Yann LeCun: Deep Learning, Convolutional Neural Networks, and Self-Supervised Learning
嘉宾：Yann LeCun
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
§01Transcript · 00:00–08:50
The following is a conversation with Jan Lekun.
He's considered to be one of the fathers of deep learning,
which if you've been hiding under a rock,
is the recent revolution in AI that's captivated the world
with the possibility of what machines can learn from data.
He's a professor at New York University,
a vice president and chief AI scientist at Facebook,
and co-recipient of the Turing Award
for his work on deep learning.
He's probably best known as the founding father
of convolutional neural networks.
In particular, their application
to optical character recognition
and the famed MNIST dataset.
He is also an outspoken personality,
unafraid to speak his mind in a distinctive French accent
and explore provocative ideas,
both in the rigorous medium of academic research
and the somewhat less rigorous medium
of Twitter and Facebook.
This is the Artificial Intelligence Podcast.
If you enjoy it, subscribe on YouTube,
give it five stars on iTunes, support it on Patreon,
or simply connect with me on Twitter at Lex Friedman,
spelled F-R-I-D-M-A-N.
And now, here's my conversation with Jan Lekun.
You said that 2001, Space Odyssey
is one of your favorite movies.
Hal 9000 decides to get rid of the astronauts
for people that haven't seen the movie, spoiler alert,
because he, it, she believes that the astronauts,
they will interfere with the mission.
Do you see Hal as flawed in some fundamental way
or even evil, or did he do the right thing?
- Neither.
There's no notion of evil in that context,
other than the fact that people die,
but it was an example of what people call
value misalignment, right?
You give an objective to a machine
and the machine strives to achieve this objective.
And if you don't put any constraints on this objective,
like don't kill people and don't do things like this,
the machine, given the power, will do stupid things
just to achieve this objective,
or damaging things to achieve this objective.
It's a little bit like, I mean, we're used to this
in the context of human society.
We put in place laws to prevent people
from doing bad things, because spontaneously
they would do those bad things, right?
So we have to shape their cost function,
their objective function, if you want,
through laws to kind of correct,
and education, obviously, to sort of correct for those.
- So maybe just pushing a little further on that point,
Hal, you know, there's a mission.
There's a, there's fuzziness around the ambiguity
around what the actual mission is,
but, you know, do you think that there will be a time
from a utilitarian perspective where an AI system,
where it is not misalignment, where it is alignment
for the greater good of society,
that an AI system will make decisions that are difficult?
- Well, that's the trick.
I mean, eventually we'll have to figure out how to do this.
And again, we're not starting from scratch
because we've been doing this with humans for millennia.
So designing objective functions for people
is something that we know how to do.
And we don't do it by, you know, programming things,
although the legal code is called code.
So that tells you something.
And it's actually the design of an objective function.
That's really what legal code is, right?
It tells you, here's what you can do,
here's what you can't do.
If you do it, you pay that much.
That's an objective function.
So there is this idea somehow that it's a new thing
for people to try to design objective functions
that are aligned with the common good.
But no, we've been writing laws for millennia
and that's exactly what it is.
So that's where, you know, the science of lawmaking
and computer science will-
- Come together.
- Will come together.
- So it's nothing, there's nothing special about HAL
or AI systems.
It's just the continuation of tools used
to make some of these difficult ethical judgments
that laws make.
- Yeah, and we have systems like this already
that make many decisions for ourselves in society
that need to be designed in a way that they,
like rules about things that sometimes have bad side effects.
And we have to be flexible enough about those rules
so that they can be broken when it's obvious
that they shouldn't be applied.
So you don't see this on the camera here,
but all the decoration in this room
is all pictures from 2001 and Space Odyssey.
- Wow, is that by accident or is there a lot-
- It's not by accident, it's by design.
(Lex laughing)
- Oh, wow.
So if you were to build HAL 10,000,
so an improvement of HAL 9,000, what would you improve?
- Well, first of all,
I wouldn't ask you to hold secrets and tell lies
because that's really what breaks it in the end.
That's the fact that it's asking itself questions
about the purpose of the mission.
And it's, you know, pieces things together
that it's heard, you know,
all the secrecy of the preparation of the mission
and the fact that it was a discovery on the lunar surface
that really was kept secret.
And one part of HAL's memory knows this,
and the other part does not know it
and is supposed to not tell anyone.
And that creates internal conflict.
- So you think there's never should be a set of things
that an AI system should not be allowed,
like a set of facts that should not be shared
with the human operators?
- Well, I think, no, I think it should be a bit like
in the design of autonomous AI systems,
there should be the equivalent of, you know,
the oath that Hippocrates oaths.
- Hippocratic oath, yeah.
- That doctors sign up to, right?
So there's certain things, certain rules
that you have to abide by,
and we can sort of hardwire this into our machines
to kind of make sure they don't go.
So I'm not, you know, an advocate of the three laws
of robotics, you know, the Asimov kind of thing,
because I don't think it's practical,
but, you know, some level of limits.
But to be clear, this is not,
these are not questions that are kind of really worth
asking today because we just don't have the technology
to do this.
We don't have autonomous intelligent machines.
We have intelligent machines,
semi-intelligent machines that are very specialized.
But they don't really sort of satisfy an objective.
They're just, you know, kind of trained to do one thing.
So until we have some idea for design
of a full-fledged autonomous intelligent system,
asking the question of how we design this objective,
I think is a little too abstract.
- It's a little too abstract.
There's useful elements to it
in that it helps us understand our own ethical codes,
humans.
So even just as a thought experiment,
if you imagine that an AGI system is here today,
how would we program it as a kind of nice thought experiment
of constructing how should we have a law,
have a system of laws for us humans.
It's just a nice practical tool.
And I think there's echoes of that idea too
in the AI systems we have today
that don't have to be that intelligent,
like autonomous vehicles.
These things start creeping in
that we're thinking about,
but certainly they shouldn't be framed as hell.
- Yeah.
- Looking back, what is the most,
I'm sorry if it's a silly question,
but what is the most beautiful or surprising idea
in deep learning or AI in general
that you've ever come across?
Sort of personally, when you said back
and just had this kind of,
oh, that's pretty cool moment.
That's nice.
That's surprising.
- I don't know if it's an idea
rather than a sort of empirical fact.
The fact that you can build gigantic neural nets,
train them on relatively small amounts of data,
relatively, with stochastic gradient descent
and that it actually works,
breaks everything you read in every textbook, right?
Every pre-deep learning textbook that told you
you need to have fewer parameters
and you have data samples.
If you have a non-convex objective function,
you have no guarantee of convergence.
All those things that you read in textbook
and they tell you to stay away from this
and they're all wrong.
- Huge number of parameters, non-convex,
and somehow which is very relative
to the number of parameters, data,
§02Transcript · 08:53–17:29
it's able to learn anything.
- Right.
- Does that still surprise you today?
- Well, it was kind of obvious to me
before I knew anything that this is a good idea.
And then it became surprising that it worked
because I started reading those textbooks.
Okay.
- So, okay, so can you talk through the intuition
of why it was obvious to you if you remember?
- Well, okay, so the intuition was,
it's sort of like those people in the late 19th century
who proved that heavier than air flight was impossible, right?
