back to indexJohn Hopfield: Physics View of the Mind and Neurobiology | Lex Fridman Podcast #76
link |
The following is a conversation with John Hopfield,
link |
professor at Princeton, whose life's work weaved beautifully
link |
through biology, chemistry, neuroscience, and physics.
link |
Most crucially, he saw the messy world of biology
link |
through the piercing eyes of a physicist.
link |
He's perhaps best known for his work
link |
on associative neural networks,
link |
now known as Hopfield networks,
link |
that were one of the early ideas that catalyzed
link |
the development of the modern field of deep learning.
link |
As his 2019 Franklin Medal in Physics Award states,
link |
he applied concepts of theoretical physics
link |
to provide new insights on important biological questions
link |
in a variety of areas, including genetics and neuroscience
link |
with significant impact on machine learning.
link |
And as John says in his 2018 article titled,
link |
Now What?, his accomplishments have often come about
link |
by asking that very question, now what?
link |
And often responding by a major change of direction.
link |
This is the Artificial Intelligence Podcast.
link |
If you enjoy it, subscribe on YouTube,
link |
give it five stars on Apple Podcast,
link |
support it on Patreon, or simply connect with me on Twitter,
link |
and Lex Friedman, spelled F R I D M A M.
link |
As usual, I'll do one or two minutes of ads now
link |
and never any ads in the middle
link |
that can break the flow of the conversation.
link |
I hope that works for you
link |
and doesn't hurt the listening experience.
link |
This show is presented by Cash App,
link |
the number one finance app in the App Store.
link |
When you get it, use code LexPodcast.
link |
Cash App lets you send money to friends, buy Bitcoin,
link |
and invest in the stock market with as little as $1.
link |
Since Cash App does fractional share trading,
link |
let me mention that the order execution algorithm
link |
that works behind the scenes
link |
to create the abstraction of fractional orders
link |
is to me an algorithmic marvel.
link |
So big props to the Cash App engineers
link |
for solving a hard problem
link |
that in the end provides an easy interface
link |
that takes a step up the next layer of abstraction
link |
over the stock market,
link |
making trading more accessible for new investors
link |
and diversification much easier.
link |
So again, if you get Cash App from the App Store,
link |
Google Play, and use code LexPodcast,
link |
and Cash App will also donate $10 to First,
link |
one of my favorite organizations
link |
that is helping advance robotics and STEM education
link |
for young people around the world.
link |
And now here's my conversation with John Hopfield.
link |
What difference between biological neural networks
link |
and artificial neural networks
link |
is most captivating and profound to you?
link |
At the higher philosophical level,
link |
let's not get technical just yet.
link |
But one of the things that very much intrigues me
link |
is the fact that neurons have all kinds of components,
link |
properties to them.
link |
And in evolutionary biology,
link |
if you have some little quirk
link |
in how a molecule works or how a cell works,
link |
and it can be made use of,
link |
evolution will sharpen it up
link |
and make it into a useful feature rather than a glitch.
link |
And so you expect in neurobiology for evolution
link |
to have captured all kinds of possibilities
link |
of getting neurons,
link |
of how you get neurons to do things for you.
link |
And that aspect has been completely suppressed
link |
in artificial neural networks.
link |
So the glitches become features
link |
in the biological neural network.
link |
Look, let me take one of the things
link |
that I used to do research on.
link |
If you take things which oscillate,
link |
they have rhythms which are sort of close to each other.
link |
Under some circumstances,
link |
these things will have a phase transition
link |
and suddenly the rhythm will,
link |
everybody will fall into step.
link |
There was a marvelous physical example of that
link |
in the Millennium Bridge across the Thames River,
link |
about, built about 2001.
link |
And pedestrians walking across,
link |
pedestrians don't walk synchronized,
link |
they don't walk in lockstep.
link |
But they're all walking about the same frequency
link |
and the bridge could sway at that frequency
link |
and the slight sway made pedestrians tend a little bit
link |
to lock into step and after a while,
link |
the bridge was oscillating back and forth
link |
and the pedestrians were walking in step to it.
link |
And you could see it in the movies made out of the bridge.
link |
And the engineers made a simple minor mistake.
link |
They assume when you walk, it's step, step, step
link |
and it's back and forth motion.
link |
But when you walk, it's also right foot left
link |
with side to side motion.
link |
And it's the side to side motion
link |
for which the bridge was strong enough,
link |
but it wasn't stiff enough.
link |
And as a result, you would feel the motion
link |
and you'd fall into step with it.
link |
And people were very uncomfortable with it.
link |
They closed the bridge for two years
link |
while they built stiffening for it.
link |
Now, nerve cells produce action potentials.
link |
You have a bunch of cells which are loosely coupled together
link |
producing action potentials at the same rate.
link |
There'll be some circumstances
link |
under which these things can lock together.
link |
Other circumstances in which they won't.
link |
Well, if they're fired together,
link |
you can be sure that other cells are gonna notice it.
link |
So you can make a computational feature out of this
link |
in an evolving brain.
link |
Most artificial neural networks
link |
don't even have action potentials,
link |
let alone have the possibility for synchronizing them.
link |
And you mentioned the evolutionary process.
link |
So the evolutionary process
link |
that builds on top of biological systems
link |
leverages the weird mess of it somehow.
link |
So how do you make sense of that ability
link |
to leverage all the different kinds of complexities
link |
in the biological brain?
link |
Well, look, in the biological molecule level,
link |
you have a piece of DNA
link |
which encodes for a particular protein.
link |
You could duplicate that piece of DNA
link |
and now one part of it can code for that protein,
link |
but the other one could itself change a little bit
link |
and thus start coding for a molecule
link |
which is slightly different.
link |
Now, if that molecule was just slightly different,
link |
had a function which helped any old chemical reaction
link |
which was important to the cell,
link |
you would go ahead and let that try,
link |
and evolution would slowly improve that function.
link |
And so you have the possibility of duplicating
link |
and then having things drift apart.
link |
One of them retain the old function,
link |
the other one do something new for you.
link |
And there's evolutionary pressure to improve.
link |
Look, there isn't in computers too,
link |
but improvement has to do with closing some companies
link |
and opening some others.
link |
The evolutionary process looks a little different.
link |
Yeah, similar timescale perhaps.
link |
Much shorter in timescale.
link |
Companies close, yeah, go bankrupt and are born,
link |
yeah, shorter, but not much shorter.
