business,
I wish you the best of luck. It's a
difficult journey,
but
it's worth it.
Anyway, right now special financing is
back. Head to netsuite.com/lex
to get their one-of-a-kind financing
program. That's netsuite.com/lex.
netsuite.com/lex
This episode is also brought to you by
Magic Spoon.
The OG, not quite OG, but really
old-school sponsor of this podcast that
I love. It's a low-carb, keto-friendly
cereal. It has 0 g of sugar. It's
delicious. I don't say that enough. It's
really is delicious. Given that it's 0 g
of sugar, it's very surprising how
delicious it is. 13 to 14 g of protein,
only 4 net g of carbs, and 140 calories
in in each serving. You can build your
own box or get a variety pack with
available flavors of cocoa, fruity,
frosted, peanut butter, blueberry, and
cinnamon.
Cocoa is my favorite. It's the flavor of
champions. I don't know why I keep
saying that, but it seems to be true.
Anyway, Magic Spoon has a 100% happiness
guarantee. So, if you don't like it,
they will refund it. Who else
will give you a 100% happiness
guarantee?
Go to magicspoon.com/lex
and use code Lex at checkout to save $5
off your order. That's
magicspoon.com/lex
and use code Lex.
This is the Lex Fridman podcast, and
here's my conversation with Yann LeCun.
>> [music]
>> 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.
Uh There is obviously a kind of learning
that humans and animals are
uh 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 efficient.
Supervised learning requires many
samples for learning anything.
And reinforcement learning requires a
ridiculously large number of trial and
errors to for, you know, a system to
learn anything.
Um
And that's why we don't have
self-driving cars.
>> [laughter]
>> That's a big leap from one to the other.
Okay, so that to solve difficult
problems, you have to have a lot of uh
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 uh even with millions of hours
of simulated practice, a self-driving
car can't actually learn to drive itself
properly.
Um and so, obviously, we're missing
something, right? And 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 to be questioned.
So, now you have to ask, you know, 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.
Um and they learn an enormous amount of
background knowledge about the world
that may be the the basis of what we
call common sense.
Uh this type of learning is not learning
a task. It's not being reinforced for
anything. It's just
observing the world and figuring out how
how it works.
Building world models. Learning world
models. Um how do we do this? And how do
we reproduce this in 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 the That's
what you mean about 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 of that basic
information.
And you're saying in 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 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,
you know, 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?
Um but if you are a sort of,
you know, tabular rasa reinforcement
learning system that doesn't have a
model of the world,
uh 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.
>> And is
>> 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 um
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, you know,
analogy of the cake.
>> Yes. [laughter]
>> Uh
the, you know, the 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 scalar. You tell the
machine you did good, you did bad. And
you and you and you only
tell this to the machine once in a
while.
When I say you, it could be the the
universe telling the machine, right? Um
but it's just one scalar. And so, as a
consequence, there is you you cannot
possibly learn something very
complicated without many, many, many
trials where you get many, many
feedbacks
uh of this type. Supervised learning,
you you give a few bits to the machine
uh at every 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 a
video and then stop the video and ask
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. It's a of uh
you know, the of the the future after
the video clip you fed it um 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
>> Sounds dumb, but can you
it's
it is possible that you can 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 a bit short at the
moment. Um so whether this will
take us all the way to, you know,
human-level intelligence or something or
just cat-level intelligence
uh is not clear, but among all the
possible approaches that people have
proposed, I think it's a bit short. So
I think this idea of
an intelligent system filling in the
blanks, either, you know, predicting the
future, inferring the past, filling in
missing information,
uh you know, I'm currently filling the
blank of what is behind your head and
what you what your head looks like and,
you know, from from the back. Uh because
I have, you know, basic knowledge about
how humans are made.
And I don't know if you're going to, you
know, what where you're going to say, at
which point you're going to speak,
whether you're going to move your head
this way or that way, which way you're
going to look, but I know you're not
going to just dematerialize and reappear
3 m
down the hall. Uh you know, because I
know what's possible and what's
impossible uh according to intuitive
physics.
>> But you have a model of what's possible
and impossible and then you'd be very
surprised if it happens and then then
you'll have to reconstruct your model.
>> Right. So that that's the model of the
world. It's what tells you you know what
fills in the blanks. So given your
partial information about the state of
the world given by your perception
your your model of the world fills in
the missing information and that
includes predicting the future,
retrodicting the past,
you know, and 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 uh
predicting
what control decision you make when
you're driving in a lane. You have a
sequence of images from a a vehicle
and then you can 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
uh method for you know, training
machines to do this kind of prediction
or filling in the blanks. So right now
uh 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 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.
That and once you pre-train that
network, you can use the
internal representation learned by it as
input to you know, something that you
train supervised or or whatever. Um
that's been incredibly successful, not
so successful in images although it's
making progress.
And uh and it's it's based on uh sort of
manual data augmentation. Uh 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 That's not
obvious. So, what In your view, what
which is harder
or perhaps are they just the same
problem? When uh
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 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 uh the prediction is never a
single thing cuz the world is not
entirely predictable.
>> Yeah.
>> It may be deterministic or stochastic.
We can get into the philosophical
discussion about it, but uh
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 uh 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,
um etc. And you know, if you are a 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
uh you can have a neural net give a
score for every word in the dictionary.
It's a big, you know, it's a big list of
numbers, you know, maybe 100,000 or so.
You can turn them into a a probability
distribution that gives that tells you
when I say a sentence, you know, the
you know, the cat is chasing the blank
in the kitchen.
You know, there are only a few words
that make sense there. You know, it
could be a mouse, it could be a lizard
spot or you know, something like that,
right?
>> [snorts]
>> Uh
and and if I if I say the the blank is
chasing the blank in the savanna, you
also have a bunch of plausible options
for those two words, right?
Um that that that because you you have
kind of a you know, underlying reality
that you can refer to to sort of fill in
those those blanks.
Um
so
you cannot say for sure in the savanna
if it's a you know, a lion or a cheetah
or whatever. You cannot know if it's a
zebra or gnu or you know, whatever.
Wildebeest, same thing. Um
but uh
but you can represent the uncertainty by
just a a long list of numbers. Now,
if I 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
somewhere representing a sort of
infinite number of plausible
continuations of multiple frames in a
you know, high-dimensional continuous
space. And we just have no idea how to
do this properly.
>> Uh finite high-dimensional.
So, like you cuz
>> It's finite high-dimensional, yes.
>> Just like the words. Like try to get it
to uh
down to
a 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 over every single
possible word for language and it works.
It It like that's a really dumb way to
do it.
Um like there seems to there there seems
to be like there should be some more
compressed representation of the
distribution over 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 you can form a
distribution over it?
>> That's one of the big questions, you
know, how do how do you do that? Right.
I mean what's what's kind of you know,
another thing that that really is
stupid about um I shouldn't say stupid
but like simplistic about current
approaches to self-supervision in in
NLP in text
is that not only do you represent a
joint distribution over words but for
multiple words that are missing those
distributions are essentially
independent of each other.
Um and you know, you don't pay too much
of a price for this. So you So you can't
So that you know, the system you know,
in the the the sentence that I gave
earlier
uh if it gives a a certain probability
for a lion and and cheetah and then a
certain probability for
uh you know, gazelle uh
wildebeest and and and zebra
uh
those two probabilities are independent
of each other.
Uh and it's not the case that those
things are independent. Lions actually
attack like bigger animals than than
cheetahs. So you know, there's a huge
independence hypothesis in in this
process which is not actually true. The
reason for this is that we don't know
how to represent uh
properly distributions over
combinatorial uh sequences of symbols
essentially when they're because the
number grows exponentially with the
length of the of the symbols.
And so we have to use tricks for this
but um those techniques can you know,
get around like don't even deal with it.
So
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
when I switch lion for gazelle uh 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.
Uh or okay, let me just say the crude
version, which is it's just statistics.
It's not intelligence. Uh 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 they're kind of
pointless. Um
so, first of all, it's quite possible
that intelligence is just statistics.
It's just statistics of a particular
kind.
>> Yes.
>> Uh
where
>> the philosophical question. Is Is kind
of is
is intel- Is it possible that
intelligence is just statistics?
>> Yeah.
But, what kind of statistics? So, uh if
you are asking the question
are the model of the world the models of
the world that we learn, um do they have
some notion of causality? Yes.
>> Mhm.
>> So, if the criticism comes from people
who say you know, uh current machine
learning system don't care about
causality, which by the way is wrong. Uh
you know, I agree with them. Yeah, you
should, you know, your model of the
world should have your actions as one of
your
the inputs.
And that will drive you to learn causal
models of the world where you know what
you know, what uh intervention in the
world will cause what result. Or you can
do this by observation of other agents
uh acting in the world and and observing
the effect.
Uh other humans, for example. So, I
think
you know, at some level of description,
uh intelligence is just statistics.
Uh
but that doesn't mean you don't you
don't you know, you won't have models
that have, you know, deep mechanistic
explanation for what goes on. Uh, the
question is how do you learn them?
That's that's the question I'm
interested in.
Uh, because
you know, a lot of people who actually
voiced a criticism
say that those mechanistic model has to
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.
Uh,
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 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 there's Is this serve as a good
analogy?
Like, are we ignoring the uh
the dark matter, the the basic low-level
mechanisms, just like we ignore the way
the operating system works, we're just
using the
uh the the high-level software, we're
ignoring that at the low level, the
neural network might be doing something
like statistics.
Like, I mean, sorry to use this word
probably incorrectly and crudely, but
doing this kind of filling the gap kind
of learning and just kind of updating
the model constantly in order to be able
to support the raw sensory information
mation to predict it and then adjust to
the prediction when it's wrong. But
like, high level when we look at our
brain at the high level, it feels like
we're doing like we're playing chess,
like we're we're like 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
uh uh 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,
um 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.
>> Yeah.
>> And everything the brain does is trying
to predict
uh predict everything from everything
else.
Okay, and that that's really sort of the
underlying principle, if you want, that
uh self-supervised learning is trying to
kind of reproduce this idea of
prediction as kind of an essential
mechanism
of uh task-independent learning, if you
want.
The next step is
what kind of intelligence are you
interested in reproducing? And of
course, you know, we all think about,
you know, trying to reproduce sort of,
you know, 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 in a cat brain.
Um you know, the most intelligent of our
intelligent systems don't don't have as
much common sense as as a house cat.
>> [snorts]
>> So, um how is it that cats learn? And
you know, cats don't do a whole lot of
uh reasoning. They certainly have causal
models. They certainly have
uh because, you know, many cats can
figure out like how they can act on the
world to get what they want. Um they
certainly have uh a fantastic model of
intuitive physics,
uh certainly of the the the dynamics of
their own bodies, but but also of prey
and things like that, right? So,
um they they're they're pretty smart.
They They only do this with about 800
million neurons.