And of course you have birds, right?
They do fly.
And so on the face of it,
it's obviously wrong as an empirical question, right?
And so we have the same kind of thing that,
we know that the brain works,
we don't know how, but we know it works.
And we know it's a large network of neurons in interaction
and that learning takes place by changing the connections.
So kind of getting this level of inspiration
without copying the details,
but sort of trying to derive basic principles,
you know, that kind of gives you a clue
as to which direction to go.
There's also the idea somehow
that I've been convinced of since I was an undergrad
that even before,
that intelligence is inseparable from learning.
So the idea somehow that you can create
an intelligent machine by basically programming,
for me, it was a non-starter from the start.
Every intelligent entity that we know about
arrives at this intelligence through learning.
So learning, you know,
machine learning was a completely obvious path.
Also because I'm lazy, so, you know,
kind of. (laughs)
- You automate basically everything
and learning is the automation of intelligence.
- Right.
- So do you think, so what is learning then?
What falls under learning?
Because do you think of reasoning as learning?
- Well, reasoning is certainly a consequence
of learning as well,
just like other functions of the brain.
The big question about reasoning is,
how do you make reasoning compatible
with gradient-based learning?
- Do you think neural networks can be made to reason?
- Yes, there is no question about that.
Again, we have a good example, right?
The question is how.
So the question is how much prior structure
do you have to put in the neural net
so that something like human reasoning
will emerge from it, you know, from learning?
Another question is,
all of our kind of model of what reasoning is
that are based on logic are discrete
and are therefore incompatible with gradient-based learning.
And I'm a very strong believer
in this idea of gradient-based learning.
I don't believe that other types of learning
that don't use kind of gradient information, if you want.
- So you don't like discrete mathematics?
You don't like anything discrete?
- Well, it's not that I don't like it,
it's just that it's incompatible with learning
and I'm a big fan of learning, right?
So in fact, that's perhaps one reason
why deep learning has been kind of looked at
with suspicion by a lot of computer scientists
because the math is very different.
The math that you use for deep learning,
you know, has more to do with cybernetics,
the kind of math you do in electrical engineering
than the kind of math you do in computer science.
And, you know, nothing in machine learning is exact, right?
Computer science is all about sort of, you know,
obsessive-compulsive attention to details
of like, you know, every index has to be right
and you can prove that an algorithm is correct, right?
Machine learning is the science of sloppiness, really.
- That's beautiful.
So, okay, maybe let's feel around in the dark
of what is a neural network that reasons
or a system that works with continuous functions
that's able to do, build knowledge,
however we think about reasoning,
build on previous knowledge, build on extra knowledge,
create new knowledge,
generalize outside of any training set ever built.
What does that look like?
If, yeah, maybe do you have inklings of thoughts
of what that might look like?
- Yeah, I mean, yes and no.
If I had precise ideas about this,
I think, you know, we'd be building it right now.
But, and there are people working on this
or whose main research interest
is actually exactly that, right?
So what you need to have is a working memory.
So you need to have some device, if you want,
some subsystem that can store a relatively large number
of factual episodic information
for a reasonable amount of time.
So in the brain, for example,
there are kind of three main types of memory.
One is the sort of memory of the state of your cortex,
and that sort of disappears within 20 seconds.
You can't remember things for more than about 20 seconds
or a minute if you don't have any other form of memory.
The second type of memory, which is longer term,
but still short term, is the hippocampus.
So you can, you know, you came into this building,
you remember where the exit is, where the elevators are.
You have some map of that building
that's stored in your hippocampus.
You might remember something about what I said,
you know, if you've been to the zoo,
you might remember something about what I said,
you know, a few minutes ago.
- I forgot it all already, but it's part.
- Of course, it's been erased, but you know,
but that would be in your hippocampus.
And then the longer term memory is in the synapse,
the synapses, right?
So what you need if you want a system
that's capable of reasoning
is that you want the hippocampus-like thing, right?
And that's what people have tried to do
with memory networks and, you know,
neural training machines and stuff like that, right?
Transformers, which have sort of a memory in there,
kind of self-attention system.
You can think of it this way.
So that's one element you need.
Another thing you need is some sort of network
that can access this memory,
get an information back, and then kind of crunch on it,
and then do this iteratively multiple times,
because a chain of reasoning
is a process by which you update your knowledge
about the state of the world,
about what's going to happen, et cetera.
And that has to be this sort of
recurrent operation, basically.
- And you think that kind of,
if we think about a transformer,
so that seems to be too small to contain the knowledge
that's to represent the knowledge
that's contained in Wikipedia, for example.
- Well, a transformer doesn't have this idea of recurrence.
It's got a fixed number of layers,
and that's the number of steps that limits,
basically, its representation.
- But recurrence would build on the knowledge somehow.
I mean, it would evolve the knowledge
and expand the amount of information, perhaps,
or useful information within that knowledge.
But is this something that just can emerge with size?
Because it seems like everything we have now is too small.
- Not just.
No, it's not clear.
I mean, how you access and write
into an associative memory inefficient
way, I mean, sort of the original memory network
maybe had something like the right architecture,
but if you try to scale up a memory network
so that the memory contains all of Wikipedia,
it doesn't quite work.
- Right.
- So there's a need for new ideas there, okay.
But it's not the only form of reasoning.
So there's another form of reasoning,
which is very classical, also, in some types of AI,
and it's based on, let's call it energy minimization.
Okay, so you have some sort of objective,
some energy function that represents
the quality or the negative quality, okay.
Energy goes up when things get bad
and they get low when things get good.
So let's say you want to figure out,
you know, what gestures do I need to do
to grab an object or walk out the door.
If you have a good model of your own body,
a good model of the environment,
using this kind of energy minimization,
you can do planning.
And in optimal control,
it's called model predictive control.
You have a model of what's going to happen in the world
as a consequence of your actions,
and that allows you to, by energy minimization,
figure out a sequence of action
that optimizes a particular objective function,
§03Transcript · 17:32–26:16
which minimizes the number of times
you're going to hit something
and the energy you're going to spend
doing the gesture and et cetera.
So that's a form of reasoning.
Planning is a form of reasoning.
And perhaps what led to the ability of humans to reason
is the fact that, or, you know,
species that appear before us
had to do some sort of planning
to be able to hunt and survive
and survive the winter in particular.
And so, you know, it's the same capacity
that you need to have.
- So in your intuition is,
if we look at expert systems,
and encoding knowledge as logic systems,
as graphs, or in this kind of way,
is not a useful way to think about knowledge?
- Graphs are a little brittle, or logic representation.
So basically, you know, variables that have values
and then constraint between them
that are represented by rules
is a little too rigid and too brittle, right?
So one of the, you know, some of the early efforts
in that respect were to put probabilities on them.
So a rule, you know, if you have this and that symptom,
you know, you have this disease with that probability,
and you should prescribe that antibiotic
with that probability, right?
That's the myosin system from the '70s.
And that's what that branch of AI led to,
you know, Bayesian networks and graphical models
and causal inference and variational, you know, method.
So there is, I mean, certainly a lot of interesting work
going on in this area.
The main issue with this is knowledge acquisition.