link |
Some companies last a century, but yeah, you're right.
link |
I mean, if you think of companies as a single organism
link |
that builds and you all know, yeah,
link |
it's a fascinating dual correspondence there
link |
between biological organisms.
link |
And companies have difficulty having a new product
link |
competing with an old product.
link |
When IBM built its first PC, you probably read the book,
link |
they made a little isolated internal unit to make the PC.
link |
And for the first time in IBM's history,
link |
they didn't insist that you build it out of IBM components.
link |
But they understood that they could get into this market,
link |
which is a very different thing
link |
by completely changing their culture.
link |
And biology finds other markets in a more adaptive way.
link |
Yeah, it's better at it.
link |
It's better at that kind of integration.
link |
So maybe you've already said it,
link |
but what to use the most beautiful aspect
link |
or mechanism of the human mind?
link |
Is it the adaptive, the ability to adapt
link |
as you've described, or is there some other little quirk
link |
that you particularly like?
link |
Adaptation is everything when you get down to it.
link |
But the difference, there are differences between adaptation
link |
where your learning goes on only over generations
link |
and over evolutionary time,
link |
where your learning goes on at the time scale
link |
of one individual who must learn from the environment
link |
during that individual's lifetime.
link |
And biology has both kinds of learning in it.
link |
And the thing which makes neurobiology hard
link |
is that a mathematical system, as it were,
link |
built on this other kind of evolutionary system.
link |
What do you mean by mathematical system?
link |
Where's the math and the biology?
link |
Well, when you talk to a computer scientist
link |
about neural networks, it's all math.
link |
The fact that biology actually came about from evolution,
link |
and the fact that biology is about a system
link |
which you can build in three dimensions.
link |
If you look at computer chips,
link |
computer chips are basically two dimensional structures,
link |
maybe 2.1 dimensions, but they really have difficulty
link |
doing three dimensional wiring.
link |
Biology is, the neocortex is actually also sheet like,
link |
and it sits on top of the white matter,
link |
which is about 10 times the volume of the gray matter
link |
and contains all what you might call the wires.
link |
But there's a huge, the effect of computer structure
link |
on what is easy and what is hard is immense.
link |
And biology does, it makes some things easy
link |
that are very difficult to understand
link |
how to do computationally.
link |
On the other hand, you can't do simple floating point
link |
arithmetic because it's awfully stupid.
link |
And you're saying this kind of three dimensional
link |
complicated structure makes, it's still math.
link |
It's still doing math.
link |
The kind of math it's doing enables you to solve problems
link |
of a very different kind.
link |
That's right, that's right.
link |
So you mentioned two kinds of adaptation,
link |
the evolutionary adaptation and the adaptation
link |
or learning at the scale of a single human life.
link |
Which do you, which is particularly beautiful to you
link |
and interesting from a research
link |
and from just a human perspective?
link |
And which is more powerful?
link |
I find things most interesting that I begin to see
link |
how to get into the edges of them
link |
and tease them apart a little bit and see how they work.
link |
And since I can't see the evolutionary process going on,
link |
But I find it just a black hole as far as trying
link |
to understand what to do.
link |
And so in a certain sense, I'm in awe of it,
link |
but I couldn't be interested in working on it.
link |
The human life's time scale is however thing
link |
you can tease apart and study.
link |
Yeah, you can do, there's developmental neurobiology
link |
which understands how the connections
link |
and how the structure evolves from a combination
link |
of what the genetics is like and the real,
link |
the fact that you're building a system in three dimensions.
link |
In just days and months, those early days
link |
of a human life are really interesting.
link |
They are and of course, there are times
link |
of immense cell multiplication.
link |
There are also times of the greatest cell death
link |
in the brain is during infancy.
link |
So what is not effective, what is not wired well enough
link |
to use at the moment, throw it out.
link |
It's a mysterious process.
link |
From, let me ask, from what field do you think
link |
the biggest breakthrough is in understanding
link |
the mind will come in the next decades?
link |
Is it neuroscience, computer science, neurobiology,
link |
psychology, physics, maybe math, maybe literature?
link |
Well, of course, I see the world always
link |
through a lens of physics.
link |
I grew up in physics and the way I pick problems
link |
is very characteristic of physics
link |
and of an intellectual background which is not psychology,
link |
which is not chemistry and so on and so on.
link |
Yeah, both of your parents are physicists.
link |
Both of my parents were physicists
link |
and the real thing I got out of that was a feeling
link |
that the world is an understandable place
link |
and if you do enough experiments and think about
link |
what they mean and structure things
link |
so you can do the mathematics of the,
link |
relevant to the experiments, you ought to be able
link |
to understand how things work.
link |
But that was, that was a few years ago.
link |
Did you change your mind at all through many decades
link |
of trying to understand the mind,
link |
of studying in different kinds of ways?
link |
Not even the mind, just biological systems.
link |
You still have hope that physics, that you can understand?
link |
There's a question of what do you mean by understand?
link |
When I taught freshman physics, I used to say,
link |
I wanted to get physics to understand the subject,
link |
to understand Newton's laws.
link |
I didn't want them simply to memorize a set of examples
link |
to which they knew the equations to write down
link |
to generate the answers.
link |
I had this nebulous idea of understanding
link |
so that if you looked at a situation,
link |
you could say, oh, I expect the ball to make that trajectory
link |
or I expect some intuitive notion of understanding
link |
and I don't know how to express that very well
link |
and I've never known how to express it well.
link |
And you run smack up against it when you do these,
link |
look at these simple neural nets,
link |
feed forward neural nets, which do amazing things
link |
and yet, you know, contain nothing of the essence
link |
of what I would have felt was understanding.
link |
Understanding is more than just an enormous lookup table.
link |
Let's linger on that.
link |
How sure you are of that?
link |
What if the table gets really big?
link |
So, I mean, asked another way,
link |
these feed forward neural networks,
link |
do you think they'll ever understand?
link |
Could answer that in two ways.
link |
I think if you look at real systems,
link |
feedback is an essential aspect
link |
of how these real systems compute.
link |
On the other hand, if I have a mathematical system
link |
with feedback, I know I can unlayer this and do it,
link |
but I have an exponential expansion
link |
in the amount of stuff I have to build
link |
if I can resolve the problem that way.
link |
So feedback is essential.