Um we are not anywhere close to
reproducing this kind of uh thing. So,
to some extent, I I could I could say
let's not even worry about like the
high-level cognition
uh and kind of, you know, long-term
planning and reasoning that humans can
do until we figure out like, you know,
can we even reproduce what Cassio is
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 there are three
challenges in the three main challenges
in machine learning. The first one is uh
you know, getting machines to learn to
represent the world um
and I'm proposing self-supervised
learning.
>> [snorts]
>> Uh
the second is
uh getting machines to reason in ways
that are compatible with
essentially gradient-based learning
because this is what deep learning is
all about, really.
Um [snorts]
and the third one is something we have
no idea how to solve, at least I have no
idea how to solve, is uh
can we get machines to learn
hierarchical representations of action
plans?
You know, like, you know, we know how to
train them to learn hierarchical
representations of perception.
>> [snorts]
>> You know, we accomplish on that sensing
like that and so on. But, what about
action plans? Can we get them to
spontaneously learn good hierarchical
representations of actions?
>> Also gradient-based.
>> Yeah, all of that has, you know, needs
to be somewhat differentiable so that
you can apply sort of gradient-based
learning, uh which is really what deep
learning is about.
>> [snorts]
>> So, it's background knowledge, ability
to reason in a way
uh
that's differentiable that
is somehow connected deeply integrated
with our background knowledge or builds
on top of that background knowledge and
then given that background knowledge, be
able to make hierarchical plans.
>> Right.
>> So, if if you take classical optimal
control, there's something in classical
optimal control called uh
model predictive control and it's, you
know, it's been around since the the
early '60s.
NASA uses that to compute trajectories
of rockets. And the basic idea is that
you You a pre- a predictive model
>> [snorts]
>> of the rocket, let's say, or whatever
system you are you intend to control,
>> [snorts]
>> 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 it would be thrust and, you
know, all the controls you can have.
Uh it gives you the state of the system
at time t plus delta t, right? So
basically differential equation,
something like that.
>> [snorts]
>> Um
and if you have this model and you have
this model in the form of some sort of
neural net or some sort of uh set of
formula that you can back propagate
gradient through, you can do what's
called model predictive control
or gradient based uh
model predictive control. So, you have
uh you can unroll that uh
that model in time, you you
you 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 uh
succeeded or matched what you wanted to
do.
Um you know, is the robot harmed? 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. Um and by
back propagation through time, and again
this was invented in the 1960s by
optimal control theories,
uh you can figure out
uh what is the optimal sequence of
actions that will
uh you know, get my system to the the
best final state.
So,
that's a form of reasoning. It's
basically planning, and a lot of
planning uh systems in robotics are
actually based on this.
Uh and uh
uh 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 uh and then, you know, employ
uh 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, you
know, all those things have names, you
know,
uh going back, you know, decades. Um
and so,
if you're not not the you know, in
classical optimal control, the model of
the world is not generally learned.
Uh this, you know, sometimes a few
parameters you have to identify. That's
called systems identification. But, uh
but generally,
the model is
mostly deterministic and mostly built by
hand. So, the big question of AI,
uh 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 its
complexity. So, it's not just trajectory
of a rocket, which you can reduce to
first principles. It's not 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,
um
you know, physical systems that involve
collective phenomena, like water, or or,
you know,
trees and you know, branches in a in a
tree or something, or or uh 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 in these three,
maybe in the in the planning stages,
the 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 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 um
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 uh
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 che- I mean, I I suppose
chess is an analogy.
So, Monte Carlo tree search.
>> Right.
>> It There is a I go, you go, I go, you
go. Like um 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 the anthology of what
defines a car door, this big
philosophical question. But to me it was
interesting cuz 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 going to
signal something. And that's a dance.
It's a
asynchronous
chess game. I don't know.
>> [laughter]
>> So,
I It feels like um
it's not just I mean, I guess you can
integrate all of that into one giant
model. Like the entirety of the the
the these little interactions. Cuz 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
uh because it's continuous, it's
uncertain in a continuous manner.
Uh
>> It doesn't feel more complicated.
>> But it doesn't feel more complicated
because that's what we are 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.
>> Yeah.
>> Um
in fact, that's why we designed it as a
as a game is to be challenging.
>> Mhm.
>> [clears throat]
>> Uh but and if there is something that,
you know, recent progress in the 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 uh
uh you know, the best Go player thought
they were maybe two or three stones
behind, you know, an ideal player that
they would call God.
Uh in fact, no, they 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 know
we're not very good at like doing this
tree exploration. You know, computers
are much better
um 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
you know, I should say
not differentiable in the sense that,
you know, we run backprop through it,
but in the sense that uh our brain has
some mechanism for estimating gradients
of of some kind.
>> Yeah.
>> And that's what, you know, makes us
efficient. So, if you have an agent that
consists of
a a model of the world, which, you know,
in the human brain is basically the
entire front half of your brain.
>> [snorts]
>> Um
an objective function
which
uh in human in in humans is a
combination of two things. There is your
sort of intrinsic motivation module,
which is in the basal ganglia, you know,
at the base of your brain. That's the
thing that measures pain and hunger and
things like that, like immediate um
uh feelings and emotions.
Um and then there is
you know, 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
of a situation will be.
And so, it's 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 your critic, your
uh Uh, you know, your 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 we'd
like an intelligent agent
uh, to do.
>> And the 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 um
logic-based reasoning uh, in your view.
>> I don't know how to make logic-based
reasoning compatible with
uh, efficient learning.
>> Yeah.
>> And Okay. I mean, there is a big
question, perhaps a philosophical
question. I mean, it's not that
philosophical, but
uh, 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.
>> Right?
Uh, so one question we may ask is uh,
is
does learning in the brain minimize an
objective function?
I mean, it could be a,
you know, a composite of multiple
objective functions, but it's still an
objective function.
Uh, second, if it does optimize an
objective function, does it do does it
do it by uh, 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.
>> [snorts]
>> Uh,
cuz 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
uh, you know, you 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 uh,
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 gradients.
>> Is it possible that some kind of
logic-based reasoning emerges in pockets
as a useful, like you said, if the brain
has an objective function, maybe it's a
mechanism for creating objective
functions.
It's It's It's a mechanism for um
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 uh intelligence. So,
first of all, I think the type of
logical reasoning that we think about um
that we are
>> [snorts]
>> you know, maybe stemming from you know,
sort of classical AI of the 1970s and
'80s.
Um
I think humans use that relatively
rarely
and are not particularly good at it.
>> But we judge each other based on our
ability to uh solve those rare problems.
It's called an IQ test.
>> so. Like I'm I'm not very good at chess.
>> Yes, I'm I'm judging you this whole time
because of it.
Well, we we actually
>> With With your with your uh you know,
heritage, I'm sure you're good at chess.
>> [laughter]
>> No, stereotypes. Not all stereotypes are
true.
>> Well, I'm terrible at chess. So, um
you know, but I think perhaps uh another
type of intelligence that I have is this
uh
uh you know, ability of sort of building
models of the world from
uh you know,
reasoning obvious obviously, but also
also data.
Uh and those those models generally are
more kind of analogical, right? So, it's
it's uh it's reasoning by simulation
and but 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 um connect it to a
situation you've encountered before. Uh
and and your reasoning is more,
you know, akin to some sort of internal
simulation. So, you you are 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 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 or nails or whatever.
Um when you are interacting with
someone, you also have a model of that
person and and sort of interact with
that person,
you know, having this model in mind uh
to kind of uh
tell the person what what you think is
useful to them. So, I think this
this ability to construct models of the
world is basically the essence the
essence of intelligence.
And the ability to use it then to
uh plan uh actions that will
uh fulfill a particular uh criterion um
of course is is is necessary as well.
>> So, I'm going to ask you a series of
impossible questions as we keep ask as
I've been doing. So, 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 you know, I think dark matter you
you put a percentage of
uh on it of
of
uh the composition of the universe and
how much of it is dark matter, how much
of it is dark energy. How much
uh information do you think is required
to to to be a house cat?
So, you have to be able to uh 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 uh great
self-awareness of the physics of your of
your own body and the and the world.
How much knowledge is required do you
think to solve it?
>> Um
>> 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
about 800,000 neurons. Uh
800 million neurons, sorry.
>> The representation does.
>> Everything. All knowledge, everything,
right?
Um
you know, it's less than a billion. A
dog is 2 billion, but a cat is less than
1 billion.
>> Right.
And
so multiply that by 1,000 and you get
the number of synapses.
And I think almost all of it is is
learned through this you know a sort of
self-supervised learning although you
know I I think a tiny sliver is learned
through reinforcement learning and
certainly very little through
you know classical supervised learning
although it's not even clear how
supervised learning actually works in
in biological world.
So I think almost all of it is
is self-supervised learning.
>> It's driven
by
the the sort of ingrained objective
functions that a cat or human have at
the base of their brain which kind of
drives their
their behavior. So you know nature tells
us
you're hungry.
It doesn't tell us how to feed a feed
ourselves. That's that's something that
the rest of our brain has to figure out
right?
>> Well it's interesting cuz there might be
more like deeper objective functions
that are lying the whole thing. So hun-
hunger may be some kind of
Now you go to like neurobiology might be
just the brain
um
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. So it
could be just like one really dumb
objective function at the core.
>> But that's how that's how a behavior is
is is driven.
The fact that you know the orbital
ganglia
drive us to do things that are 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 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 of
uh
uh
It's not true of orangutans. Orangutans
are solitary animals. Um they don't seek
the company of others. In fact, they
avoid them.
Uh in fact, they scream at them when
they come too close because they're
territorial.
Uh cuz for for their survival, you know,
uh evolution has figured out that's the
best thing.
I mean, they're occasionally social, of
course, for, you know, um
reproduction and stuff like that. But um
but but they're mostly solitary. So,
uh 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 And they get to be really smart in
less than a less than a year, in like
half a year.
Um
you know, in a year or they're adults.
In 2 years, they're dead. So,
um
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 for 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
um
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
uh to kind of,
you know, get it to do the thing and
learn the appropriate things, right? But
uh
um like for example, you know, uh
baby humans are are driven to
learn to stand up and walk.
Okay, you know, it's not that's kind of
this desire is hardwired. How to do it
precisely is not. That's learned.
>> Mhm.
>> But the desire to
>> To walk?
>> move around and stand up.
Um that's sort of
hardwired. It's very simple to hardwire
this kind of stuff.
>> What 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 I think it was probably socially
imposed by society that you need to
walk. All the other bipedal
>> There are 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 cuz you suck at bipedal walking at
first. It seems crawling is much safer.
Much more
like why are you in a hurry?
>> [laughter]
>> Well, because because you have this
thing that drives you to do it, you
know? Um
which is sort of part of uh
the sort of
human development.
>> Is that understood actually what
>> Not entirely. No.
>> What is the What's the reason to get on
two feet? It's really hard. Like most
animals don't get on two feet.
>> They get on four feet. You know, many
mammals get on four feet.
>> Yeah, they feel
>> 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.