How do you reduce a bunch of data to a graph of this type?
- Yeah, it relies on the expert to,
on the human being to encode, to add knowledge.
- And that's essentially impractical.
- Yeah, it's not scalable.
- That's a big question.
The second question is,
do you want to represent knowledge as symbols,
and do you want to manipulate them with logic?
And again, that's incompatible with learning.
So one suggestion with, you know,
Geoff Hinton has been advocating for many decades
is replace symbols by vectors.
Think of it as pattern of activities
in a bunch of neurons or units
or whatever you want to call them,
and replace logic by continuous functions.
Okay, and that becomes now compatible.
There's a very good set of ideas
by, written in a paper about 10 years ago
by Leon Boutou, who is here at Facebook.
The title of the paper is
"From Machine Learning to Machine Reasoning."
And his idea is that a learning system
should be able to manipulate objects
that are in a space,
and then put the result back in the same space.
So it's this idea of working memory, basically.
And it's very enlightening.
- And in a sense, that might learn something
like the simple expert systems.
I mean, you can learn basic logic operations there.
- Yeah, quite possibly.
There's a big debate on sort of how much prior structure
you have to put in for this kind of stuff to emerge.
That's the debate I have with Gary Marcus
and people like that.
- Yeah, yeah.
So, and the other person,
so I just talked to Judea Pearl,
from the, you mentioned causal inference world.
So his worry is that the current neural networks
are not able to learn
what causes what causal inference between things.
- So I think he's right and wrong about this.
If he's talking about the sort of classic
type of neural nets,
people sort of didn't worry too much about this.
But there's a lot of people now working on causal inference.
And there's a paper that just came out last week
by Leon Boutou, among others,
David Lopez-Paz and a bunch of other people,
exactly on that problem of how do you kind of,
get a neural net to sort of pay attention
to real causal relationships,
which may also solve issues of bias in data
and things like this.
- I'd like to read that paper because that ultimately,
the challenge there is also seems to fall back
on the human expert
to ultimately decide causality between things.
- People are not very good at establishing causality,
first of all.
So first of all, you talk to physicists
and physicists actually don't believe in causality
because look at all the basic laws of macrophysics
are time reversible.
So there is no causality.
- The arrow of time is not real.
- It's as soon as you start looking at macroscopic systems,
where there is unpredictable randomness,
where there is clearly an arrow of time,
but it's a big mystery in physics actually,
well, how that emerges.
- Is it emergent or is it part of the fundamental fabric
of reality?
- Or is it a bias of intelligent systems that,
because of the second law of thermodynamics,
we perceive a particular arrow of time,
but in fact, it's kind of arbitrary, right?
- So yeah, physicists, mathematicians,
they don't care about, I mean,
the math doesn't care about the flow of time.
- Well, certainly macrophysics doesn't.
People themselves are not very good at establishing
causal relationships.
If you ask, I think it was in one of Seymour Papert's book
on like children learning,
he studied with Jean Piaget,
he's the guy who co-authored the book "Perceptron"
with Marvin Minsky that kind of killed
the first wave of neural nets,
but he was actually a learning person.
He, in the sense of studying learning in humans
and machines, that's why he got interested in perceptron.
And he wrote that if you ask a little kid
about what is the cause of the wind,
a lot of kids will say, they will think for a while
and they'll say, "Oh, it's the branches in the trees.
They move and that creates wind."
So they get the causal relationship backwards.
And it's because their understanding of the world
and intuitive physics is not that great.
I mean, these are like four or five year old kids.
It gets better and then you understand that this,
it can't be.
But there are many things which we can,
because of our common sense understanding of things,
what people call common sense,
and our understanding of physics,
there's a lot of stuff that we can figure out
'cause even with diseases, we can figure out
what's not causing what often.
There's a lot of mystery, of course,
but the idea is that you should be able
to encode that into systems.
'Cause it seems unlikely they'd be able
to figure that out themselves.
- Well, whenever we can do intervention,
but all of humanity has been completely deluded
for millennia, probably since its existence,
about a very, very wrong causal relationship
where whatever you can explain,
you attribute it to some deity, some divinity.
And that's a cop-out.
That's a way of saying, "I don't know the cause,
so God did it."
- So you mentioned Marvin Minsky and the irony of
maybe causing the first AI winter.
You were there in the '90s, you were there in the '80s,
of course.
In the '90s, why do you think people lost faith
in deep learning in the '90s and found it again
a decade later, over a decade later?
- Yeah, it wasn't called deep learning yet.
It was just called neural nets.
- Neural networks.
- Yeah, they lost interest.
I mean, I think I would put that around 1995,
at least the machine learning community.
There was always a neural net community,
but it became kind of disconnected
from mainstream machine learning, if you want.
There were, it was basically electrical engineering
that kept at it.
And computer science.
- Just gave up.
- Gave up on neural nets.
I don't know.
I was too close to it to really sort of analyze it
with sort of an unbiased eye, if you want.
But I would make a few guesses.
So the first one is, at the time, neural nets were,
it was very hard to make them work,
in the sense that you would implement backprop
in your favorite language.
And that favorite language was not Python.
It was not MATLAB.
It was not any of those things,
'cause they didn't exist, right?
You had to write it in Fortran or C
or something like this, right?
So you would experiment with it.
§04Transcript · 26:18–34:03
You would probably make some very basic mistakes,
like, you know, barely initialize your weights,
make the network too small
because you read in a textbook, you know,
you don't want too many parameters, right?
And of course, you know, and you would train on XOR
because you didn't have any other dataset to train on.
And of course, you know, it works half the time.
So you would say, "I give up."
Also, you would train it with batch gradient,
which, you know, isn't really sufficient.
So there was a lot of, there was a bag of tricks
that you had to know to make those things work,
or you had to reinvent.
And a lot of people just didn't,
and they just couldn't make it work.
So that's one thing.
The investment in software platform
to be able to kind of, you know, display things,
figure out why things don't work,
kind of get a good intuition for how to get them to work,
have enough flexibility so you can create, you know,
network architectures like convolutional nets
and stuff like that.
It was hard.
I mean, you had to write everything from scratch.
And again, you didn't have any Python
or MATLAB or anything, right?
- I read that, sorry to interrupt,
but I read that you wrote in Lisp,
your first versions of Linnet
with the convolutional networks,
which by the way, one of my favorite languages.
That's how I knew you were legit.
Turing Award, whatever.
You programmed in Lisp.
- It's still my favorite language.
But it's not that we programmed in Lisp,
it's that we had to write our Lisp interpreter.
Okay, 'cause it's not like we used one that existed.
So we wrote a Lisp interpreter
that we hooked up to, you know,
a backend library that we wrote
also for sort of neural net computation.
And then after a few years, around 1991,
we invented this idea of basically having modules
that know how to forward propagate
and back propagate gradients,
and then interconnecting those modules in a graph.
Leombo 2 had made proposals on this,
about this in the late 80s,
and we were able to implement this using our Lisp system.
Eventually we wanted to use that system
to build production code
for character recognition at Bell Labs.
So we actually wrote a compiler for that Lisp interpreter
so that Petrissi Mart, who is now at Microsoft,
kind of did the bulk of it with Leon and me.