link |
So we can talk even about recurrent neural nets,
link |
so recurrence, but do you think all the pieces are there
link |
to achieve understanding through these simple mechanisms?
link |
Like back to our original question,
link |
what is the fundamental, is there a fundamental difference
link |
between artificial neural networks and biological
link |
or is it just a bunch of surface stuff?
link |
Suppose you ask a neurosurgeon, when is somebody dead?
link |
So we'll probably go back to saying,
link |
well, I can look at the brain rhythms
link |
and tell you this is a brain
link |
which has never could have functioned again.
link |
This is one of the, this other one is one of the stuff
link |
we treat it well is still recoverable.
link |
And then just do that by some electrodes
link |
looking at simple electrical patterns,
link |
which don't look in any detail at all
link |
what individual neurons are doing.
link |
These rhythms are utterly absent
link |
from anything which goes on at Google.
link |
Yeah, but the rhythms.
link |
But the rhythms what?
link |
So, well, that's like comparing, okay, I'll tell you,
link |
it's like you're comparing the greatest classical musician
link |
in the world to a child first learning to play.
link |
The question I'm at, but they're still both
link |
playing the piano.
link |
I'm asking, is there, will it ever go on at Google?
link |
Do you have a hope?
link |
Because you're one of the seminal figures
link |
in both launching both disciplines,
link |
both sides of the river.
link |
I think it's going to go on generation after generation.
link |
The way it has where what you might call
link |
the AI computer science community says,
link |
let's take the following.
link |
This is our model of neurobiology at the moment.
link |
Let's pretend it's good enough
link |
and do everything we can with it.
link |
And it does interesting things.
link |
And after a while it sort of grinds into the sand
link |
and you say, ah, something else is needed for neurobiology.
link |
And some other grand thing comes in
link |
and enables you to go a lot further.
link |
What will go into the sand again?
link |
And I think it could be generations of this evolution.
link |
I don't know how many of them.
link |
And each one is going to get you further
link |
into what a brain does.
link |
And in some sense, past the Turing test longer
link |
and in more broad aspects.
link |
And how many of these are going to have to be
link |
before you say, I've made something,
link |
I've made a human, I don't know.
link |
But your sense is it might be a couple.
link |
My sense is it might be a couple more.
link |
And going back to my brainwaves as it were.
link |
Yes, from the AI point of view,
link |
they would say, ah, maybe these are an epiphenomenon
link |
and not important at all.
link |
The first car I had, a real wreck of a 1936 Dodge,
link |
go above about 45 miles an hour and the wheels would shimmy.
link |
Good speedometer that.
link |
Now, nobody designed the car that way.
link |
The car is malfunctioning to have that.
link |
But in biology, if it were useful to know
link |
when are you going more than 45 miles an hour,
link |
you just capture that.
link |
And you wouldn't worry about where it came from.
link |
It's going to be a long time before that kind of thing,
link |
which can take place in large complex networks of things
link |
is actually used in the computation.
link |
Look, how many transistors are there
link |
in your laptop these days?
link |
Actually, I don't know the number.
link |
It's on the scale of 10 to the 10.
link |
I can't remember the number either.
link |
And all the transistors are somewhat similar.
link |
And most physical systems with that many parts,
link |
all of which are similar, have collective properties.
link |
Sound waves in air, earthquakes,
link |
what have you, have collective properties.
link |
There are no collective properties used
link |
in artificial neural networks, in AI.
link |
If biology uses them,
link |
it's going to take us to more generations of things
link |
for people to actually dig in
link |
and see how they are used and what they mean.
link |
See, you're very right.
link |
We might have to return several times to neurobiology
link |
and try to make our transistors more messy.
link |
At the same time, the simple ones will conquer big aspects.
link |
And I think one of the most, biggest surprises to me was
link |
how well learning systems
link |
because they're manifestly nonbiological,
link |
how important they can be actually,
link |
and how important and how useful they can be in AI.
link |
So if we can just take a stroll to some of your work.
link |
If we can just take a stroll to some of your work
link |
that is incredibly surprising,
link |
that it works as well as it does,
link |
that launched a lot of the recent work with neural networks.
link |
If we go to what are now called Hopfield networks,
link |
can you tell me what is associative memory in the mind
link |
for the human side?
link |
Let's explore memory for a bit.
link |
Okay, what do you mean by associative memory is,
link |
ah, you have a memory of each of your friends.
link |
Your friend has all kinds of properties
link |
from what they look like, what their voice sounds like,
link |
to where they went to college, where you met them,
link |
go on and on, what science papers they've written.
link |
And if I start talking about a 5 foot 10 wire,
link |
cognitive scientist who's got a very bad back,
link |
it doesn't take very long for you to say,
link |
oh, he's talking about Jeff Hinton.
link |
I never mentioned the name or anything very particular.
link |
But somehow a few facts that are associated
link |
with a particular person enables you to get a hold
link |
of the rest of the facts.
link |
Or not the rest of them, another subset of them.
link |
And it's this ability to link things together,
link |
link experiences together, which goes under
link |
the general name of associative memory.
link |
And a large part of intelligent behavior
link |
is actually just large associative memories at work,
link |
as far as I can see.
link |
What do you think is the mechanism of how it works?
link |
What do you think is the mechanism of how it works
link |
Is it a mystery to you still?
link |
Do you have inklings of how this essential thing
link |
for cognition works?
link |
What I made 35 years ago was, of course,
link |
a crude physics model to actually enable you
link |
to understand my old sense of understanding
link |
as a physicist, because you could say,
link |
ah, I understand why this goes to stable states.
link |
It's like things going downhill.
link |
And that gives you something with which to think
link |
in physical terms rather than only in mathematical terms.
link |
So you've created these associative artificial networks.
link |
Now, if you look at what I did,
link |
I didn't at all describe a system which gracefully learns.
link |
I described a system in which you could understand
link |
how learning could link things together,
link |
how very crudely it might learn.
link |
One of the things which intrigues me
link |
as I reinvestigate that system now to some extent is,
link |
look, I see you, I'll see you every second
link |
for the next hour or what have you.
link |
Each look at you is a little bit different.
link |
I don't store all those second by second images.
link |
I don't store 3,000 images.
link |
I somehow compact this information.