>> It Well, technically we can go into
anthology. They have four. I guess they
have two feet.
>> They have two feet.
>> Chickens.
>> You know, dinosaurs had two feet. Many
of them.
>> Allegedly.
>> [laughter]
>> I'm just now learning that T-Rex was
eating grass, not other animals. T-Rex
might have been a friendly
friendly pet. What do you think about uh
I don't know if you looked at
the test for general intelligence that
François Chollet put together. I don't
know if you got a chance to look at that
kind of thing. Like what
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 uh 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 uh you know, to
recognize handwritten digits on top of
this will only require a few samples.
And we observe this in in humans, right?
You you show a a young child a a picture
book with a couple picture of an
elephant and that's it. The
The child knows what an elephant is. And
we we we see this today with practical
systems that we you know, we train
[snorts] image recognition systems with
uh enormous amounts of of images either
either completely self-supervised or
very weakly supervised. For example, you
can uh
train a neural net to predict uh
whatever hashtag people type on
Instagram, right? And you can do this
with billions of images cuz there's
billions per day that are showing up. Um
so, the the amount of training data
there is essentially unlimited. And
[snorts] then you take the output
representation, you know, a couple
layers down from the output uh 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? [clears throat] Uh or
weakly supervised uh transfer learning.
Uh people are making very very fast
progress using self-supervised learning
uh for for the with this kind of
scenario as well.
>> [snorts]
>> Um and you know, my guess is that that's
that's going to be the the future.
>> For self-supervised learning, how much
cleaning do you think is needed for
filtering um
uh malicious signal or what's a better
term? But like a lot of people use
hashtags on Instagram
to uh get like good SEO
>> Mhm.
>> 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 kind of
Why would you put science? That's not
very good SEO.
>> The way the way my my colleagues who
worked on this project at
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
or situations. Like you know, that has
some visual content.
Um
so you know
you wouldn't have like hash TBT or
anything like that.
>> Oh, so they they keep a very select set
of hashtags is what you're saying.
>> Yeah.
>> Okay.
>> But it's still 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 contrastive
learning
for
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 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, etc.
>> Add noise.
>> Add noise, etc. And it 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 years because
a lot of self-supervised learning
techniques to pre-train vision systems
are based on data augmentation. And the
the basic techniques is originally
inspired by
techniques that I worked on in the early
'90s and Jeff Hinton worked on also in
the early '90s. They were sort of
parallel work. I used to call this
Siamese network. So, basically you take
uh two identical copies of the same
network. They share the same weights.
And you show two uh 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,
uh 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 be basically learn a function
that will that will be invariant, that
will not change, whose output will not
change when you transform those inputs
uh in in those in those particular ways,
right?
>> [snorts]
>> Um 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?
>> Mhm.
>> Because if you don't have a specific
provision for this,
uh the system will just ignore the
inputs uh 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?
>> Yes.
>> That's called a collapse.
>> Uh-huh.
>> Now, how do you avoid collapse? So,
there's two ideas.
Uh one idea that I proposed in the early
'90s uh with my colleagues at Bell Labs,
uh Jan Bromley and a couple other
people,
uh 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 uh
those two copies, and then you you push
the two output vectors away from each
other.
>> Mhm.
>> And that would eventually guarantee that
things that are semantically similar
produce similar representations, and
things that are different produce
different representations.
>> Mhm.
>> Uh we actually came up with this idea
for a project of doing uh signature
verification. So, we would collect
signature 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 for different signatures.
This was actually the the problem was
proposed by people from a what was a
subsidiary of AT&T at the time called
NCR.
And they were interested in storing
representation of the signature on the
80 bytes of the magnetic strip of a
credit card.
>> [snorts]
>> So, we came up with this idea of having
a neural net with 80 outputs.
You know, that we would quantize on
bytes to so that we could encode the the
>> And then encoding was then used to
compare whether signature matches or
not.
>> That's right. So, then you would you
know sign you 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.
>> [laughter]
>> Um because nobody cares actually. I mean
the
American you know finish payment system
is
incredibly lax in that respect compared
to Europe for example.
>> the signatures. What's the purpose of
signatures anyway? This is very
>> Nobody looks at them. Nobody cares, you
know.
>> And and it's uh yeah.
>> Yeah, no. So,
uh so that that's contrastive running,
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'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 and so you would need lots and lots
and lots of negative pairs.
Uh so, there is a a particular
implementation of this which is
relatively recent from actually the
Google Toronto group
uh where you know, Jeff Hinton is the
the senior member there. It's called
SimCLR, S I M C L C L R.
>> Mhm.
>> Uh and it you know, 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-contrasting methods.
So, other ways to guarantee that
>> [snorts]
>> uh uh the representations will be
different for different uh
different inputs.
And it's actually based on an idea that
um
Jeff Hinton proposed in the early '90s
with his student at the time, Subbarao.
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.
>> [snorts]
>> And you train the two networks to be
informative,
but also to be as informative
uh of each other as possible. So,
basically, one representation has to be
predictable from the other, essentially.
Uh
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 at FAIR.
Uh
particularly a postdoc called Stéphane
Deny, who's now a
junior professor in in Finland at
University of Alto.
Um
we came [snorts] up with something
called uh that we called Barlow Twins.
And it's a a particular way of
maximizing the information content of uh
of a vector, you know, using some
uh hypotheses.
And we have uh kind of a
another version of it that's more recent
now called VICReg, VICREG. That that
means variance invariance covariance
regularization. it's the thing I'm the
most excited about in in machine
learning in the last 15 years. I mean,
I'm I'm not I'm really really excited
about this.
>> What uh
kind of data augmentation is useful for
that non-contrastive 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 a important
step, and it's also an annoying step
because you need to have that knowledge
of what the augmentation
uh
you can do that do not change the nature
of the of the object.
>> Yeah.
>> And so um the standard scenario which
you know a lot of people working in this
area are using is you use uh
the the type of distortion So so
basically you do uh geometric
distortion. So one basically just shift
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 kind of do a shift
in color balance or something like that.
Um saturation another one sort of blurs
it. Another one adds noise. So you have
like a a catalog of kind of standard
things and people try to use the same
ones for different algorithms so that
they can compare.
Um but some algorithms uh some
self-supervised algorithm actually can
deal with much bigger
like more aggressive data augmentation
and some can't don't. So that kind of
makes the whole thing difficult. But but
that's the kind of distortions we're
talking about and
and so you you you you train with those
distortions. And then uh
you you chop off the last layer or
couple layers of the of the the network.
And you use the representation as input
to a classifier. You train the
classifier.
>> Mhm.
>> [clears throat]
>> Um
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.
>> [laughter]
>> And uh as a consequence you cannot use
those the representations in those
systems for things like object detection
and localization because that
information is gone.
>> [snorts]
>> So the type of data augmentation you
need to do depends on the task you want
eventually
the system to 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
>> Okay, help me out and
what uh localization is So, you're
saying it's just not good at the
negative
uh like at cla- 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 you you give it an image
and then you give it the same image
shifted and scaled. You tell it that's
the same image.
>> Yeah.
>> So, the system basically is trying to
eliminate the information about position
and size.
So, now And now you want to use that
>> Oh, like literally
>> out where an object is and what size it
is.
>> bounding box. It did be able to actually
Okay. It can still find 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, there that's an interesting
sort of philosophical question. How
important How important is object
localization anyway? We're 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 an
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
um recognizing the nature of of a scene
or an object
and localizing objects. So, you use the
first pathway called the ventral pathway
for, you know, telling what you're
looking at.
>> [snorts]
>> Uh the other pathway, 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
>> [snorts]
>> localization and detection.
>> Is similarity learning or contrastive
learning or 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
Does 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
uh use techniques of that type, perhaps
very different, but you know, of the
same nature,
to uh train a system from from video to
do video prediction, essentially.
Um
I think we'll have a path, um you know,
towards
I you know, I wouldn't say unlimited,
but but a path towards some level of
uh
you know, physical common sense in in
machines. And I also think that
>> [snorts]
>> um
that
ability to learn how the world works
from a sort of high-throughput channel
like like vision
>> Mhm.
>> is a necessary step towards uh
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.
>> Mhm.
>> So, for example, let let's uh and and
you know, people have attempted to do
this for for 30 years, right? The The
Cyc project and things like that, right?
Of basically kind of writing down all
the facts that are known and
>> Mhm.
>> uh hoping that some some sort of common
sense will emerge. Um I think it's
basically hopeless. Uh but let me take
an example. You take an object, I 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.
Um 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
um whatever it is, it's never going to
learn about this.
Um that information is just is not
present in any text.
>> Well, the the question like with the Cyc
project, the dream I think is to have
like
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 uh
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? A cat knows that.
>> No, but they evolved. So
>> No, they learn like us.
>> The the sorry, the physics of stuff?
>> Yeah.
>> [sighs and gasps]
>> Well, yeah, so you're saying it's
uh so you're putting a lot of
intelligence onto the nurture side, not
the nature.
>> Yes.
>> Cuz we've seen 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,
uh
the question is how Okay, so
the question is how fundamental is that
that the nature of the whole hardware
uh 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
uh big beautiful sexy background model
that you're talking about just by
sitting there.
Um
then okay, then you need to then like
maybe
uh it is all supervised learning all the
way down.
It's all supervised learning, say.
>> Whatever it is that makes
uh 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.
>> Yeah.
>> Uh and
>> Yeah, it it
>> you know, it it and it only involves,
you know,
less than 1% of our genome, right? Which
is the difference between human genome
and chimps or whatever.
So,
uh
it can't be that complicated, you know,
it can't be that fundamental. I mean,
the most of the so complicated stuff
already exists in cats and dogs and you
know, certainly primates, non-human
primates.
>> Yeah, that little thing with humans
might be just a something about social
interaction and ability to maintain
ideas across like a collective of
people.
It it's it sounds very dramatic and very
impressive, but it probably isn't
mechanistically speaking.
>> It is, but we're not there yet. Like, we
you know, we 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 is is
is number one. What do you um
Just a quick
tangent on data augmentation. So, a lot
of it is
uh 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 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.
>> Huh.
>> Um so, what people are working on now is
is two things. One is
uh the type of self-supervised learning
like trying to to translate the type of
self-supervised learning people use in
language, translating this to images,
which is basically a denoising
autoencoder method. Right? So, you you
take an image, you you block you mask
some parts of it.
Uh and then you you train some giant
neural net to reconstruct the parts that
you that are that are missing. And
until very recently
there was no 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 at Menlo Park that
actually works very well. So,
that doesn't require the data
augmentation, that requires only
masking.
>> Okay. Only masking for images
Okay. [snorts]
>> 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 you can
the transformer represents the image as
uh non-overlapping patches, so it's easy
to mask patches and things like that.
>> Okay, then my question transfers to that
problem then masking. Like why should
the mask be a square or rectangle?