And so we could write our system in Lisp
and then compile to C,
and then we'll have a self-contained compute system
that could kind of do the entire thing.
Neither PyTorch nor TensorFlow can do this today.
Yet.
Okay, it's coming.
- Yeah.
(laughs)
- I mean, there's something like that in PyTorch
called TorchScript.
And so we had to write our Lisp interpreter,
we had to write our Lisp compiler,
we had to invest a huge amount of effort to do this.
And not everybody,
if you don't completely believe in the concept,
you're not going to invest the time to do this.
Now, at the time also,
or today, this would turn into Torch or PyTorch
or TensorFlow or whatever.
We'd put it in open source,
everybody would use it and realize it's good.
Back before 1995, working at AT&T,
there's no way the lawyers would let you
release anything in open source of this nature.
And so we could not distribute our code, really.
- And on that point,
and sorry to go on a million tangents,
but on that point,
I also read that there was some almost patent,
like a patent on convolutional neural networks.
- Yes, there was.
- So that, first of all,
I mean, just--
- There's two, actually.
- That ran out.
- Thankfully, in 2007.
- In 2007.
Can we just talk about that for a second?
I know you're at Facebook,
but you're also at NYU.
What does it mean to patent ideas
like these software ideas, essentially?
Or mathematical ideas?
Or what are they?
- Okay, so they're not mathematical ideas.
They are algorithms.
And there was a period where the US patent office
would allow the patent of software
as long as it was embodied.
The Europeans are very different.
They don't quite accept that.
They have a different concept, but you know.
I don't, I no longer,
I mean, I never actually strongly believed in this,
but I don't believe in this kind of patent.
Facebook basically doesn't believe in this kind of patent.
Google fires patents because they've been burned with Apple.
And so now they do this for defensive purpose,
but usually they say,
"We're not gonna sue you if you infringe."
Facebook has a similar policy.
They say, "We fire patents on certain things
"for defensive purpose.
"We're not gonna sue you if you infringe,
"unless you sue us."
So the industry does not believe in patents.
They are there because of the legal landscape
and various things,
but I don't really believe in patents
for this kind of stuff.
- So that's a great thing.
So I-
- I'll tell you a worse story, actually.
So what happens was the first patent about convolutional net
was about kind of the early version of convolutional net
that didn't have separate pooling layers.
It had convolutional layers
with stride more than one, if you want, right?
And then there was a second one on convolutional nets
with separate pooling layers, training with backprop.
And they were filed in '89 and 1990
or something like this.
At the time, the life of a patent was 17 years.
So here's what happened over the next few years
is that we started developing
character recognition technology around convolutional nets.
And in 1994, a check reading system
was deployed in ATM machines.
In 1995, it was for large check reading machines
in back offices, et cetera.
And those systems were developed by an engineering group
that we were collaborating with at AT&T.
And they were commercialized by NCR,
which at the time was a subsidiary of AT&T.
Now AT&T split up in 1996, early 1996.
And the lawyers just looked at all the patents
and they distributed the patents among the various companies.
They gave the convolutional net patent to NCR
because they were actually selling products that used it.
But nobody at NCR had any idea what a convolutional net was.
- Yeah.
- Okay.
So between 1996 and 2007,
so there's a whole period until 2002
where I didn't actually work on machine learning
or convolutional net.
I resumed working on this around 2002.
And between 2002 and 2007, I was working on them,
crossing my finger that nobody at NCR would notice.
And nobody noticed.
- Yeah, and I hope that this kind of somewhat,
as you said, lawyers aside,
relative openness of the community now will continue.
- It accelerates the entire progress of the industry.
And the problems that Facebook and Google
and others are facing today
is not whether Facebook or Google or Microsoft or IBM
or whoever is ahead of the other,
is that we don't have the technology
to build the things we want to build.
We want to build intelligent virtual assistants
that have common sense.
We don't have monopoly on good ideas for this.
We don't believe we do.
Maybe others believe they do, but we don't.
Okay.
If a startup tells you they have the secret
to human level intelligence and common sense,
don't believe them.
They don't.
And it's going to take the entire work
of the world research community for a while
to get to the point where you can go off
and each of those companies
is going to start to build things on this.
We're not there yet.
- It's absolutely.
This calls to the gap between the space of ideas
and the rigorous testing of those ideas
of practical application that you often speak to.
You've written advice saying,
§05Transcript · 34:05–42:26
"Don't get fooled by people who claim
"to have a solution to artificial general intelligence,
"who claim to have an AI system
"that works just like the human brain
"or who claim to have figured out how the brain works.
"Ask them what the error rate they get
"on MNIST or ImageNet."
- Yeah, this is a little dated by the way.
(laughs)
- I mean, five years, who's counting?
Okay.
But I think your opinion is still MNIST and ImageNet,
yes, may be dated.
There may be new benchmarks, right?
But I think that philosophy is one you still
and somewhat hold that benchmarks
and the practical testing, the practical application
is where you really get to test the ideas.
- Well, it may not be completely practical.
Like for example, it could be a toy dataset,
but it has to be some sort of task
that the community as a whole has accepted
as some sort of standard kind of benchmark, if you want.
It doesn't need to be real.
So for example, many years ago here at FAIR,
people, Jason West, Antoine Born and a few others
proposed the Babi tasks,
which were kind of a toy problem to test
the ability of machines to reason actually
to access working memory and things like this.
And it was very useful, even though it wasn't a real task.
MNIST is kind of halfway real task.
So, toy problems can be very useful.
It's just that I was really struck by the fact that
a lot of people, particularly a lot of people
with money to invest would be fooled by people telling them,
oh, we have the algorithm of the cortex
and you should give us 50 million.
- Yes, absolutely.
So there's a lot of people who try to take advantage
of the hype for business reasons and so on.
But let me sort of talk to this idea
that sort of new ideas,
the ideas that push the field forward
may not yet have a benchmark,
or it may be very difficult to establish a benchmark.
- I agree.
That's part of the process.
Establishing benchmarks is part of the process.
- So what are your thoughts about,
so we have these benchmarks on,
around stuff we can do with images,
from classification to captioning
to just every kind of information you can pull off
from images and the surface level.
There's audio data sets, there's some video.
What can we start, natural language,
what kind of benchmarks do you see that start creeping
on to more something like intelligence,
like reasoning, like maybe you don't like the term,
but AGI, echoes of that kind of formulation?
- A lot of people are working on interactive environments
in which you can train and test intelligent systems.
So there, for example,
the classical paradigm of supervised learning
is that you have a data set,
you partition it into a training set,
validation set, test set,
and there's a clear protocol, right?
But what if, that assumes that the samples
are statistically independent,
you can exchange them,
the order in which you see them shouldn't matter,
things like that.
But what if the answer you give
determines the next sample you see,
which is the case, for example, in robotics, right?
You robot does something
and then it gets exposed to a new room
and depending on where it goes,
the room would be different.
So that creates the exploration problem.
The what if the samples,
so that creates also a dependency between samples, right?
If you can only move in space,
the next sample you're gonna see
is gonna be probably in the same building, most likely.
So all the assumptions about the validity
of this training set, test set hypothesis break
whenever a machine can take an action
that has an influence in the world
and it's what it's gonna see.