link |
So I now have a view of you,
link |
It doesn't slavishly remember anything in particular,
link |
but it compacts the information into useful chunks,
link |
which are somehow these chunks,
link |
which are not just activities of neurons,
link |
bigger things than that,
link |
which are the real entities which are useful to you.
link |
Which are useful to you.
link |
Useful to you to describe,
link |
to compress this information coming at you.
link |
And you have to compress it in such a way
link |
that if the information comes in just like this again,
link |
I don't bother to rewrite it or efforts to rewrite it
link |
simply do not yield anything
link |
because those things are already written.
link |
And that needs to be not,
link |
look this up, have I stored it somewhere already?
link |
There'll be something which is much more automatic
link |
in the machine hardware.
link |
Right, so in the human mind,
link |
how complicated is that process do you think?
link |
So you've created,
link |
feels weird to be sitting with John Hotfield
link |
calling him Hotfield Networks, but.
link |
Yeah, but nevertheless, that's what everyone calls him.
link |
So that's a simplification.
link |
That's what a physicist would do.
link |
You and Richard Feynman sat down
link |
and talked about associative memory.
link |
Now, if you look at the mind
link |
where you can't quite simplify it so perfectly,
link |
do you think that?
link |
Well, let me backtrack just a little bit.
link |
Biology is about dynamical systems.
link |
Computers are dynamical systems.
link |
You can ask, if you want to model biology,
link |
if you want to model neurobiology,
link |
what is the time scale?
link |
There's a dynamical system in which,
link |
of a fairly fast time scale in which you could say,
link |
the synapses don't change much during this computation,
link |
so I'll think of the synapses fixed
link |
and just do the dynamics of the activity.
link |
Or you can say, the synapses are changing fast enough
link |
that I have to have the synaptic dynamics
link |
working at the same time as the system dynamics
link |
in order to understand the biology.
link |
Most, if you look at the feedforward artificial neural nets,
link |
they're all done as learnings.
link |
First of all, I spend some time learning, not performing,
link |
and I turn off learning and I turn off learning,
link |
and I turn off learning and I perform.
link |
That's not biology.
link |
And so as I look more deeply at neurobiology,
link |
even as associative memory,
link |
I've got to face the fact that the dynamics
link |
of the synapse change is going on all the time.
link |
And I can't just get by by saying,
link |
I'll do the dynamics of activity with fixed synapses.
link |
So the synaptic, the dynamics of the synapses
link |
is actually fundamental to the whole system.
link |
And there's nothing necessarily separating the time scales.
link |
When the time scale's gonna be separated,
link |
it's neat from the physicist's
link |
or the mathematician's point of view,
link |
but it's not necessarily true in neurobiology.
link |
So you're kind of dancing beautifully
link |
between showing a lot of respect to physics
link |
and then also saying that physics
link |
cannot quite reach the complexity of biology.
link |
So where do you land?
link |
Or do you continuously dance between the two points?
link |
I continuously dance between them
link |
because my whole notion of understanding
link |
is that you can describe to somebody else
link |
how something works in ways which are honest and believable
link |
and still not describe all the nuts and bolts in detail.
link |
I can describe weather
link |
as 10 to the 32 molecules colliding in the atmosphere.
link |
I can simulate weather that way if I have a big enough machine.
link |
I'll simulate it accurately.
link |
It's no good for understanding.
link |
If I want to understand things, I want to understand things
link |
in terms of wind patterns, hurricanes,
link |
pressure differentials, and so on,
link |
all things as they're collective.
link |
And the physicist in me always hopes
link |
that biology will have some things
link |
that can be said about it which are both true
link |
and for which you don't need all the molecular details
link |
as the molecules colliding.
link |
That's what I mean from the roots of physics,
link |
So what did, again, sorry,
link |
but Hopfield Networks help you understand
link |
what insight did give us about memory, about learning?
link |
They didn't give insights about learning.
link |
They gave insights about how things having learned
link |
could be expressed, how having learned a picture of you,
link |
a picture of you reminds me of your name.
link |
That would, but it didn't describe a reasonable way
link |
of actually doing the learning.
link |
They only said if you had previously learned
link |
the connections of this kind of pattern,
link |
would now be able to,
link |
behave in a physical way was to say,
link |
ah, if I put the part of the pattern in here,
link |
the other part of the pattern will complete over here.
link |
I could understand that physics,
link |
if the right learning stuff had already been put in.
link |
And it could understand why then putting in a picture
link |
of somebody else would generate something else over here.
link |
But it did not have a reasonable description
link |
of the learning that was going on.
link |
It did not have a reasonable description
link |
of the learning process.
link |
But even, so forget learning.
link |
I mean, that's just a powerful concept
link |
that sort of forming representations
link |
that are useful to be robust,
link |
you know, for error correction kind of thing.
link |
So this is kind of what the biology does
link |
we're talking about.
link |
Yeah, and what my paper did was simply enable you,
link |
there are lots of ways of being robust.
link |
If you think of a dynamical system,
link |
you think of a system where a path is going on in time.
link |
And if you think for a computer,
link |
there's a computational path,
link |
which is going on in a huge dimensional space
link |
of ones and zeros.
link |
And an error correction system is a system,
link |
which if you get a little bit off that trajectory,
link |
will push you back onto that trajectory again.
link |
So you get to the same answer in spite of the fact
link |
that there were things,
link |
so that the computation wasn't being ideally done
link |
all the way along the line.
link |
And there are lots of models for error correction.
link |
But one of the models for error correction is to say,
link |
there's a valley that you're following, flowing down.
link |
And if you push a little bit off the valley,
link |
just like water being pushed a little bit by a rock,
link |
it gets back and follows the course of the river.
link |
And that basically the analog
link |
in the physical system, which enables you to say,
link |
oh yes, error free computation and an associative memory
link |
are very much like things that I can understand
link |
from the point of view of a physical system.
link |
The physical system is, can be under some circumstances,
link |
an accurate metaphor.
link |
It's not the only metaphor.
link |
There are error correction schemes,
link |
which don't have a valley and energy behind them.
link |
But those are error correction schemes,
link |
which a mathematician may be able to understand,
link |
So there's the physical metaphor that seems to work here.
link |
That's right, that's right.
link |
So these kinds of networks actually led to a lot of the work
link |
that is going on now in neural networks,
link |
artificial neural networks.
link |
So the follow on work with restricted Boltzmann machines
link |
and deep belief nets followed on from these ideas
link |
of the Hopfield network.