>> Yeah, so it doesn't matter. Like you
know, I think we're going to come up
probably in the future with sort of uh
you know, ways to mask that are it's 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
an data augmentation or masking
is 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
Boltzmann machine training. You
you know, every every every time you see
a percept you also
you you can you can perturb it in some
way and then
the the principle of the training
procedure
is to minimize the difference of the
output of the representation between the
the clean version and the corrupted
version essentially, right? And you you
you can do this in real time, right? So,
you know, Boltzmann machine works like
this, right? You you you show a percept,
uh you tell the machine that's a good
combination of activities of your input
neurons.
Uh and then you either uh
let them
go their merry way without
clamping them to values or you only do
this with a subset.
>> Yeah.
>> [snorts]
>> And what you do is you're training the
system so that the the stable state of
the entire network is the same
regardless of whether it's his entire
input or whether it's his only part of
it.
Um you know, the noise in autoencoder
method is basically the same thing,
right? You you're you're training a
system to reproduce the input the
complete input and filling the blanks
regardless of which which parts are 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
uh
oscillate, right? So, they
they take their activity and then
temporarily they kind of shut off
to um you know, force the rest of the
system to basically
reconstruct the input without their
help, you know? And and and
I mean, you you could imagine, you know,
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 do as
much as you want.
>> Yeah.
>> And and get better over time.
Cuz I was thinking like you might want
to be clever about the way you do all of
these procedures,
you know, but that's only if 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.
This data augmentation by waiting, which
is, you know, the the sort of video
prediction.
>> Right.
>> Um you're
observing a video clip observing the
you know the
continuation of that video clip and try
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 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 it the
amount of data is not the constraint. I
mean
>> it would require some selection I think.
>> Some actually
>> Some some selection of maybe the right
type of data you need some
>> the rabbit hole of just cat videos that
might you might need to watch some
lectures or something.
>> No, you
>> How matter would that be
if it like watches lectures about
intelligence and then learns watches
your lectures and NYU and learns
[clears throat] from that how to be
intelligent.
>> I don't think it would be enough.
>> [laughter]
[gasps]
>> What's your do you find multimodal
learning interesting we've been talking
about visual language like if combining
those together maybe audio all those
kinds of things.
>> 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 relative 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 or useless or
easy
once we figured out how to do
self-supervised representation learning
or 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 the next
stage. But of course there is like a
huge amount of you know, very
interesting work to do in sort of
practical questions that that have you
know, short-term impact.
>> Well, you know,
it's it's difficult to talk about that
the temporal scale because all of human
civilization will eventually be
destroyed because the the the 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 um
>> That's going to take a while though.
>> So but 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 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
uh 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 got any
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'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 then trying to
figure out do we add task? Do we like we
focus on each individual task
separately? In fact, half So the I would
say I'll classify Andrej Karpathy's talk
into it. 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 you have to have like a a giant
suite of benchmarks to understand how
well your your system actually works.
>> with him. I mean, he's
he's he's very sensible guy. Um
Now,
okay, it's it's very clear that if
you're faced with a 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 you you might think about the
the fact that the the the right thing to
do and the long-term solution involves,
you know, some fancy
self-supervisioning, but you have
you know, Elon Musk breathing down your
neck, uh and, you know, this involves
uh you know, human lives, and so you you
have to
basically just do the systematic
engineering and, you know,
uh
fine-tuning and refinements and trial
and error and and and all that stuff.
Um there's nothing wrong with that.
That's that's called engineering. That's
called, you know, uh
putting technology out
uh in the
yeah, in the world, um and and you have
to kind of iron clad it before before
you do this, you know, um
uh
so much for, you know, grand grand ideas
and principles.
Um
But, you know what? I'm placing myself
sort of, you know, some
you know, upstream of this, you know,
quite a bit upstream of this.
>> You're Plato, thinking about Platonic
forms. You're You're
>> No, Platonic because eventually I want
us stuff to get used, but uh it's okay
if it takes 5 or 10 years for the
community to realize this is the right
thing to do. I've I've done this before.
It's been the case before that, you You
I've made that 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."
>> [snorts]
>> Um
you know, I can use convolutional nets,
or I can use some more conventional uh
kind of computer vision techniques, you
know, using interest point detectors or
SIFT dense features, and you know,
sticking an SVM on top. At that time,
the data sets were so small that
those methods that used more engineering
worked better than ConvNets. There was
just not enough data for ConvNets. And
ConvNets 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 uh
basically when, you know,
data sets became bigger and and GPUs
became available. That that's what, you
know, the
two of the main factors that basically
made people change their change their
mind.
Um
>> [snorts]
>> 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.
Um you know,
be it
optical character recognition, speech
recognition,
computer vision, like image recognition
in general,
uh natural language understanding, like,
you know, translation, things like that,
right? You start to engineer the hell
out of it.
Um you start to acquire all the
knowledge that 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, Zernike moments,
you know, whatever, right? People have
come up with thousands of ways of
representing images so that you 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 preprocess uh,
uh, speech signals 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.
Um,
you know, [snorts] catch all
coefficients or whatever, right? Um,
and same for for text, right? Uh, you do
named entity recognition and you parse
and you you you do uh, tagging of uh, of
of the parts of speech and, you know,
you do this sort of tree representation
of uh, clauses and all that stuff,
right? Before you can do anything.
Um,
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 hand crafted 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 and, you know, speech
recognition systems nowadays,
uh, or OCR systems are completely end to
end. It's the uh, 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
you know, 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.
Um, and uh,
I I think uh,
I mean, it's true in biology as well.
>> So, I mean, we might disagree about
this, maybe not uh, in the 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 to form the self-supervised
learning,
well, you can't just give it a giant sea
of data.
>> Okay, I agree and I disagree.
>> Okay.
>> I agree in the sense that I think uh
I agree I agree in two ways. The first
The first way I agree is that if you
want uh
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 be to learn causal models of
the world.
>> Well, so that's not 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
hardwire that part, right? And you know,
mirror neurons and all that stuff,
right? So,
um
and it's not clear to me how you would
do this in a machine. So,
um
um so, I think the the action part would
be necessary for having causal models of
of the world.
Uh the second reason it may be necessary
or at least more efficient is that uh
active learning basically, you know,
goes for the jugular of what you're what
you don't know, right? Mhm. It's It's
you know, obvious areas of uncertainty
uh about your your world and about the
how the world behaves.
And you can resolve this uncertainty by
systematic exploration of that part that
you don't you don't know. And if you
know that you don't know, then, you
know, it makes you curious. You kind of
look into situations that And
you know, across the animal world,
different species have different levels
of curiosity, right? [laughter]
Depending on how they're built, right?
So,
you know, [snorts] 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 would be useful, but curiosity
just makes the process faster. It
doesn't make the process exist.
The So, what process what learning
process
is it that
active learning makes more efficient?
And I'm asking that first question, you
know,
>> [snorts]
>> you know, we haven't answered that
question yet. So, you know, I'll worry
about active learning once this question
is uh
>> So, that'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? Like
>> That's true.
>> But fundamentally it's still the same
thing and building up the intuition
about how to
in a self-supervised way to construct
background models, efficient or
inefficient,
is
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
uh
>> That'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
uh 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 And so, on your retina, the
image that forms is an the 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 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 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.
Um
I have [snorts] kind of a
unorthodox folk
uh
speculative
hypothesis about consciousness. So,
we're talking about this idea of world
model.
>> Mhm.
>> [clears throat]
>> And uh I think, you know, our entire
prefrontal cortex basically is
uh the engine for our world model.
Uh 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.
Uh that engine is configurable to the
situation at hand. So, we're building a
box out of wood, or we are, you know,
driving uh
down the highway playing chess, we we we
basically have uh 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,
um
it 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 uh chess
grandmaster,
uh
you know, I get wiped out in, you know,
10
10 plays, right? Uh
and, you know, I have to think about my
move for,
you know, like 15 minutes.
Uh
and the person in front of me, the
grandmaster, you know, would just like
react within seconds, right? Uh,
you know, he doesn't need to think about
it. That's become part of his
subconscious because, you know, it's
basically just
pattern recognition at this point.
>> [snorts]
>> Um, same, you know, you the first few
hours you drive a car, you are 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.
Um, 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
uh, executive kind of overseer that
configures your world model for the
situation at hand.
And that that leads to the kind of the
really curious concept that uh,
consciousness is not a consequence of
the power of our minds, but of the
limitation of our brains.
That because we have only one world
model, we have to be conscious. If we
had as many world models as
there are as 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 the hard problem of
consciousness, there's some kind of
chemicals and biology that's creating a
feeling
like it feels to experience some of
these things.
That that's kind of like the hard
question is
what the heck [clears throat] 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.
Um, yeah, it could be just a very nice
side effect of um,
of the way we evolved that's just very
um, useful to
to uh, 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, depends who you talk about, but I
think
So certainly there is uh
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 in our minds.
Um things like you know, the world is
three-dimensional for example. Is that
hardwired?
Things like uh you know, object
permanence. Is it something that we
learn uh
you know, before the age of 3 months or
so? Or or are we born with it? And there
are you know, very you know,
wide disagreement among the you know,
the cognitive scientists uh
for this.
I think those things are actually very
simple to learn.
Um
you know, 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
are actually will train edge detectors.
So
Um and again, those are
things that can be learned within
minutes of
opening your eyes, right? I mean uh
>> [snorts]
>> you know, since the 1990s we have uh
algorithms that can learn oriented edge
detectors completely unsupervised with
the equivalent of a few minutes of real
time. So uh so those things have to be
learned.
Um and there's also those you know, MIT
experiments where you kind of
plug the
optical nerve on the auditory cortex of
a baby ferret, right? And that auditory
cortex become a visual cortex
essentially.
So you know, clearly um there's learning
taking place there.
Um 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're easy to learn.
>> Do you 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. They They They are the things
that, you know, make humans
uh human or make, you know,
cats different from dogs, right? It's
the the basic drives that are kind of
hardwired in our
basal ganglia. Um
I mean, there are people who are working
on on this kind of stuff. It's called
intrinsic motivation in the context of
reinforcement learning. Um 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
uh
comfort or discomfort.
Uh and and because it's your brain
computing this, presumably it knows also
how to estimate gradient and so that's
right. So, um
uh so, it's easier to to learn when your
objective is is intrinsic.
Um so, that has to be hardwired.
Uh the critic that makes long-term
prediction of the outcome, which is the
eventual
result of this, that's learned.
Uh and perception is learned and your
model of the world is learned. But, let
me take take an example of, you know,
why the critic uh I mean, example of why
the critic may be learned, right? If I
uh if I come to you um
you know, I reach across the table and I
I pinch your arm, right? Complete
surprise for you.
You would not have expected this from
>> expecting that the whole time, but yes,
right. [laughter] Let's say for the the
sake of the story, yes.
>> Um okay, your basal ganglia is going to
light up cuz it's going to hurt, right?