So people are setting up artificial environments
where that takes place, right?
The robot runs around a 3D model of a house
and can interact with objects and things like this.
So you do robotics by simulation,
you have those, you know,
opening a gym type thing
or MuJoCo kind of simulated robots
and you have games, you know, things like that.
So that's where the field is going really,
this kind of environment.
Now, back to the question of AGI,
like I don't like the term AGI
because it implies that human intelligence is general
and human intelligence is nothing like general.
It's very, very specialized.
We think it's general, we like to think of ourselves
as having general intelligence, we don't.
We're very specialized.
We're only slightly more general than--
- Why does it feel general?
So you kind of, the term general,
I think what's impressive about humans
is ability to learn, as we were talking about learning,
to learn in just so many different domains.
It's perhaps not arbitrarily general,
but just you can learn in many domains
and integrate that knowledge somehow.
- Okay. - The knowledge persists.
- So let me take a very specific example.
It's not an example, it's more like
a quasi-mathematical demonstration.
So you have about one million fibers
coming out of one of your eyes, okay, two million total,
but let's talk about just one of them.
It's one million nerve fibers, your optical nerve.
Let's imagine that they are binary,
so they can be active or inactive, right?
So the input to your visual cortex is one million bits.
Now, they're connected to your brain in a particular way,
and your brain has connections
that are kind of a little bit like a convolutional net,
they're kind of local in space and things like this.
Now imagine I play a trick on you.
It's a pretty nasty trick, I admit.
I cut your optical nerve,
and I put a device that makes a random perturbation,
a permutation of all the nerve fibers.
So now what comes to your brain
is a fixed but random permutation of all the pixels.
There's no way in hell that your visual cortex,
even if I do this to you in infancy,
will actually learn vision
to the same level of quality that you can.
- Got it, and you're saying
there's no way you've relearned that?
- No, because now two pixels that are nearby in the world
will end up in very different places in your visual cortex,
and your neurons there have no connections with each other
because they're only connected locally.
- So this whole, our entire,
the hardware is built in many ways to support--
- The locality of the real world.
- Yeah.
- Yes, that's specialization.
- Yeah, but it's still pretty damn impressive.
So it's not perfect generalization.
It's not even close.
- No, no, it's not that it's not even close.
It's not at all.
- Yeah, it's not.
It's specialized, yeah.
- So how many Boolean functions,
so let's imagine you want to train your visual system
to recognize particular patterns of those one million bits.
Okay, so that's a Boolean function, right?
Either the pattern is here or not here.
It's a two-way classification
with one million binary inputs.
How many such Boolean functions are there?
Okay, you have two to the one million
combinations of inputs.
For each of those, you have an output bit.
And so you have two to the one million
Boolean functions of this type, okay?
Which is an unimaginably large number.
How many of those functions can actually be computed
by your visual cortex?
And the answer is a tiny, tiny, tiny, tiny, tiny, tiny sliver
like an enormously tiny sliver.
- Yeah, yeah.
- So we are ridiculously specialized.
- But, okay.
(laughing)
Okay, that's an argument against the word general.
I think there's a,
I agree with your intuition, but I'm not sure it's,
it seems the brain is impressively
capable of adjusting to things, so.
- It's because we can't imagine tasks
that are outside of our comprehension, right?
So we think we are general
because we are general of all the things
that we can apprehend.
But there is a huge world out there
of things that we have no idea.
We call that heat, by the way.
- Heat. - Heat.
§06Transcript · 42:27–50:45
So, at least physicists call that heat,
or they call it entropy, which is kind of,
you have a thing full of gas, right?
- Closed system full of gas.
- Right?
Closed or not closed.
It has pressure, it has temperature,
it has, and you can write equations,
PV equal NRT, things like that, right?
When you reduce the volume, the temperature goes up,
the pressure goes up, things like that, right?
For a perfect gas, at least.
Those are the things you can know about that system.
And it's a tiny, tiny number of bits
compared to the complete information
of the state of the entire system.
Because the state of the entire system
will give you the position and momentum
of every molecule of the gas.
And what you don't know about it is the entropy,
and you interpret it as heat.
The energy contained in that thing is what we call heat.
Now, it's very possible that, in fact,
there is some very strong structure
in how those molecules are moving.
It's just that they are in a way
that we are just not wired to perceive.
- Yeah, we're ignorant to it.
And there's an infinite amount of things
we're not wired to perceive.
And you're right, that's a nice way to put it.
We're general to all the things we can imagine,
which is a very tiny subset of all things that are possible.
- So it's like Kolmogorov complexity
or the Kolmogorov-Chaitin summa of complexity.
Every bit string or every integer is random,
except for all the ones that you can actually write down.
(both laughing)
- Yeah, okay, so beautifully put.
So we can just call it artificial intelligence.
We don't need to have a general.
- Or human level.
Human level intelligence is good.
Anytime you touch human, it gets interesting
because we attach ourselves to human
and it's difficult to define what human intelligence is.
Nevertheless, my definition is maybe
damn impressive intelligence, okay?
Damn impressive demonstration of intelligence, whatever.
And so on that topic, most successes in deep learning
have been in supervised learning.
What is your view on unsupervised learning?
Is there a hope to reduce involvement of human input
and still have successful systems
that have practical use?
- Yeah, I mean, there's definitely a hope.
It's more than a hope, actually.
It's mounting evidence for it.
And that's basically all I do.
The only thing I'm interested in at the moment is,
I call it self-supervised learning, not unsupervised,
'cause unsupervised learning is a loaded term.
People who know something about machine learning
tell you, so you're doing clustering or PCA,
which is not the case.
And the wide public, when you say unsupervised learning,
oh my God, machines are gonna learn by themselves
and without supervision.
- Where's the parents?
- Yeah, so I call it self-supervised learning
because in fact, the underlying algorithms that are used
are the same algorithms
as the supervised learning algorithms,
except that what we train them to do
is not predict a particular set of variables,
like the category of an image,
and not to predict a set of variables
that have been provided by human labelers.
But what you train the machine to do is basically
reconstruct a piece of its input
that is being masked out, essentially.
You can think of it this way, right?
So show a piece of video to a machine
and ask it to predict what's gonna happen next.
And of course, after a while, you can show what happens
and the machine will kind of train itself
to do better at that task.
You can do, like all the latest, most successful models
in natural language processing
use self-supervised learning.
You know, sort of BERT-style systems, for example, right?
You show it a window of a dozen words on a test corpus,
you take out 15% of the words,
and then you train a machine
to predict the words that are missing.
That's self-supervised learning.
It's not predicting the future,
it's just predicting things in the middle,
but you could have it predict the future,
that's what language models do.
- So you construct, so in an unsupervised way,
you construct a model of language.
Do you think-
- Or video, or the physical world, or whatever, right?
- How far do you think that can take us?
Do you think BERT understands anything?
- To some level.
It has a shallow understanding of text,
but it needs to, I mean,
to have kind of true human level intelligence,
I think you need to ground language in reality.
So some people are attempting to do this, right?
Having systems that kind of have some visual representation
of what is being talked about,
which is one reason you need
those interactive environments, actually.