link |
So what do you think about this continued progress
link |
of that work towards now re revigorated exploration
link |
of feed forward neural networks
link |
and recurrent neural networks
link |
and convolutional neural networks
link |
and kinds of networks that are helping solve
link |
image recognition, natural language processing,
link |
all that kind of stuff.
link |
It always intrigued me that one of the most long lived
link |
of the learning systems is the Boltzmann machine,
link |
which is intrinsically a feedback network.
link |
And with the brilliance of Hind and Sinowski
link |
to understand how to do learning in that.
link |
And it's still a useful way to understand learning
link |
and the learning that you understand in that
link |
has something to do with the way
link |
that feed forward systems work.
link |
But it's not always exactly simple
link |
to express that intuition.
link |
But it's always amuses me to see Hinton
link |
going back to the will yet again
link |
on a form of the Boltzmann machine
link |
because really that which has feedback
link |
and interesting probabilities in it
link |
is a lovely encapsulation of something in computational.
link |
Something computational?
link |
Something both computational and physical.
link |
Computational and it's very much related
link |
to feed forward networks.
link |
Physical in that Boltzmann machine learning
link |
is really learning a set of parameters
link |
for a physics Hamiltonian or energy function.
link |
What do you think about learning in this whole domain?
link |
Do you think the aforementioned guy,
link |
Jeff Hinton, all the work there with backpropagation,
link |
all the kind of learning that goes on in these networks,
link |
if we compare it to learning in the brain, for example,
link |
is there echoes of the same kind of power
link |
that backpropagation reveals
link |
about these kinds of recurrent networks?
link |
Or is it something fundamentally different
link |
going on in the brain?
link |
I don't think the brain is as deep
link |
as the deepest networks go,
link |
the deepest computer science networks.
link |
And I do wonder whether part of that depth
link |
of the computer science networks is necessitated
link |
by the fact that the only learning
link |
that's easily done on a machine is feed forward.
link |
And so there's the question of to what extent
link |
is the biology, which has some feed forward
link |
and some feed back,
link |
been captured by something which has got many more neurons
link |
but much more depth than the neurons in it.
link |
So part of you wonders if the feedback is actually
link |
more essential than the number of neurons or the depth,
link |
the dynamics of the feedback.
link |
The dynamics of the feedback.
link |
Look, if you don't have feedback,
link |
it's a little bit like a building a big computer
link |
and running it through one clock cycle.
link |
And then you can't do anything
link |
until you reload something coming in.
link |
How do you use the fact that there are multiple clock cycles?
link |
How do I use the fact that you can close your eyes,
link |
stop listening to me and think about a chessboard
link |
for two minutes without any input whatsoever?
link |
Yeah, that memory thing,
link |
that's fundamentally a feedback kind of mechanism.
link |
You're going back to something.
link |
Yes, it's hard to understand.
link |
It's hard to introspect,
link |
let alone consciousness.
link |
Oh, let alone consciousness, yes, yes.
link |
Because that's tied up in there too.
link |
You can't just put that on another shelf.
link |
Every once in a while I get interested in consciousness
link |
and then I go and I've done that for years
link |
and ask one of my betters, as it were,
link |
their view on consciousness.
link |
It's been interesting collecting them.
link |
What is consciousness?
link |
Let's try to take a brief step into that room.
link |
Well, ask Marvin Minsky,
link |
his view on consciousness.
link |
consciousness is basically overrated.
link |
It may be an epiphenomenon.
link |
After all, all the things your brain does,
link |
but they're actually hard computations
link |
you do nonconsciously.
link |
And there's so much evidence
link |
that even the simple things you do,
link |
you can make decisions,
link |
you can make committed decisions about them,
link |
the neurobiologist can say,
link |
he's now committed, he's going to move the hand left
link |
before you know it.
link |
So his view that consciousness is not,
link |
that's just like little icing on the cake.
link |
The real cake is in the subconscious.
link |
Subconscious, nonconscious.
link |
Nonconscious, what's the better word, sir?
link |
It's only that Freud captured the other word.
link |
Yeah, it's a confusing word, subconscious.
link |
Nicholas Chaiter wrote an interesting book.
link |
I think the title of it is The Mind is Flat.
link |
Flat in a neural net sense, might be flat
link |
as something which is a very broad neural net
link |
without any layers in depth,
link |
whereas a deep brain would be many layers
link |
In the same sense that if you push Minsky hard enough,
link |
he would probably have said,
link |
consciousness is your effort to explain to yourself
link |
that which you have already done.
link |
Yeah, it's the weaving of the narrative
link |
around the things that have already been computed for you.
link |
That's right, and so much of what we do
link |
for our memories of events, for example.
link |
If there's some traumatic event you witness,
link |
you will have a few facts about it correctly done.
link |
If somebody asks you about it, you will weave a narrative
link |
which is actually much more rich in detail than that
link |
based on some anchor points you have of correct things
link |
and pulling together general knowledge on the other,
link |
but you will have a narrative.
link |
And once you generate that narrative,
link |
you are very likely to repeat that narrative
link |
and claim that all the things you have in it
link |
are actually the correct things.
link |
There was a marvelous example of that
link |
in the Watergate slash impeachment era of John Dean.
link |
John Dean, you're too young to know,
link |
had been the personal lawyer of Nixon.
link |
And so John Dean was involved in the coverup
link |
and John Dean ultimately realized
link |
the only way to keep himself out of jail for a long time
link |
was actually to tell some of the truths about Nixon.
link |
And John Dean was a tremendous witness.
link |
He would remember these conversations in great detail
link |
and very convincing detail.
link |
And long afterward, some of the tapes,
link |
the secret tapes as it were from which these,
link |
Don was, Gene was recalling these conversations
link |
were published, and one found out that John Dean
link |
had a good but not exceptional memory.
link |
What he had was an ability to paint vividly
link |
and in some sense accurately the tone of what was going on.
link |
By the way, that's a beautiful description of consciousness.
link |
Do you, like where do you stand in your today?
link |
So perhaps it changes day to day,
link |
but where do you stand on the importance of consciousness
link |
in our whole big mess of cognition?
link |
Is it just a little narrative maker
link |
or is it actually fundamental to intelligence?
link |
That's a very hard one.
link |
When I asked Francis Crick about consciousness,
link |
he launched forward in a long monologue
link |
about Mendel and the peas and how Mendel knew
link |
that there was something and how biologists understood
link |
that there was something in inheritance,
link |
which was just very, very different.