Uh
>> [snorts]
>> and now your model of the world includes
the fact that I may pinch you if I
approach my uh
my uh
>> Don't trust humans. Right. My hand to
your arm. So, if I try again, you're
going to recoil. And that's your critic,
uh your predictive
you know, your predictor of your uh
ultimate pain
uh
uh
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
uh you know, define some goals, right?
So,
um the fact that you know, you're a
school child, you wake up in the morning
and you go to school and you know,
um
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. He's the
philosopher who wrote the book The
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.
Uh that's what makes us unique from
cats. Cats are just surviving. They do
not have a
deep under- like um 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
the basically
this idea that
all of human civilization, everything we
create is kind of uh trying to forget if
even for a brief moment that we're going
to die.
When When do you think humans understand
that they're going to die? Is it learned
early on
also, like
>> I don't know to at what point I mean,
it's a 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?
>> Well, so the the pushback on that what
Ernest Becker says and um 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
uh 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's like 7 or 8 years, something
like that. Yeah.
>> You realize
holy crap,
this is
uh 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, of the woods, looking all
around you. There's darkness full of
terror. I mean, that's that realization
says, "Okay, I'm going to go go back in
the comfort of my mind where there is a
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 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. And you
kind of know you're going to die, yeah,
and it's going to be sad, but you don't
really understand that you're going to
die. And so that's that's their idea.
And if I find that compelling
because it does seem to be a core unique
aspect of human nature that we we're
able to think that we're go we're able
to really understand that this life is
finite. I find that important.
>> This There's a bunch of different things
there. So, first of all, I don't think
there is a qualitative difference
between between us and cats in the term.
I think the difference is that we just
have a better long-term
ability to predict, you know, in the
long term. And so, we have better
understanding of how the world works.
So, we have better understanding of, you
know, finiteness of life and things like
that.
>> So, we have a better planning engine
than cats?
>> Yeah.
>> Okay.
>> Um you know, perhaps
>> what's the motivation for
>> 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, you know, the essence of
intelligence is the ability to predict
and so the 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.
>> Um you say religion helps with that. I
think religion hurts actually.
Um It makes people worry about like, you
know, what's going to happen after their
death, etc.
If you believe that, you know,
you just don't exist after death. Like,
you know, 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.
>> worry about the about you know, this
life because that's the only one you
have.
>> I think it's well, I don't know I don't
know. If I were to say what Ernest
Becker says then I'll say I agree with
him more
than not is um
you do deeply worry.
Uh if you if you believe there's no God,
there's still a deep worry like 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 in this being and then
>> Science keeps bringing humanity down
from its pedestal.
And that's that's just another 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 what
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
experienced the deepest trauma that it
earlier in life, they often before they
to 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 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.
>> And
>> Uh but that didn't have any impact on my
reflection on that topic.
>> So, I'm I'm basically just playing a bit
of devil's advocate pushing back and
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
uh not just avoid
falling off the roof or something like
that, but
ponder
the the end of the ride.
Uh if you listen to the Stoics,
it's uh it's a great motivator. It adds
a sense of urgency. So, maybe a to truly
fear death or be cognizant of it might
give
uh a deeper meaning and urgency to the
moment to live
fully.
>> I may I may
Maybe I don't disagree with that.
I mean, I think what motivates me here
is uh
you know, knowing more about about human
nature. I mean, I think
human nature and human intelligence is a
big mystery. It's a scientific mystery.
>> Mhm. [snorts]
>> Uh in addition to, you know,
philosophical and etc., but, you know,
I'm a true believer in science, so
um
I and and and I do have kind of a belief
that for complex systems like like the
brain and the mind,
the 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 um I've used this analogy
before with you, I believe. Um
the same way [clears throat] we we only
started to understand uh
aerodynamics when we started building
airplanes and that helped us understand
how birds fly.
Uh, you know, so I think there's kind of
a a similar process here where
we don't have a theory of 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
uh
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, so you're trying to build novel
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 you going to be
like most humans
and be
once again unhappy to be brought down
from a pedestal of consciousness /
intelligence?
>> No, I'm I'll be very happy to
understand
uh
more about uh human nature, human mind,
and human intelligence through the
construction of machines that
uh
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
[laughter] fine with it. That's just
reality of life. Um, so I'm fine with
that. Now, you're asking me about things
that uh opinions I have that a lot of
people may disagree with.
I think
uh
if we think about the design of
autonomous intelligence systems, so
assuming that we are somewhat successful
at some at some level
of getting machines to learn models of
the world, predictive models of the
world, uh we have we build intrinsic
motivation objective functions to drive
the behavior of that system. Uh 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 tried to pinch you,
um
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 uh intrinsic
motivation, by objectives,
if it has
a critic that allows it to predict in
advance whether the outcome of a
of a situation is going to be good or
bad, it's going to have emotions. It's
going to have fear
>> Yes.
>> uh when it predicts that the outcome is
going to is going to be bad
and and something to avoid, it's going
to have elation when it predicts that
it's going to be good.
>> [snorts]
>> Um uh if it has drives to relate with
humans, um
you know, in some ways, the way humans
have,
um
you know, it's it's going to be social,
right? And so it's going to have
emotions about attachment and and things
of that type. So,
um so I think uh you know, the the sort
of
sci-fi
thing where, you know, you see Commander
Data like having an emotion chip that he
can turn off, right? Um 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
um okay, forget the movement, but a
discussion at the Supreme Court
that particular kinds of robots,
you know, particular kinds of systems,
um, deserve the same rights as humans
because they can suffer just as humans
can.
Um, all those kinds of things.
>> Well,
perhaps, perhaps not. Like imagine that
humans were that that you could, uh,
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
>> [laughter]
>> there's always a backup, you can always
restore, maybe like the the importance
of murder will go down one notch.
>> That's right. But also the,
uh, your your
you know, desire to do dangerous things
like, you know,
you know, skydiving or you know, uh,
or or or, you know, race car driving,
you know, car racing or that kind of
stuff, you know, would probably
increase.
Or or, you know, airplane aerobatics or
that kind of stuff, right? You 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
uh, today's technology and you can you
can always have a backup.
>> So, it's possible
I don't know if you like video games,
but there's a there's a game called
Diablo and, um,
>> Oh, my my sons are huge fans of this.
>> Yes. Uh,
and
>> In fact, they made a game that's
inspired by it.
>> Awesome. Like built a game?
>> My my three sons have a a game design
studio between them, yeah.
>> That's awesome.
>> They came out with a game
>> Like just came out with a game?
>> Last year. No, this was last year, early
this year, about a year ago.
>> That's awesome. But so, in Diablo,
there's a something called hardcore mode
which if you die, there's no
you're gone.
>> Right.
>> That's it. And 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 for cuz if 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
possible.
>> but cloning is not copying, right? I
mean, you don't reproduce the the the
mind of the person and experience.
>> It's just a delayed twin, so.
>> But then it's well but we were talking
about with computers that you will be
able to copy. You You will be able to
perfectly save, pickle the the the mind
state. And it's possible that that will
be illegal because that goes against um
that will destroy the motivations of the
system.
>> Okay, so let's say you you have a
domestic robot, okay,
sometime in the future.
>> Yes.
>> And uh
the domestic robot, you know, comes to
you kind of somewhat pre-trained, you
know, it can do a bunch of things, 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 5
years, you've 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
maybe it's a virtual assistant that
lives in your, you know, augmented
reality glasses or whatever, right?
>> Yeah.
>> Uh you know, the the Her movie type
thing, right? Um
and that system to some extent the 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 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
uh live his or her own life and, you
know, the fact that they learned stuff
from you doesn't mean that you have any
ownership of it, right? But if it's a
robot
uh that you've trained, perhaps you have
some
intellectual property claim about
>> Intellectual property? Oh, I thought you
meant like uh permanent value in the
sense that it's part of you is in
>> Well, there is permanent value, right?
So, you would lose a lot if that robot
were to be destroyed and you you had no
backup, you would lose a lot. You know,
you would lose a lot of investment, you
know, kind of like a
uh you know, a person dying, you know?
Um that that uh a friend of a friend of
yours dying or or co-worker or something
like that.
Um [snorts]
>> But also uh you have like intellectual
property rights in the sense that 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.
>> Yes.
>> Or kind of, you know, thinks uh
life with you is sort of untenable or
whatever.
>> Right.
>> Um
>> [snorts]
>> Now, all the all the things that that
system learned from you,
uh
you know, how
can you like, you know, delete all the
personal information that that system
knows about you?
>> Yeah.
>> Uh
I mean, that would be kind of an ethical
question, like, you know, can you erase
the the mind of a of a intelligent robot
uh to protect your 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, that's the
problem with the relationships, you
know? If they you break up, you can't
you can't erase the other human.
With robots, I think it'll have to be
the same thing with robots. That that
risk that
that has to be um,
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 you were.
>> But then, are you allowed to, you know,
murder the robot to protect your private
information if the robot decides to
leave?
>> I have this intuition that for robots
with with certain
like it's it's almost like, uh,
regulation. If you declare your robot to
be, let's call it sentient or something
like that. Like, this 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 you do preserve on the on a on
a hard drive or the equivalent in the
future?
>> That might be illegal. Just like it's
like piracy piracy is illegal.
>> No, but it's your own it's your own
robot, right?
>> But you can't you don't
>> But then but then you can wipe out,
>> [laughter]
>> its brain so the 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 we like
the robots and the humans are the same.
Uh, we can both suffer, we can both
hope. We can both
all of those all those kinds of things.
Raise families, all that kind of stuff.
It's it's, uh, interesting for these
just like you said, emotion seems to be
a fascinatingly powerful aspect of human
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, uh, really good questions.
>> questions, yeah. But you I mean you ask
about you asked about the Chinese room
type argument, you know, is it real if
it looks real?
>> Yeah.
>> I think the Chinese room argument is a
ridiculous one. So
>> So so for people don't know, Chinese
room is
uh you you can 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 um
if you understand how that works, you
have this giant nearly infinite lookup
table, is that really intelligence? Cuz
intelligence seems to be a mechanism
that's much more interesting and complex
than this lookup table.
>> I don't think so. So the I mean the real
question comes down to do you think uh
you know, you can you can make an AI's
uh intelligence in some way, even if if
that involves uh learning.
And the answer is of course yes. There's
no question.
Uh there's a second question then, which
is
uh assuming you can uh reproduce
intelligence in sort of different
hardware than biological hardware, you
know, what I computers,
um
can you
uh you know
match uh human intelligence in
all the domains in which humans are
intelligent?
Is it possible, right? So this is the
quite the hypothesis of strong AI. The
answer to this
in my opinion is a unqualified yes. This
will 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 going to take a long time
uh [snorts] regardless of what, you
know,
Elon and others have claimed or
believed. Um this is a lot a lot harder
than many of many of those guys think it
is.
Um and many of those guys who thought it
was simpler than that years you know, 5
years ago now think it's hard because
it's been 5 years and they realize it's
it's going to take a lot longer.
That includes [snorts] a bunch of people
at DeepMind for example, but um
>> Oh, interesting. I haven't actually uh
touched base with the DeepMind folks.