But it's like a huge technical problem that is not solved,
and that explains why self-supervised learning works
in the context of natural language,
but does not work in the context,
or at least not well,
in the context of image recognition and video,
although it's making progress quickly.
And the reason, that reason is the fact that
it's much easier to represent uncertainty in the prediction
in the context of natural language
than it is in the context of things like video and images.
So for example, if I ask you to predict
what words are missing,
you know, 15% of the words that are taken out.
- The possibility is just small.
I mean- - It's small, right?
There is a hundred thousand words in the lexicon,
and what the machine spits out
is a big probability vector, right?
It's a bunch of numbers between zero and one
that's onto one.
And we know how to do this with computers.
So there, representing uncertainty in the prediction
is relatively easy,
and that's, in my opinion,
why those techniques work for NLP.
For images, if you ask,
if you block a piece of an image
and you ask the system,
reconstruct that piece of the image,
there are many possible answers
that are all perfectly legit, right?
And how do you represent that,
this set of possible answers?
You can't train a system to make one prediction.
You can't train a neural net to say,
here it is, that's the image,
because there's a whole set of things
that are compatible with it.
So how do you get the machine to represent
not a single output,
but a whole set of outputs?
And similarly with video prediction,
there's a lot of things that can happen
in the future of a video.
You're looking at me right now,
I'm not moving my head very much,
but I might turn my head to the left or to the right.
If you don't have a system that can predict this,
and you train it with least square
to kind of minimize the error with a prediction
on what I'm doing,
what you get is a blurry image of myself
in all possible future positions that I might be in,
which is not a good prediction.
- But so there might be other ways
to do the self-supervision, right?
For visual scenes.
- Like what?
(laughs)
- If I knew, I wouldn't tell you.
Publish it first, I don't know.
- No, there might be.
- So, I mean, these are kind of,
there might be artificial ways of like self-play in games
to where you can simulate part of the environment.
You can-
- Oh, that doesn't solve the problem.
It's just a way of generating data.
- But because you have more of a control,
like maybe you can control,
yeah, it's a way to generate data.
That's right.
And because you can do huge amounts of data generation,
that doesn't, you're right.
Well, it creeps up on the problem from the side of data,
and you don't think that's the right way to creep up.
- It doesn't solve this problem
of handling uncertainty in the world, right?
So if you have a machine learn a predictive model
of the world in a game that is deterministic
or quasi-deterministic, it's easy, right?
Just give a few frames of the game to a ConvNet,
put a bunch of layers,
§07Transcript · 50:47–59:40
and then have the game generates the next few frames.
And if the game is deterministic, it works fine.
And that includes feeding the system with the action
that your little character is gonna take.
The problem comes from the fact that the real world
and most games are not entirely predictable.
And so there you get those blurry predictions,
and you can't do planning with blurry predictions.
Right, so if you have a perfect model of the world,
you can, in your head, run this model
with a hypothesis for a sequence of actions,
and you're going to predict the outcome
of that sequence of actions.
But if your model is imperfect, how can you plan?
- Yeah, it quickly explodes.
What are your thoughts on the extension of this,
which topic I'm super excited about,
it's connected to something you were talking about
in terms of robotics, is active learning.
So as opposed to sort of completely unsupervised
or self-supervised learning,
you ask the system for human help
for selecting parts you want annotated next.
So if you think about a robot exploring a space
or a baby exploring a space,
or a system exploring a dataset,
every once in a while asking for human input,
do you see value in that kind of work?
- I don't see transformative value.
It's going to make things that we can already do
more efficient, or they will learn slightly more efficiently,
but it's not going to make machines
sort of significantly more intelligent.
And by the way, there is no opposition,
there's no conflict between self-supervised learning,
reinforcement learning, and supervised learning,
or imitation learning, or active learning.
I see self-supervised learning as a preliminary
to all of the above.
- Yes.
- So the example I use very often is,
how is it that, so if you use
classical reinforcement learning,
deep reinforcement learning, if you want,
the best methods today,
so-called model-free reinforcement learning,
to learn to play Atari games,
take about 80 hours of training to reach the level
that any human can reach in about 15 minutes.
They get better than humans, but it takes them a long time.
AlphaStar, okay, the, you know,
Aurelien Wigner and his team's system to play StarCraft,
plays, you know, a single map, a single type of player,
and can reach better than human level
with about the equivalent of 200 years of training
playing against itself.
It's 200 years, right?
It's not something that no human can, could ever do.
- I mean, I'm not sure what lesson to take away from that.
- Okay, now take those algorithms,
the best Aurel algorithms we have today,
to train a car to drive itself.
It would probably have to drive millions of hours.
It will have to kill thousands of pedestrians.
It will have to run into thousands of trees.
It will have to run off cliffs.
And it had to run off cliffs multiple times
before it figures out that it's a bad idea, first of all.
And second of all, before it figures out how not to do it.
And so, I mean, this type of learning, obviously,
does not reflect the kind of learning
that animals and humans do.
There is something missing
that's really, really important there.
And my hypothesis,
which I've been advocating for like five years now,
is that we have predictive models of the world
that include the ability to predict under uncertainty.
And what allows us to not run off a cliff
when we learn to drive,
most of us can learn to drive
in about 20 or 30 hours of training
without ever crashing, causing any accident.
And if we drive next to a cliff,
we know that if we turn the wheel to the right,
the car is gonna run off the cliff
and nothing good is gonna come out of this.
Because we have a pretty good model of intuitive physics
that tells us the car is gonna fall.
We know about gravity.
Babies learn this around the age of eight or nine months,
that objects don't float, they fall.
And we have a pretty good idea
of the effect of turning the wheel of the car.
And we know we need to stay on the road.
So there's a lot of things that we bring to the table,
which is basically our predictive model of the world.
And that model allows us to not do stupid things
and to basically stay within the context
of things we need to do.
We still face unpredictable situations
and that's how we learn,
but that allows us to learn really, really, really quickly.
So that's called model-based reinforcement learning.
There's some imitation and supervised learning
because we have a driving instructor
that tells us occasionally what to do.
But most of the learning
is learning the model, learning physics
that we've done since we were babies.
That's where almost all the learning-
- And the physics is somewhat transferable from,
is transferable from scene to scene.
Stupid things are the same everywhere.
- Yeah.
I mean, if you have experience of the world,
you don't need to be from a particularly intelligent species
to know that if you spill water from a container,
the rest is gonna get wet.
And you might get wet.
So, you know, cats know this, right?
- Yeah.
- So the main problem we need to solve
is how do we learn models of the world?
That's, and that's what I'm interested in.
That's what self-supervised learning is all about.
- If you were to try to construct a benchmark for,
let's look at MNIST.
I love that dataset.
Do you think it's useful, interesting/possible
to perform well on MNIST
with just one example of each digit?
And how would we solve that problem?
- The answer is probably yes.
The question is what other type of learning
are you allowed to do?
So if what you're allowed to do
is train on some gigantic dataset of labeled digit,
that's called transfer learning.
And we know that works, okay?
We do this at Facebook, like in production, right?
We train large convolutional nets
to predict hashtags that people type on Instagram.
And we train on billions of images, literally billions.
And then we chop off the last layer
and fine tune on whatever task we want.