link |
And the fact that inherited traits didn't just wash out
link |
into a gray, but this or this and propagated
link |
that that was absolutely fundamental to the biology.
link |
And it took generations of biologists to understand
link |
that there was genetics and it took another generation
link |
or two to understand that genetics came from DNA.
link |
But very shortly after Mendel, thinking biologists
link |
did realize that there was a deep problem about inheritance.
link |
And Francis would have liked to have said,
link |
and that's why I'm working on consciousness.
link |
But of course, he didn't have any smoking gun
link |
in the sense of Mendel.
link |
And that's the weakness of his position.
link |
If you read his book, which he wrote with Koch, I think.
link |
Yeah, Christoph Koch, yeah.
link |
I find it unconvincing for the smoking gun reason.
link |
So I'm going on collecting views without actually having taken
link |
a very strong one myself,
link |
because I haven't seen the entry point.
link |
Not seeing the smoking gun from the point of view
link |
of physics, I don't see the entry point.
link |
Whereas in neurobiology, once I understood the idea
link |
of a collective, an evolution of dynamics,
link |
which could be described as a collective phenomenon,
link |
I thought, ah, there's a point where what I know
link |
about physics is so different from any neurobiologist
link |
that I have something that I might be able to contribute.
link |
And right now, there's no way to grasp at consciousness
link |
from a physics perspective.
link |
From my point of view, that's correct.
link |
And of course, people, physicists, like everybody else,
link |
think very muddily about things.
link |
You ask the closely related question about free will.
link |
Do you believe you have free will?
link |
Physicists will give an offhand answer,
link |
and then backtrack, backtrack, backtrack,
link |
where they realize that the answer they gave
link |
must fundamentally contradict the laws of physics.
link |
Natural, answering questions of free will
link |
and consciousness naturally lead to contradictions
link |
from a physics perspective.
link |
Because it eventually ends up with quantum mechanics,
link |
and then you get into that whole mess
link |
of trying to understand how much,
link |
from a physics perspective, how much is determined,
link |
already predetermined, how much is already deterministic
link |
about our universe, and there's lots of different things.
link |
And if you don't push quite that far, you can say,
link |
essentially, all of neurobiology, which is relevant,
link |
can be captured by classical equations of motion.
link |
Right, because in my view of the mysteries of the brain
link |
are not the mysteries of quantum mechanics,
link |
but the mysteries of what can happen
link |
when you have a dynamical system, driven system,
link |
with 10 to the 14 parts.
link |
That that complexity is something which is,
link |
that the physics of complex systems
link |
is at least as badly understood
link |
as the physics of phase coherence in quantum mechanics.
link |
Can we go there for a second?
link |
You've talked about attractor networks,
link |
and just maybe you could say what are attractor networks,
link |
and more broadly, what are interesting network dynamics
link |
that emerge in these or other complex systems?
link |
You have to be willing to think
link |
in a huge number of dimensions,
link |
because in a huge number of dimensions,
link |
the behavior of a system can be thought
link |
as just the motion of a point over time
link |
in this huge number of dimensions.
link |
And an attractor network is simply a network
link |
where there is a line and other lines
link |
converge on it in time.
link |
That's the essence of an attractor network.
link |
In a highly dimensional space.
link |
And the easiest way to get that
link |
is to do it in a highly dimensional space,
link |
where some of the dimensions provide the dissipation,
link |
which, if I have a physical system,
link |
trajectories can't contract everywhere.
link |
They have to contract in some places and expand in others.
link |
There's a fundamental classical theorem
link |
of statistical mechanics,
link |
which goes under the name of Liouville's theorem,
link |
which says you can't contract everywhere.
link |
If you contract somewhere, you expand somewhere else.
link |
In interesting physical systems,
link |
you've got driven systems
link |
where you have a small subsystem,
link |
which is the interesting part.
link |
And the rest of the contraction and expansion,
link |
the physicists would say it's entropy flow
link |
in this other part of the system.
link |
But basically, attractor networks are dynamics
link |
that are funneling down so that you can't be any,
link |
so that if you start somewhere in the dynamical system,
link |
you will soon find yourself
link |
on a pretty well determined pathway, which goes somewhere.
link |
If you start somewhere else,
link |
you'll wind up on a different pathway,
link |
but I don't have just all possible things.
link |
You have some defined pathways which are allowed
link |
and onto which you will converge.
link |
And that's the way you make a stable computer,
link |
and that's the way you make a stable behavior.
link |
So in general, looking at the physics
link |
of the emergent stability in networks,
link |
what are some interesting characteristics that,
link |
what are some interesting insights
link |
from studying the dynamics of such high dimensional systems?
link |
Most dynamical systems, most driven dynamical systems,
link |
are driven, they're coupled somehow to an energy source.
link |
And so their dynamics keeps going
link |
because it's coupling to the energy source.
link |
Most of them, it's very difficult to understand at all
link |
what the dynamical behavior is going to be.
link |
You have to run it.
link |
You have to run it.
link |
There's a subset of systems which has
link |
what is actually known to the mathematicians
link |
as a Lyapunov function, and those systems,
link |
you can understand convergent dynamics
link |
by saying you're going downhill on something or other.
link |
And that's what I found with ever knowing
link |
what Lyapunov functions were in the simple model
link |
I made in the early 80s, was an energy function
link |
so you could understand how you could get this channeling
link |
on the pathways without having to follow the dynamics
link |
in infinite detail.
link |
You started rolling a ball at the top of a mountain,
link |
it's gonna wind up at the bottom of a valley.
link |
You know that's true without actually watching
link |
the ball roll down.
link |
There's certain properties of the system
link |
that when you can know that.
link |
And not all systems behave that way.
link |
Most don't, probably.
link |
Most don't, but it provides you with a metaphor
link |
for thinking about systems which are stable
link |
and who to have these attractors behave
link |
even if you can't find a Lyapunov function behind them
link |
or an energy function behind them.
link |
It gives you a metaphor for thought.
link |
Yeah, speaking of thought,
link |
if I had a glint in my eye with excitement
link |
and said I'm really excited about this something
link |
called deep learning and neural networks
link |
and I would like to create an intelligent system
link |
and came to you as an advisor, what would you recommend?