But some of it Elon or
uh Demis Hassabis, I mean sometimes in
your role you have to kind of create
deadlines that are nearer than farther
away to to kind of create an urgency cuz
you know, 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 it's a
weird You can't be too cynical if you
want to get something done.
>> Absolutely. I I I agree with that, but
um
I mean you have to inspire people,
right? To work on certain ambitious
things.
Um
Uh so, you know, it's it's uh it's
certainly uh a lot harder than we
believe, but there's no question in my
mind that this will this will happen.
And now, you know, people are kind of
worried about what does that mean for
humans? They they are going to be
brought down from their pedestal, you
know, a bunch of notches.
With that and uh
you know, is that going to be good or
bad? I mean, it's just going to give
more power, right? To sort of amplify
your 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. Uh
looking back, what has been the
successes, the failures, the lessons
learned from the 8 years of FAIR? And
maybe you can also give context of
the newly minted Meta AI fit into uh how
does it relate to FAIR?
>> Right. So, let me tell you a little bit
about the organization of all this. Uh
uh Yeah, FAIR was created almost exactly
8 years ago. Uh it wasn't called FAIR
yet. It took that name a few a few
months later.
Um and at the time uh I joined uh
Facebook, um there was a a group called
the AI group that had about 12 engineers
[snorts] and a few sci- a few
scientists, like, you know, 10 engineers
and two scientists or something like
that.
>> Mhm.
>> I run it for 3 and 1/2 years as a
director.
You know, hired the first few scientists
and kind of set up the culture and
organized it, you know, explained to the
Facebook leadership what what
fundamental research was about and how
it can work within
uh
uh industry and how it needs to be open
and everything.
Um
and
uh I think it's been an an qualified
success
in the sense that
FAIR has simultaneously produced um you
know, top-level research and advanced
the the science and the technology,
provided tools, open-source tools like
PyTorch and many others.
Uh but at the same time has had a
a direct or mostly indirect impact on
uh Facebook at the time, now Meta.
>> Mhm.
>> Uh in the sense that a lot of systems
that are that Meta is built around
um now are
are based on
research projects that uh started at uh
at FAIR.
So, if [snorts] you were to take out,
you know, deep learning out of uh
Facebook services now and and Meta more
generally, um I mean, the company would
would literally crumble. I mean, it's
completely built around
uh around AI these days. It's really
essential to the operations.
So, what happened uh after 3 and 1/2
years is that uh
I changed role. I became chief
scientist, so I'm I'm not doing
day-to-day management of uh of FAIR
anymore.
I'm more of a kind of
you know, think about strategy and
things like that and and I carry my I 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.
>> [snorts]
>> Um so, now uh FAIR is run by, uh, Joelle
Pineau and, uh, Antoine Bord
>> [snorts]
>> uh, together because FAIR is kind of
split in two now. There's something
called FAIR Labs, which is sort of
bottom-up scientist-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, Joelle leads FAIR Lab and
Antoine Bord leads FAIR Excel.
>> Where are they located?
>> Uh, it's
decentralized all over. Um,
so there's there's there's no question
that the leadership, uh,
of the company believes that
this was a very worthwhile
investment, and what that means is that
uh,
it's it's there for the long run.
Right? So, there is, uh,
if you if you want to talk in these
terms, which I don't like, this is a
this is a business model, if you want,
where where
uh, FAIR, despite being a very
fundamental research lab, brings a lot
of value to the company, either mostly
indirectly through other groups.
>> Mhm.
>> Um, now, what happened 3 and 1/2 years
ago when I stepped down is, uh, was also
the creation of Facebook AI, which was
basically a larger organization that,
uh,
covers FAIR, so FAIR is included in it,
but also has other organizations that
are
uh,
focused on applied research or advanced
development
uh, of AI technology that is more, you
know, focused on the
>> Right.
>> products of the company.
>> So, less emphasis on fundamental
research.
>> Less fundamental, but it's still a
research. I mean, there's a lot of
papers coming out of those organizations
and uh, the people are some awesome and,
you know, wonderful to interact with,
and but it it serves, uh, as kind of,
uh,
uh, a way to kind of scale up, if you
want, um,
sort of AI technology which, you know,
may be very experimental and and sort of
lab prototypes into things that are
usable.
>> So, FAIR is [snorts] a subset of Meta
AI. Is FAIR become like KFC? It it'll
just keep the F. Nobody cares what the F
stands for.
>> We'll know soon enough by probably
probably by the end of the of 2021.
>> not a giant change, mayor fair.
>> Well, mayor doesn't sound too good, but
you know, the the brand people are kind
of deciding on this. And they've been
hesitating for for a while now, and they
you know, they tell us they're going to
come up with an answer as to whether
fair is going to change name or whether
we're going to 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's a good one.
>> be my preference, you know, I would turn
I would turn the F into fundamental.
>> Oh, that's good.
>> Fundamentally I research.
>> Oh, that's really good. Yeah, yeah.
>> Within meta AI. So, this would be
sort of meta fair.
>> Yeah. But, you know, people will call it
fair, right?
>> Yeah, exactly. I like it.
>> meta AI
uh
is part of the reality lab. So, you
know, uh
uh meta now the new Facebook right is
called meta and it's kind of
divided into
you know, Facebook, Instagram, WhatsApp.
>> [snorts]
>> And
reality lab. And reality lab is about,
you know, AR, VR,
um
you know, telepresence, communication
part of uh technology and stuff like
that. It's kind of the
you can think of it as the sort of a
combination of um
sort of new products and and technology
part of uh
of uh meta.
>> Is that where the touch sensing for
robots? I saw that you were posting
about. That's that's
>> sensing for robot is part of fair,
actually. That's that's a fair part.
>> Okay, cool.
>> Yeah. This is also the No, but there is
the the other way the
the haptic glove, right? That that has
like
>> more reality lab.
>> That's that's reality lab research.
>> Reality lab research. But, by the way,
the touch sensors are super interesting.
Uh like integrating that modality into
the whole
uh sensing uh suite is very interesting.
So,
uh
what do you think about the metaverse?
What do you think about this whole
this whole kind of expansion of the view
of the role of Facebook and Meta in the
world?
>> Well, a metaverse really should be
thought of as the next step in the
internet, right? Sort of
trying to
uh kind of, you know, make the
experience more compelling um of, you
know, being
connected either with other people or
with content.
Um and you know, we are
evolved and trying to
evolve in, you know, 3D environments
where
uh you know, we can see other people, we
can talk to them when you're when when
near them or
you know, and other people who are far
away can't hear us, you know, things
like that, right? So, it it's uh there's
a lot of social conventions that exist
in the real world that we can try to
transpose. Now, what is going to be
eventually the
the uh how compelling is it going to be?
Like are, you know,
uh is it going to be the case that
people are going to be willing to
do this if they have to wear, you know,
a huge pair of goggles all day? Maybe
not. Um
>> 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, technology makes sufficient
progress for that.
Um um you know, AR is a much easier
concept to grasp that you're going to
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 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 in VR.
>> Yeah, but you you you can imagine how,
you know, the metaverse being
uh
>> a mix.
>> A mix, right? Or 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 so far.
>> This was easy.
Uh
the Facebook, now Meta, the social
network has been painted by the media as
a 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 the description, the
company that is being described in the
in some media
uh 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 the company, but
you know, I'm a vice president. I mean,
[snorts] I have a pretty good vision of
what goes on, you know, I don't know
everything obviously. I'm not involved
in in in everything, but certainly not
in decision about like, you know,
content moderation or anything like
this, but
but I have, you know, some decent vision
of what 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 in the world and you know,
the the reason your friend believe crazy
stuff.
Um you know, there's an easy scapegoat,
right? In the
uh in in in in social media in general,
Facebook in particular.
But you have to look at the data. Like,
is it the case
that uh Facebook, for example, uh
polarizes people politically?
Um are there academic studies that show
this?
Is it the case that uh you know,
teenagers
uh think of themselves less if they use
Instagram more?
Is it the case that uh
you know, people get more riled up
against uh
you know, opposite sides in a in a
debate or political opinion if they if
they are more on Facebook or if they are
less?
And uh study after study
show that none of this is true.
This is independent studies by academic
that are not funded by Facebook or Meta.
Um, you know, studied by Stanford, by
some of my colleagues at NYU actually
with whom I have no connection.
Um, you know, there's a study recently
they they they paid people, I think it
was in um
in in in the former Yugoslavia, I'm not
exactly sure in what what part, but they
paid people to not use Facebook for a
while in the
period um uh
before the anniversary of the Srebrenica
>> Mhm.
>> 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 [snorts] a bunch of people
to not use Facebook for a few weeks.
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.
Um, there's a study, you know, from
Stanford of
uh economists at Stanford that tried 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.
>> Uh continuously. And um and uh so the 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
uh evolution by some measure of uh
polarization in the US.
And then you compare this with other
countries like
uh
>> [snorts]
>> uh the the the west half of Germany
because you can go 40 years in the east
east side or Denmark or or other
countries.
Uh 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, uh a
causal relationship there
um
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 riled me up is that
people now are accusing Facebook of bad
deeds that are done by others and those
others are we're not doing anything
about them.
>> Mhm.
>> 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
Sheryl Mike Sheryl for on this podcast
and also Mark Zuckerberg and probably
these are conversations you can have
with them.
Uh cuz it's very interesting to me even
if Facebook has an
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's a question. You can't just say like
uh there's an increase in division. Yes,
probably Google search engine has
created increase in division. We 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, we have to
look at the full context of the world
and did it make a better world? You have
to
>> the printing press has created more
division, right?
>> Exactly.
>> So, you know, when the
printing press was invented,
the first books that were 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, you know, the Protestant
movement and 200 years of religious
persecution and wars. So, that's a bad
side effect of the printing press. You
know, social networks aren't being
nearly as bad as the printing press, but
nobody would say the printing press was
a bad idea.
>> [sighs and gasps]
>> Yeah, a lot of it is perception and
there's a lot of different incentives
operating here.
Um maybe a quick comment since you're
one of the top leaders at Facebook and
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, the I
mean, it's just a huge amount.
Uh maybe can you give me context about
how much of Sherp'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 as Zuckerberg, Mark
Zuckerberg.
>> They really focus on AI. I mean,
certainly, uh,
in the, uh,
in the run-up of the creation of Fair
and for, you know, at least a year after
that, if not more,
uh, Mark was was very, very much focused
on on AI and was spending quite a lot of
effort, um, on it. And that's his style.
When he gets interested in something, he
reads everything about it.
You know, he read some of my papers, for
example, before he joined. Um,
um, and, uh, so he he learned a lot
about it.
>> notes.
>> [laughter]
>> Right? Um,
>> [snorts]
>> and, uh, uh,
you know, Sherp
was really tweeting also. I mean, Sherp
is really kind of,
um,
you know, has
something I've tried to preserve also,
despite my, uh, not so young age, which
is a a sense of wonder about, uh,
science and technology. And he certainly
he certainly has that.