That works really well.
You can beat the ImageNet record with this.
We actually open-sourced the whole thing
like a few weeks ago.
- Yeah, that's still pretty cool.
But yeah, so what would be impressive?
And what's useful and impressive?
What kind of transfer learning would be useful and impressive?
Is it Wikipedia?
That kind of thing?
- No, no, so I don't think transfer learning
is really where we should focus.
We should try to do,
you know, have a kind of scenario for benchmark
where you have unlabeled data
and you can, and it's very large number of unlabeled data.
It could be video clips.
It could be where you do, you know, frame prediction.
It could be images where you could choose to,
you know, mask a piece of it.
Could be whatever, but they're unlabeled
and you're not allowed to label them.
So you do some training on this
and then you train on a particular supervised task,
ImageNet or NIST,
and you measure how your test error decrease
or validation error decreases
as you increase the number of labeled training samples.
Okay, and what you'd like to see is that,
you know, your error decreases much faster
than if you train from scratch, from random weights.
So that to reach the same level of performance
in a completely supervised, purely supervised system
would reach, you would need way fewer samples.
So that's the crucial question
because it will answer the question to like, you know,
people interested in medical image analysis.
Okay, you know, if I want to get to a particular
level of error rate for this task,
I know I need a million samples.
Can I do, you know, self-supervised pre-training
to reduce this to about a hundred or something?
- And you think the answer there
is self-supervised pre-training?
- Yeah, some form, some form of it.
- Telling you, active learning, but you disagree.
- No, it's not useless.
It's just not gonna lead to a quantum leap.
It's just gonna make things that we already do.
- So you're way smarter than me.
I just disagree with you.
But I don't have anything to back that.
It's just intuition.
So I worked with a lot of large scale datasets
§08Transcript · 59:42–1:07:22
and there's something that might be magic
in active learning, but okay.
Now at least I said it publicly.
(laughs)
At least I'm being an idiot publicly.
Okay.
- It's not being an idiot.
It's, you know, working with the data you have.
I mean, certainly people are doing things like,
okay, I have 3000 hours of, you know,
imitation learning for self-driving car,
but most of those are incredibly boring.
What I like is select, you know, 10% of them
that are kind of the most informative.
And with just that, I would probably reach the same.
So it's a weak form of active learning, if you want.
- Yes, but there might be a much stronger version.
- Yeah, that's right.
- That's what, and that's an open question if it exists.
The question is how much stronger it can get.
Elon Musk is confident.
Talked to him recently.
He's confident that large scale data
and deep learning can solve the autonomous driving problem.
What are your thoughts on the limits,
possibilities of deep learning in this space?
- It's obviously part of the solution.
I mean, I don't think we'll ever have a self-driving system,
or at least not in the foreseeable future,
that does not use deep learning.
Let me put it this way.
Now, how much of it?
So in the history of sort of engineering,
particularly sort of AI-like systems,
there's generally a first phase
where everything is built by hand.
Then there is a second phase,
and that was the case for autonomous driving,
you know, 20, 30 years ago.
There's a phase where there's,
a little bit of learning is used,
but there's a lot of engineering that's involved
in kind of, you know, taking care of corner cases
and putting limits, et cetera,
because the learning system is not perfect.
And then as technology progresses,
we end up relying more and more on learning.
That's the history of character recognition,
it's the history of speech recognition,
now computer vision, natural language processing.
And I think the same is going to happen
with autonomous driving, that currently
the methods that are closest to providing
some level of autonomy, some, you know,
decent level of autonomy, where you don't expect
a driver to kind of do anything,
is where you constrain the world.
So you only run within, you know,
100 square kilometers or square miles in Phoenix,
but the weather is nice and the roads are wide,
which is what Waymo is doing.
You completely over-engineer the car
with tons of lidars and sophisticated sensors
that are too expensive for consumer cars,
but they're fine if you just run a fleet.
And you engineer the hell out of everything else,
you map the entire world,
so you have a complete 3D model of everything.
So the only thing that the perception system
has to take care of is moving objects
and construction and sort of, you know,
things that weren't in your map.
And you can engineer a good, you know,
SLAM system and all that stuff, right?
So that's kind of the current approach
that's closest to some level of autonomy,
but I think eventually the long-term solution
is gonna rely more and more on learning
and possibly using a combination
of self-supervised learning and model-based reinforcement
or something like that.
- But ultimately learning will be not just at the core,
but really the fundamental part of the system.
- Yeah, it already is, but it'll become more and more.
- What do you think it takes to build a system
with human level intelligence?
You talked about the AI system in the movie "Her"
being way out of reach, our current reach.
This might be outdated as well, but-
- It's still way out of reach.
- It's still way out of reach.
What would it take to build "Her"?
Do you think?
- So I can tell you the first two obstacles
that we have to clear,
but I don't know how many obstacles there are after this.
So the image I usually use is that
there is a bunch of mountains that we have to climb
and we can see the first one,
but we don't know if there are 50 mountains
behind it or not.
And this might be a good sort of metaphor
for why AI researchers in the past
have been overly optimistic about the result of AI.
For example, Newell and Simon
wrote the general problem solver
and they called it a general problem solver.
- General problem solver.
- And of course, the first thing you realize
is that all the problems you want to solve are exponential.
And so you can't actually use it for anything useful.
But you know.
- Yeah, so yeah, all you see is the first peak.
So what are the first couple of peaks for "Her"?
- So the first peak,
which is precisely what I'm working on,
is self-supervised learning.
How do we get machines to learn models of the world
by observation,
kind of like babies and like young animals?
So we've been working with cognitive scientists.
So this Emmanuelle Dupoux, who is at FAIR in Paris,
is a half-time, is also a researcher in French University.
And he has this chart that shows
which, how many months of life
baby humans can learn different concepts.
And you can measure this in sort of various ways.
So things like distinguishing animate objects
from inanimate objects,
you can tell the difference at age two, three months.
Whether an object is going to stay stable,
is going to fall, you know,
about four months, you can tell.
You know, there are various things like this.
And then things like gravity,
the fact that objects are not supposed to float in the air,
but are supposed to fall,
you learn this around the age of eight or nine months.
If you look at a lot of, you know, eight-month-old babies,
you give them a bunch of toys on their high chair.
First thing they do is they throw them on the ground
and they look at them.
It's because, you know,
they're learning about, actively learning
about gravity. - Gravity, yeah.
- Okay, so they're not trying to annoy you,
but they, you know, they need to do the experiment, right?
So, you know, how do we get machines to learn like babies?
Mostly by observation with a little bit of interaction
and learning those models of the world,
because I think that's really a crucial piece
of an intelligent autonomous system.
So if you think about the architecture
of an intelligent autonomous system,
it needs to have a predictive model of the world.
So something that says, here is a world at time T,
here is a state of the world at time T plus one
if I take this action.
And it's not a single answer, it can be a--
- Yeah, it can be a distribution, yeah.
- Yeah, well, but we don't know how to represent
distributions in high dimensional space.
So it's gotta be something weaker than that, okay?
But with some representation of uncertainty.
If you have that, then you can do
what optimal control theorists call
model predictive control,
which means that you can run your model
with a hypothesis for a sequence of action
and then see the result.