link |
Is it a hopeless pursuit to use neural networks
link |
to achieve thought?
link |
Is it, what kind of mechanisms should we explore?
link |
What kind of ideas should we explore?
link |
Well, you look at the simple networks,
link |
the one past networks.
link |
They don't support multiple hypotheses very well.
link |
As I have tried to work with very simple systems
link |
which do something which you might consider to be thinking,
link |
thought has to do with the ability to do mental exploration
link |
before you take a physical action.
link |
Almost like we were mentioning, playing chess,
link |
visualizing, simulating inside your head different outcomes.
link |
And now you would do that in a feed forward network
link |
because you've pre calculated all kinds of things.
link |
But I think the way neurobiology does it
link |
hasn't pre calculated everything.
link |
It actually has parts of a dynamical system
link |
in which you're doing exploration in a way which is.
link |
There's a creative element.
link |
There's a creative element.
link |
And in a simple minded neural net,
link |
you have a constellation of instances
link |
of which you've learned.
link |
And if you are within that space,
link |
if a new question is a question within this space,
link |
you can actually rely on that system pretty well
link |
to come up with a good suggestion for what to do.
link |
If on the other hand,
link |
the query comes from outside the space,
link |
you have no way of knowing how the system
link |
There are no limitations on what can happen.
link |
And so with the artificial neural net world
link |
is always very much,
link |
I have a population of examples.
link |
The test set must be drawn from the equivalent population.
link |
If the test set has examples,
link |
which are from a population which is completely different,
link |
there's no way that you could expect
link |
to get the answer right.
link |
Yeah, what they call outside the distribution.
link |
That's right, that's right.
link |
And so if you see a ball rolling across the street at dusk,
link |
if that wasn't in your training set,
link |
the idea that a child may be coming close behind that
link |
is not going to occur to the neural net.
link |
there's something in your biology that allows that.
link |
Yeah, there's something in the way
link |
of what it means to be outside of the population
link |
of the training set.
link |
The population of the training set
link |
isn't just sort of this set of examples.
link |
There's more to it than that.
link |
And it gets back to my question of,
link |
what is it to understand something?
link |
You know, in a small tangent,
link |
you've talked about the value of thinking
link |
of deductive reasoning in science
link |
versus large data collection.
link |
So sort of thinking about the problem.
link |
I suppose it's the physics side of you
link |
of going back to first principles and thinking,
link |
but what do you think is the value of deductive reasoning
link |
in the scientific process?
link |
Well, there are obviously scientific questions
link |
in which the route to the answer to it
link |
comes through the analysis of one hell of a lot of data.
link |
Cosmology, that kind of stuff.
link |
And that's never been the kind of problem
link |
in which I've had any particular insight.
link |
Though I must say, if you look at,
link |
cosmology is one of those.
link |
If you look at the actual things that Jim Peebles,
link |
one of this year's Nobel Prize in physics,
link |
ones from the local physics department,
link |
the kinds of things he's done,
link |
he's never crunched large data.
link |
Never, never, never.
link |
He's used the encapsulation of the work of others
link |
But it ultimately boiled down to thinking
link |
through the problem.
link |
Like what are the principles under which
link |
a particular phenomenon operates?
link |
And look, physics is always going to look
link |
for ways in which you can describe the system
link |
in a way which rises above the details.
link |
And to the hard dyed, the wool biologist,
link |
biology works because of the details.
link |
In physics, to the physicists,
link |
we want an explanation which is right
link |
in spite of the details.
link |
And there will be questions which we cannot answer
link |
as physicists because the answer cannot be found that way.
link |
There's, I'm not sure if you're familiar
link |
with the entire field of brain computer interfaces
link |
that's become more and more intensely researched
link |
and developed recently, especially with companies
link |
like Neuralink with Elon Musk.
link |
Yeah, I know there have always been the interests
link |
both in things like getting the eyes
link |
to be able to control things
link |
or getting the thought patterns
link |
to be able to move what had been a connected limb
link |
which is now connected through a computer.
link |
So in the case of Neuralink,
link |
they're doing 1,000 plus connections
link |
where they're able to do two way,
link |
activate and read spikes, neural spikes.
link |
Do you have hope for that kind of computer brain interaction
link |
in the near or maybe even far future
link |
of being able to expand the ability
link |
of the mind of cognition or understand the mind?
link |
It's interesting watching things go.
link |
When I first became interested in neurobiology,
link |
most of the practitioners thought you would be able
link |
to understand neurobiology by techniques
link |
which allowed you to record only one cell at a time.
link |
People like David Hubel,
link |
very strongly reflected that point of view.
link |
And that's been taken over by a generation,
link |
a couple of generations later,
link |
by a set of people who says not until we can record
link |
from 10 to the four, 10 to the five at a time,
link |
will we actually be able to understand
link |
how the brain actually works.
link |
And in a general sense, I think that's right.
link |
You have to begin to be able to look
link |
for the collective modes, the collective operations of things.
link |
It doesn't rely on this action potential or that cell.
link |
It relies on the collective properties of this set of cells
link |
connected with this kind of patterns and so on.
link |
And you're not going to succeed in seeing
link |
what those collective activities are
link |
without recording many cells at once.
link |
The question is how many at once?
link |
What's the threshold?
link |
Yeah, and look, it's being pursued hard
link |
in the motor cortex.
link |
The motor cortex does something which is complex,
link |
and yet the problem you're trying to address
link |
Now, neurobiology does it in ways that differ
link |
from the way an engineer would do it.
link |
An engineer would put in six highly accurate stepping motors
link |
are controlling a limb rather than 100,000 muscle fibers,
link |
each of which has to be individually controlled.
link |
And so understanding how to do things in a way
link |
which is much more forgiving and much more neural,
link |
I think would benefit the engineering world.
link |
The engineering world, a touch.
link |
Let's put in a pressure sensor or two,
link |
rather than an array of a gazillion pressure sensors,
link |
none of which are accurate,
link |
all of which are perpetually recalibrating themselves.
link |
So you're saying your hope is,
link |
your advice for the engineers of the future
link |
is to embrace the large chaos of a messy, air prone system
link |
like those of the biological systems.
link |
Like that's probably the way to solve some of these.
link |
I think you'll be able to make better computations
link |
slash robotics that way than by trying to force things
link |
into a robotics where joint motors are powerful
link |
and stepping motors are accurate.