Um, he's also a wonderful person. I
mean, in terms of, uh, like, uh,
as a manager, like, dealing with people
and everything. Mark also, actually.
Um, so I mean, they're very like, you
know, very human people.
For in the case of Mark, it's, uh,
shockingly human, you know, given his,
uh,
his trajectory.
Uh,
I mean, the personality of of him that
is painted in the press is just
completely wrong.
>> Yeah. But you have to know how to play
the press. So that's
I I put some of that responsibility on
him, too. You have to
It's like, um,
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
is a depth and kindness to you.
>> And he's probably And he's probably not
the best at it. So, yeah.
>> [laughter]
>> You have to learn. And it's it's sad to
see I'll talk to him about it, but the
strap 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 up.
>> You know, I think he's I think he's done
the thing he
set out to do and, you know, he's he's
got, you know,
uh
family priorities and stuff like that
and
um
I I understand, you know, after 13 years
or something.
>> It's been a good run.
>> Which in Silicon Valley is basically a
[laughter] lifetime.
>> Yeah.
>> You know, cuz you know, it's dog years.
>> So, NeurIPS the conference just wrapped
up.
Let [clears throat] me just go back to
something else. You posted a paper you
co-authored was rejected from NeurIPS.
As you said, proudly in quotes,
rejected.
>> [laughter]
>> No joke.
>> Yeah, I know.
Uh can you describe this paper and like
what what 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 talk about
the review process afterwards.
Um the paper is called VICReg. So, this
is I mentioned that before,
variance-invariance covariance
regularization. And it's a technique in
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 uh
uh
Let me back up a little bit, right? So,
if you want to do self-supervised
learning,
uh you can you can do it by prediction.
Um so, let's say you want to train a
system to predict video, right? You show
it a video clip and uh and you train the
system to predict the next the
continuation of that video clip. Now,
because you need to handle uncertainty,
cuz there There of you, 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.
>> [snorts]
>> 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.
>> Mhm.
>> Okay. So, [clears throat] that's called
I call this a
a generative latent variable model.
>> Got it.
>> Okay? Now, there is an alternative to
this to handle uncertainty.
And instead of directly predicting the
the next frames of the of the of the
clip,
you also run those
through another neural net.
Uh so, that you now have two neural
nets. One that looks at the
the
you know, the initial segment of the
video clip. And another one that looks
at the the continuation during training,
right?
>> [snorts]
>> And what 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.
Uh what you want is
uh informative representative represent
you know, informative representations of
the two video clips that are mutually
predictable.
>> Mhm.
>> [clears throat]
>> What that means is that there's a lot of
details in the second video clip that
are irrelevant.
Uh
you know, um
I let's say video clip consists in, you
know, a camera panning the scene.
Uh there's going to be a piece of that
room that is going to be revealed, and I
can somewhat predict what the 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 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
uh whenever the the the the continuation
video clip varies over all the possible
continuations,
>> Mhm.
>> 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.
>> Yeah.
>> So, 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.
Uh
I used to be a big fan of the first
approach, and in fact, in this paper
with Ishan Mishra, this this blog post,
>> Mhm.
>> uh the dark matter intelligence, I was
kind of advocating for this. And in the
last [clears throat]
year and a half, I've completely changed
my mind. I'm now a big fan of the second
one.
>> Mhm.
>> Uh and it's because of
uh
uh a small collection of algorithms that
have been proposed over the last uh
year and a half or so, 2 years,
uh to do this, uh including VICReg, uh
its predecessor called Barlow Twins,
which I mentioned,
uh
a method from our friends at DeepMind
called BYOL,
um and and and there's a bunch of others
now that kind of work similarly. So,
they're 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 like BYOL work, but we don't
really know why. And there's been like
lots of theoretical papers about why
Barlow Twins works. No, it's not that
because we take it out and it still
works and
you know, blah blah blah. I mean, so
there's like a big a big debate, but um
uh 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.
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 uh
predictive world models
and at the same time learn how I pick up
representations of the world where what
matters about the world is preserved and
what is irrelevant is eliminated.
>> Uh by the way, the representations
before and after is
in the space in a sequence of images or
is it for single images?
>> Uh it would be either for a single
image, for a sequence. It could 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 at
you know, I'm looking for
methods that are generally applicable,
that are not specific to you know, one
particular modality. You know, it could
be audio or whatever.
>> Got it. So, what's the story behind this
paper? This This paper is what? Is Is
describing one of the one such method?
>> This this VicReg method. So, this is
co-authored uh the first author is uh a
student uh called Adrien Bard, who is a
resident PhD student at Fair Paris.
He's co-advised by me and Jean Ponce, um
who's a professor at École Normale
Supérieure, also uh
research director at Inria.
Um so, this is wonderful program in
France where PhD students can basically
do their PhD in industry, and that's
kind of what what's happening here.
Um and this paper is a follow-up on uh
the the this Barlow Twins paper by yeah,
my former postdoc now Stefano De uh with
Li Jing and and Yurish Montal and bunch
of other people from from from Fair.
And one of the main criticism from
reviewers is that VicReg is not
different enough from Barlow Twins. But,
you know, my impression is that
it's, you know,
bottom twins with a few bugs fixed
essentially. And
in the end, this is what people will
use.
>> Right. So,
>> But, you know, I'm used to stuff
>> Yeah.
>> that I submit being rejected for a
while.
>> might be rejected and actually accepted
and well-cited cuz people use it.
>> Well, it's already decided like a bunch
of times.
>> So, I mean, the the 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 the journals, but it's
accelerated significantly. Well, not
significantly, but it goes fast.
It's a nice way to get stuff out
quickly,
to peer review 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 cuz it's a
limited number of people look at it.
There's bias in the following like
if you if you want to do new ideas,
you're going to get pushback.
Um there's self-interested people that
kind of
can infer who submitted it
and kind of
you know, be cranky about it, all that
kind of stuff.
>> Yeah, I mean, there's a lot of, you
know, social [snorts] phenomena there.
Um 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 start saturating, you you
have less of that problem
of reviewers being very uh
inexperienced.
A consequence of this is that, you know,
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 that task as finding flaws
in papers and most papers have flaws,
even the good ones.
So it's it's easy to you know
to do that. You you 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
>> [snorts]
>> even if they make it better, you know,
eventually. I think that's
that's really what 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 these numbers on uh
on benchmarks
and and it's much easier to get a paper
accepted if it brings an incremental
improvement on a
sort of mainstream well-accepted
uh method
or problem.
And those are to me boring papers. I
mean they're not useless, right? Because
industry, you know, strives on 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 I'm a huge believer that
review should be done by thousands of
people, not two people.
>> 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.
>> not a question of being elitist or not.
It's a question of being basically
recommendation and sort 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.
Um
that saves you time cuz you know you
don't have to like scrutinize
the paper as much, you know, it is
brought to your attention. I mean it's
the whole idea of sort of you know
collective recommendation system, right?
So I actually thought about this a lot
um
you know, about 10 15 years ago. Of
course the what discussions at
NIPS and you know and we're about to
create ICLR with Yoshua Bengio and so I
wrote a document
um kind of describing a reviewing system
which basically was you know, you post
your paper on some repository, let's say
archive or now could be open review. Um
and then you can form
a reviewing entity which is equivalent
to a reviewing board, you know, of a
journal or
um
program committee of a conference. Um
you have to list the members
and then um
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.
Um
or may choose not to.
Um and then you know, given 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 the reviewing entity,
you know it's one of the members of
reviewing entity. So, if the reviewing
entity is, you know,
Lex Fridman's, you know, preferred
papers, right? You know it's Lex Fridman
writing the review.
>> Yes. What So, for me one that's a
beautiful
system, I think, but what in addition to
that it feels like there should be a
reputation system for the reviewers.
>> Absolutely. For the reviewing entities.
Not the reviewers individually.
>> The reviewing entities, sure.
>> But even that within that the reviewers
too, because uh
I 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 in the academic setting,
the incentive is kind of uh
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 That's a big part of my proposal,
actually. You know, I describe that as
you know, if if your evaluation of
papers is predictive of future success,
>> Yes. Yes. That's a good one.
>> then your reputation should go up as a
reviewing entity.
Um
So, yeah, exactly. I mean, that that um
I even had a master's student who was um
a master's student in uh library science
and computer science. Actually, kind of
worked out exactly how should that
should work with formulas and
everything, but
>> So, in terms of implementation, do you
think that's something that's doable?
>> I mean, I've been sort of, you know,
talking about this to sort of various
people like, you know, Andrew McCallum
who started OpenReview.
>> [snorts]
>> And the reason why we picked OpenReview
for ICLR initially, even though it was
very early for them, um is because my
hope was that ICLR was it was eventually
going to kind of
inaugurate this type of uh system. So, I
clear kept the idea of open reviews.
Uh, so whether reviews are you know,
published uh with the paper, which I
think is very useful.
But, in many ways has kind of reverted
to kind of more of a conventional type
conferences for everything else. And
that
I mean, I
I don't run I clear. I'm just the
president of the uh of the foundation,
but um
you know, people who run it should make
decisions about how to run it and I'm
not going to tell them um because
they're volunteers and I'm really
thankful that they do that. So, but I'm
saddened by the fact that uh we're not
being um innovative enough.
>> Yeah, me too. I hope that changes. Uh,
yeah. Cuz the communication [snorts]
science broadly, but communication
computer science ideas
is is is how you make those ideas have
impact, I think.
>> Yeah, and I think you know, a lot of
this is um 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 and give credits
accurately.
>> Mhm.
>> But, that comes at the expense of the
progress of science.
>> Yeah.
>> 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 have biases, you know, we're
doing you know, a double blind review,
but
uh
you know, the the the biases are are
still there. They're different different
kinds of biases.
>> You write that the phenomenon of
emergence, collective behavior exhibited
by 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 uh 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. It also is a mystery
for physicists. It's a mystery for
biologists. Um
you know, how is it that uh
>> [snorts]
>> the
uh universe around us seems to be
increasing in complexity and not
decreasing?
I mean, that that that is a kind of
curious property of uh physics that
despite the second law of
thermodynamics,
we seem to be, you know, evolution and
and learning and uh etc. seems to be
kind of at least locally
uh to increase uh
complexity and not decrease it. So,
perhaps the ultimate purpose of the
universe is uh to just get more complex.
>> Have these uh I mean,
uh small pockets of beautiful
complexity.
Uh
does that the cellular automata, these
kinds of emergence and 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, you got you got me into it. You
know, I I discovered the existence of uh
the perceptron when I was a college
student. Um
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 in the book and I I first time I
heard about Turing machine. Right? So, I
started digging the literature and I
found those paper those books which were
basically transcri- transcription of,
you know,
workshops or conferences
uh from the '50s and '60s about
self-organizing systems.
Mhm. So, there were there was a series
of conferences on self-organizing
systems and the these books on this um
some of them are you can actually get
them at the internet archive, you know,
the
digital version.