Now, what you need, the other thing you need
is some sort of objective that you want to optimize.
Am I reaching the goal of grabbing this object?
Am I minimizing energy?
Am I whatever, right?
So there is some sort of objective
that you have to minimize.
And so in your head, if you have this model,
you can figure out the sequence of action
that will optimize your objective.
That objective is something that ultimately
is rooted in your basal ganglia,
at least in the human brain,
that's what it's basal ganglia,
computes your level of contentment or miscontentment.
I don't know if that's a word.
Unhappiness, okay?
- Yeah, yeah.
- Discontentment.
- Discontentment, maybe.
- And so your entire behavior is driven
towards minimizing that objective,
which is maximizing your contentment,
computed by your basal ganglia.
And what you have is an objective function,
which is basically a predictor
of what your basal ganglia is going to tell you.
So you're not going to put your hand on fire
because you know it's going to burn.
And you're going to get hurt.
§09Transcript · 1:07:23–1:15:38
And you're predicting this because of your model of the world
and your sort of predictor of this objective, right?
So if you have those three components,
you have four components,
you have the hardwired contentment objective computer,
if you want, calculator.
And then you have those three components.
One is the objective predictor,
which basically predicts your level of contentment.
One is the model of the world.
And there's a third module I didn't mention,
which is the module that will figure out
the best course of action
to optimize an objective given your model.
Okay?
- Yeah.
- Call this a policy network or something like that, right?
Now, you need those three components
to act autonomously, intelligently.
And you can be stupid in three different ways.
You can be stupid because your model of the world is wrong.
You can be stupid because your objective is not aligned
with what you actually want to achieve.
Okay?
In humans, that would be a psychopath.
- Right.
- And then the third way you can be stupid
is that you have the right model,
you have the right objective,
but you're unable to figure out a course of action
to optimize your objective given your model.
- Right.
- Okay.
Some people who are in charge of big countries
actually have all three that are wrong.
- All right.
(laughs)
Which countries?
I don't know.
Okay, so if we think about this agent,
if we think about the movie "Her,"
you've criticized the art project that is Sophia the Robot.
And what that project essentially does
is uses our natural inclination to anthropomorphize
things that look like human and give them more.
Do you think that could be used by AI systems
like in the movie "Her"?
So do you think that body is needed
to create a feeling of intelligence?
- Well, if Sophia was just an art piece,
I would have no problem with it,
but it's presented as something else.
- Let me add that comment real quick.
If creators of Sophia could change something
about their marketing or behavior in general,
what would it be?
What's--
- I'm just about everything.
(laughs)
- I mean, don't you think, here's a tough question.
Let me, so I agree with you.
So Sophia is not, the general public
feels that Sophia can do way more than she actually can.
- That's right.
- And the people who created Sophia
are not honestly publicly communicating,
trying to teach the public.
- Right.
- But here's a tough question.
Don't you think the same thing
is scientists in industry and research
are taking advantage of the same misunderstanding
in the public when they create AI companies
or publish stuff?
- Some companies, yes.
I mean, there is no sense of,
there's no desire to delude.
There's no desire to kind of overclaim
what something is done.
Right, you publish a paper on AI
that has this result on ImageNet,
it's pretty clear.
I mean, it's not even interesting anymore,
but I don't think there is that.
I mean, the reviewers are generally not very forgiving
of unsupported claims of this type.
And, but there are certainly quite a few startups
that have had a huge amount of hype around this
that I find extremely damaging
and I've been calling it out when I've seen it.
So yeah, but to go back to your original question,
like the necessity of embodiment.
I think, I don't think embodiment is necessary.
I think grounding is necessary.
So I don't think we're gonna get machines
that really understand language
without some level of grounding in the real world.
And it's not clear to me that language
is a high enough bandwidth medium
to communicate how the real world works.
I think for this--
- Can you talk to ground, what grounding means?
- So grounding means that,
so there is this classic problem of common sense reasoning,
you know, the Winograd schema, right?
And so I tell you the trophy doesn't fit in a suitcase
because it's too big,
or the trophy doesn't fit in a suitcase
because it's too small.
And the it in the first case refers to the trophy
in the second case to the suitcase.
And the reason you can figure this out
is because you know what the trophy and the suitcase are,
you know, one is supposed to fit in the other one,
and you know the notion of size
and the big object doesn't fit in a small object
unless it's a TARDIS, you know, things like that, right?
So you have this knowledge of how the world works,
of geometry and things like that.
I don't believe you can learn everything about the world
by just being told in language how the world works.
I think you need some low-level perception of the world,
you know, be it visual touch, you know, whatever,
but some higher bandwidth perception of the world.
- So by reading all the world's text,
you still may not have enough information.
- That's right.
There's a lot of things that just will never appear in text
and that you can't really infer.
So I think common sense will emerge from, you know,
certainly a lot of language interaction,
but also with watching videos
or perhaps even interacting in virtual environments
and possibly, you know, robot interacting in the real world.
But I don't actually believe necessarily
that this last one is absolutely necessary,
but I think there's a need for some grounding.
- But the final product doesn't necessarily
need to be embodied, you're saying.
It just needs to have an awareness, a grounding.
- Right, but it needs to know how the world works
to have, you know, to not be frustrating to talk to.
- And you talked about emotions being important.
That's a whole nother topic.
- Well, so, you know, I talked about this,
the basal ganglia as the, you know,
the thing that calculates your level of miscontentment,
and then there is this other module
that sort of tries to do a prediction
of whether you're gonna be content or not.
That's the source of some emotion.
So fear, for example, is an anticipation
of bad things that can happen to you, right?
You have this inkling that there is some chance
that something really bad is gonna happen to you,
and that creates fear.
When you know for sure that something bad
is gonna happen to you, you kind of give up, right?
It's not there anymore.
It's uncertainty that creates fear.
- So the punchline is, we're not gonna have
autonomous intelligence without emotions.
- Whatever the heck emotions are.
So you mentioned very practical things of fear,
but there's a lot of other mess around it.
- But they are kind of the results of, you know, drives.
- Yeah, there's deeper biological stuff going on,
and I've talked to a few folks on this.
There's fascinating stuff that ultimately
connects to our brain.
If we create an AGI system, sorry.
- Human level intelligence.
- Human level intelligence system,
and you get to ask her one question,
what would that question be?
- You know, I think the first one we'll create
will probably not be that smart.
They'd be like a four year old.
- So you would have to ask her a question
to know she's not that smart?
- Yeah.
- Well, what's a good question to ask, you know,
to be impressed? - What is the cause of wind?
And if she answers, oh, it's because the leaves
of the tree are moving and that creates wind,
she's onto something.
- And if she says that's a stupid question,
she's really onto something.
- No, and then you tell her, actually, you know,
here is the real thing, and she says,
oh, yeah, that makes sense.
- So questions that reveal the ability
to do common sense reasoning about the physical world.
- Yeah, and you'll sum it up with a causal inference.
- Causal inference.
Well, it was a huge honor.
Congratulations on the Turing Award.
Thank you so much for talking today.
- Thank you. - Appreciate it.
(upbeat music)
§10Transcript · 1:15:41–1:15:51
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)