link |
But then the physicists, the physicist in you
link |
will be lost forever in such systems
link |
because there's no simple fundamentals to explore
link |
in systems that are so large and messy.
link |
Well, you say that, and yet there's a lot of physics
link |
in the Navier Stokes equations,
link |
the equations of nonlinear hydrodynamics,
link |
huge amount of physics in them.
link |
All the physics of atoms and molecules has been lost,
link |
but it's been replaced by this other set of equations,
link |
which is just as true as the equations at the bottom.
link |
Now those equations are going to be harder to find
link |
in general biology, but the physicist in me says
link |
there are probably some equations of that sort.
link |
They're out there.
link |
They're out there, and if physics
link |
is going to contribute anything,
link |
it may contribute to trying to find out
link |
what those equations are and how to capture them
link |
Would you say that's one of the main open problems
link |
of our age is to discover those equations?
link |
Yeah, if you look at, there's molecules
link |
and there's psychological behavior,
link |
and these two are somehow related.
link |
They're layers of detail, they're layers of collectiveness,
link |
and to capture that in some vague way,
link |
several stages on the way up to see how these things
link |
can actually be linked together.
link |
So it seems in our universe, there's a lot of elegant
link |
equations that can describe the fundamental way
link |
that things behave, which is a surprise.
link |
I mean, it's compressible into equations.
link |
It's simple and beautiful, but it's still an open question
link |
whether that link is equally between molecules
link |
and the brain is equally compressible
link |
into elegant equations.
link |
But your sense, well, you're both a physicist
link |
and a dreamer, you have a sense that...
link |
Yeah, but I can only dream physics dreams.
link |
There was an interesting book called Einstein's Dreams,
link |
which alternates between chapters on his life
link |
and descriptions of the way time might have been but isn't.
link |
The linking between these being important ideas
link |
that Einstein might have had to think about
link |
the essence of time as he was thinking about time.
link |
So speaking of the essence of time in your biology,
link |
you're one human, famous, impactful human,
link |
but just one human with a brain living the human condition.
link |
But you're ultimately mortal, just like all of us.
link |
Has studying the mind as a mechanism
link |
changed the way you think about your own mortality?
link |
It has, really, because particularly as you get older
link |
and the body comes apart in various ways,
link |
I became much more aware of the fact
link |
that what is somebody is contained in the brain
link |
and not in the body that you worry about burying.
link |
And it is to a certain extent true
link |
that for people who write things down,
link |
equations, dreams, notepads, diaries,
link |
fractions of their thought does continue to live
link |
after they're dead and gone,
link |
after their body is dead and gone.
link |
And there's a sea change in that going on in my lifetime
link |
between when my father died, except for the things
link |
which were actually written by him, as it were.
link |
Very few facts about him will have ever been recorded.
link |
And the number of facts which are recorded
link |
about each and every one of us, forever now,
link |
as far as I can see, in the digital world.
link |
And so the whole question of what is death
link |
may be different for people a generation ago
link |
and a generation further ahead.
link |
Maybe we have become immortal under some definitions.
link |
Last easy question, what is the meaning of life?
link |
Looking back, you've studied the mind,
link |
us weird descendants of apes.
link |
What's the meaning of our existence on this little earth?
link |
What's the meaning of our existence on this little earth?
link |
Oh, that word meaning is as slippery as the word understand.
link |
Interconnected somehow, perhaps.
link |
Is there, it's slippery, but is there something
link |
that you, despite being slippery,
link |
can hold long enough to express?
link |
I've been amazed at how hard it is
link |
to define the things in a living system
link |
in the sense that one hydrogen atom
link |
is pretty much like another,
link |
but one bacterium is not so much like another bacterium,
link |
even of the same nominal species.
link |
In fact, the whole notion of what is the species
link |
gets a little bit fuzzy.
link |
And do species exist in the absence
link |
of certain classes of environments?
link |
And pretty soon one winds up with a biology
link |
which the whole thing is living,
link |
but whether there's actually any element of it
link |
which by itself would be said to be living
link |
becomes a little bit vague in my mind.
link |
So in a sense, the idea of meaning
link |
is something that's possessed by an individual,
link |
like a conscious creature.
link |
And you're saying that it's all interconnected
link |
in some kind of way that there might not even
link |
We're all kind of this complicated mess
link |
of biological systems at all different levels
link |
where the human starts and when the human ends is unclear.
link |
Yeah, yeah, and we're in neurobiology where the,
link |
oh, you say the neocortex is the thinking,
link |
but there's lots of things that are done on the spinal cord.
link |
And so where's the essence of thought?
link |
Is it just gonna be neocortex?
link |
Can't be, can't be.
link |
Yeah, maybe to understand and to build thought
link |
you have to build the universe along with the neocortex.
link |
It's all interlinked through the spinal cord.
link |
John, it's a huge honor talking today.
link |
Thank you so much for your time.
link |
I really appreciate it.
link |
Well, thank you for the challenge of talking with you.
link |
And it'll be interesting to see whether you can win
link |
five minutes out of this with just coherence
link |
Thanks for listening to this conversation
link |
with John Hopfield and thank you
link |
to our presenting sponsor, Cash App.
link |
Download it, use code LexPodcast.
link |
You'll get $10 and $10 will go to FIRST,
link |
an organization that inspires and educates young minds
link |
to become science and technology innovators of tomorrow.
link |
If you enjoy this podcast, subscribe on YouTube,
link |
get five stars on Apple Podcast, support on Patreon,
link |
or simply connect with me on Twitter at Lex Friedman.
link |
And now let me leave you with some words of wisdom
link |
from John Hopfield in his article titled, Now What?
link |
Choosing problems is the primary determinant
link |
of what one accomplishes in science.
link |
I have generally had a relatively short attention span
link |
in science problems.
link |
Thus, I have always been on the lookout
link |
for more interesting questions,
link |
either as my present ones get worked out
link |
or as they get classified by me as intractable,
link |
given my particular talents.
link |
He then goes on to say,
link |
what I have done in science relies entirely
link |
on experimental and theoretical studies by experts.
link |
I have a great respect for them,
link |
especially for those who are willing to attempt
link |
communication with someone who is not an expert in the field.
link |
I would only add that experts are good
link |
at answering questions.
link |
If you're brash enough, ask your own.
link |
Don't worry too much about how you found them.
link |
Thank you for listening and hope to see you next time.