Uh
and there are like uh fascinating
articles in there by this guy whose name
has been largely forgotten, Heinz von
Foerster.
So, it's a German physicist who
immigrated to the US and uh worked on
self-organizing systems
in the 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 that lab never kind of strive
very much but but he wrote a bunch of
papers about self-organization and about
the 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 it's a bit of a mystery
you know how
like what what is possible with this
you know pattern formation in physical
systems
in chaotic system and things like that
you know you know the emergence of life
you know things like that so you know
how does
does that happen?
It's a
it's big puzzle for for physicists as
well.
>> It feels like understanding this the 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 um
>> But there's there's a missing concept
theoretical concept that we are we don't
have.
>> Yeah.
>> And it's 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 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-Solomonoff 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, right?
Um I've been fascinated by that concept.
But, the problem with that is that
that complexity is defined up to a
constant, which can be very large.
Right? Uh there's there are similar
concepts that are derived from, you
know,
you know, Bayesian probability theory,
where, you know, the complexity of
something is uh 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 uh
of of the distribution.
Uh 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
your Kolmogorov complexity.
And so, every measure of complexity we
have has some arbitrariness in it.
Um
you know, an additive constant which
is 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 this one way
that people study this in space biology,
the people that study the origin of life
or try to recreate the life in in
laboratory.
And the the more interesting one is the
alien one is when we go to other
planets,
how do we recognize that it's life?
Cuz 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
um measuring the level of complexity we
see to 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 um
if I give you uh an image of the MNIST
digits, right? And I I flip through
MNIST digits. There is some obviously
some
structure to it because
local structure, you know, neighboring
pixels are correlated uh
across the entire data set.
Now, imagine that um
I I apply a random permutation to all
the pixels.
>> Mhm.
>> A fixed random permutation.
Now, 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.
>> Okay.
>> [laughter]
>> 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 is random
uh fluctuation. It's heat.
>> Yeah.
>> Uh so
>> in the eye of the beholder.
>> Depends what kind of glasses you're
wearing.
>> 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 eye of the beholder.
>> [sighs]
>> 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.
>> Um because their notion of locality may
be different from ours.
>> Yeah.
>> This actually connects with uh
fascinating questions in physics at the
moment, like modern physics.
Uh
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 that, right?
And uh and that relies on notions of
complexity. Um
Yeah, which you know, I find I find it
fascinating.
>> Can you describe your personal quest to
build an expressive electronic wind
instrument, EWI? What is it? What does
it take
to uh to build it?
>> Well, I'm a tinkerer. I like building
things. Uh I like building things with
combinations of electronics and, you
know, mechanical stuff.
Um you know, I have a bunch of different
hobbies, but um
you know, probably my first one was
little was building model airplanes and
stuff like that, and I still do that to
some extent.
Uh but also electronics. I taught myself
electronics before I studied it. Um
and the reason I taught myself
electronics is because of music. Um
my cousin
was an inspiring electronic musician and
he had an analog synthesizer, and I was,
you know, basically modifying it for
him.
>> Mhm. [snorts]
>> Uh and building sequencers and stuff
like that, right, for him. I was I was
in high school when I was doing this.
>> And the interest in like progressive
rock, like '80s, like what's what's the
greatest band of all time according to
Yann LeCun?
>> Um there's too there's too many of them,
but you know, it's a combination of
uh
uh
you [clears throat] know, Mahavishnu
Orchestra, uh
Weather Report, Yes, Genesis, uh you
know,
>> Genesis.
>> pre-Peter Gabriel,
>> Uh-huh.
>> um
>> [laughter]
>> Gentle Giant, you know, things like
that.
>> Great. Okay, so this this love of
electronics and this love of music
combine together.
>> Right. So, I was actually trying to play
uh
uh Baroque and Renaissance uh music, and
I I played in a uh
orchestra when I was in high school and
uh first year of college.
And I played the recorder, crumhorn, uh
uh a little bit of oboe, you know,
things like that. So, I'm a I'm a wind
instrument player. But, I always wanted
to play improvised music, even though I
don't know anything about it. Uh and the
only way I figured, you know, short of
like learning to play the saxophone was
to play electronic wind instrument. So,
they behave you know, the fingering is
similar to a saxophone, but you know,
you have a wide variety of sound because
you control the synthesizer with it. So,
I I had a a bunch of those, you know,
going back to the late '80s
>> [snorts]
>> uh from either Yamaha or Akai. They
They're both kind of the main
manufacturers of those. So, they were
classically, you know, going back
several decades.
Uh 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,
uh
you have like various parameters you can
you can vary with with fingers, but they
they they're not really as expressive as
a acoustic instrument, right? You
you hear John Coltrane play two notes
and you hear you know it's John
Coltrane, you know it's got a unique
sound.
>> Mhm.
>> Uh or or Miles Davis, right? You can
hear it's Miles Davis
uh playing the trumpet because the the
sound
reflects their
you know, physiognomy, basically, the
shape of the vocal tract
um
kind of shapes the the sound. So, how
how do you do this with uh electronic
wind instrument? And it was um many
years ago I met a guy called David
Wessel. He he was a professor at
Berkeley and created the
the center for like, you know, music
technology there.
And he was interested in that question.
>> Mhm.
>> [clears throat]
>> 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 uh room and and and
home office. And uh
>> This is in New Jersey?
>> In New Jersey. And um I started uh
really being serious about you know
building my own EWI instrument.
>> What else is going on in that New Jersey
workshop? Is there some Is there some
crazy stuff you built that just or or
like left on the workshop floor? Left
behind.
>> Uh a lot of crazy stuff is uh you know
electronics we built with
microcontrollers of various kinds and
you know weird flying contraptions.
>> [laughter]
>> So you still love flying?
>> It's a family disease. My My dad got me
into it when I was a kid.
And
he was building model airplanes when he
was a kid and
and he was a mechanical engineer. He
taught himself electronics also so he he
built his early radio control systems in
the
late '60s, early '70s.
Um and so that that 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 and
appreciation of flight in other forms
like with drones, quadrocopters or do
you Is it Is it Is it model airplane the
the thing that
>> I you know before drones were you know
kind of
consumer products
um you know I built my own you know with
also building a microcontroller with
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.
Um
>> Yeah.
You were doing it before it was cool.
>> Yeah.
>> What
advice would you give to a young person
today in high school in 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? Uh
what is the universe made of? What's
life all about? Things like that.
Um
Like even like crazy big questions like
um what's time? Like nobody knows what
time is.
Um
And
uh and then
learn
basic things like basic methods either
from math from physics or from
engineering.
Uh things that have a long shelf life.
Um like if you have a choice between
like, you know,
learning uh you know, mobile programming
on iPhone
or quantum mechanics, take quantum
mechanics.
Um because you're going to learn things
that you have no idea exist. It and it
you may not you you you may never be a
quantum physicist
but you will learn about path integrals
and path integrals are used um
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 or within some
of these kind of more solidified fields
will have a longer shelf life. They will
you somehow use indirectly in in your
work.
>> Learn classical mechanics like you learn
about Lagrangians for example.
>> Mhm.
>> Um
which is like a a huge hugely useful
concept, you know, for all kinds of
different things. Uh
learn uh statistical physics because uh
all the math that comes out of you know,
for machine learning
uh basically comes out of was figured
out by statistical physicist in the you
know, late 19th early 20th century,
right? So
um and for some of them actually more
recently for by people like like Giorgio
Parisi who just got the Nobel Prize for
the replica method among other things.
Um it's used for a lot of different
things. Uh you know, variational
inference uh
that math comes from statistical
physics.
Um so um so a lot of those kind of, you
know,
basic courses, you know, you'll you'll
If you do electrical engineering, you
take signal processing, you'll you'll
learn about Fourier transforms.
Again, something super useful. It's at
the basis of things like graph neural
nets, which is a an entirely new
sub-area of, you know, AI machine
learning deep learning, which I think is
super promising for all kinds of
applications.
Um, something very promising if you're
more interested in applications is the
applications of, uh, AI machine learning
and deep learning to science.
Or to
science that can help solve big problems
in the world. I've I've colleagues at,
uh, at Meta at FAIR who started this
project called Open Catalyst, and it's,
um, it's an open project collaborative.
And the idea is to use, uh, deep
learning to
help design, uh, 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 you solve climate change.
It's as simple as that, because, um,
you you cover, you know, some random
desert with, uh, solar panels,
uh, and you have it have them work all
day, produce hydrogen, and then you ship
the hydrogen wherever it's needed. You
don't need anything else.
Um, you know, you you you have
controllable power that's, you know,
what can be transported, uh, anywhere.
So, if we if we have a large-scale
efficient, uh,
energy storage technology like producing
hydrogen,
uh, we we solve climate change. Here's
another way to solve climate change is,
uh, if you can figure 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,
um, you know, it uh, 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
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 for cars or airplanes or
things like that. Maybe better fuel
cells. 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 or cars.
We wouldn't have
emission problem CO2 emission problems
for for
air transportation anymore. So, this
there's a lot of those things I think
where AI you know, can be used it
and and this is not even talking about
all the sort of medicine, biology and
and everything like that, right?
You know, like
protein folding, you know, figuring out
like how can you design a protein that
it sticks to another protein that a
particular sites because that's how you
design drugs in the end.
So, you know, deep learning would be
useful and those are 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
a hexagonal mesh and then make this
single single atom thick.
You put another one on top. You twist
them by some magic
number of degrees, 3 degrees or
something.
It becomes superconductor.
Nobody has any idea why.
>> [laughter]
>> I want to 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 but there is
a hint perhaps that with machine
learning we could train a system to
basically be a phenomenological model of
some complex emergent phenomenon which,
you know,
superconductivity is one of those.
Uh
where, you know, think this collective
phenomenon is too difficult to describe
from first principles with the current,
you know,
the usual sort of, you know,
reductionist type method. But we could
have
uh deep learning systems that predict
the properties
of a system from a description of it uh
after being trained with sufficiently uh
many samples.
Um
There's um this guy Pascal Fua at EPFL.
He has a startup company um that
>> [snorts]
>> where he he
basically trained a
um
a convolutional net essentially to
predict the aerodynamic properties of
solids.
And you can generate as much data as you
want by just running uh computational
fluid dynamics, right? So, you give a
like a
a wing
airfoil or something shape of some kind,
and you run computational fluid
dynamics, you get as a result the drag
and, you know,
uh lift and all that stuff, right? And
you can 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 and lift as a
function of the shape of that solid, and
so you can do back gradient descent, you
can optimize the shape so you get the
properties you want.
>> [snorts]
>> Um
>> Yeah, that's incredible. That's
incredible. And then 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, cuz
after all all of these technologies will
have to work in a human world.
>> Yes.
>> And the human world is complicated.
>> It is.
So, yeah.
>> Yeah, and this is uh
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 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 It was
a pleasure.
>> Thanks for listening to this
conversation with Yann LeCun. 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.