John Hopfield is professor at Princeton, whose life's work weaved beautifully through biology, chemistry, neuroscience, and physics. Most crucially, he saw the messy world of biology through the piercing eyes of a physicist. He is perhaps best known for his work on associate neural networks, now known as Hopfield networks that were one of the early ideas that catalyzed the development of the modern field of deep learning.

EPISODE LINKS:
Now What? article: http://bit.ly/3843LeU
John wikipedia: https://en.wikipedia.org/wiki/John_Hopfield
Books mentioned:
- Einstein's Dreams: https://amzn.to/2PBa96X
- Mind is Flat: https://amzn.to/2I3YB84

This episode is presented by Cash App. Download it & use code "LexPodcast":
Cash App (App Store): https://apple.co/2sPrUHe
Cash App (Google Play): https://bit.ly/2MlvP5w

PODCAST INFO:
Podcast website:
https://lexfridman.com/podcast
Apple Podcasts:
https://apple.co/2lwqZIr
Spotify:
https://spoti.fi/2nEwCF8
RSS:
https://lexfridman.com/feed/podcast/
Full episodes playlist:
https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4
Clips playlist:
https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41

OUTLINE:
0:00 - Introduction
2:35 - Difference between biological and artificial neural networks
8:49 - Adaptation
13:45 - Physics view of the mind
23:03 - Hopfield networks and associative memory
35:22 - Boltzmann machines
37:29 - Learning
39:53 - Consciousness
48:45 - Attractor networks and dynamical systems
53:14 - How do we build intelligent systems?
57:11 - Deep thinking as the way to arrive at breakthroughs
59:12 - Brain-computer interfaces
1:06:10 - Mortality
1:08:12 - Meaning of life

CONNECT:
- Subscribe to this YouTube channel
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/LexFridmanPage
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

The following is a conversation with John Hopfield, professor at Princeton, whose life's work weaved beautifully through biology, chemistry, neuroscience, and physics.

Most crucially, he saw the messy world of biology through the piercing eyes of a physicist.

He's perhaps best known for his work on associative neural networks, now known as Hopfield networks, that were 1 of the early ideas that catalyzed the development of the modern field of deep learning.

As his 2019 Franklin Medal in Physics Award states, he applied concepts of theoretical physics to provide new insights on important biological questions in a variety of areas, including genetics and neuroscience with significant impact on machine learning.

And as John says in his 2018 article titled, Now What?

His accomplishments have often come about by asking that very question, now what?

And often responding by a major change of direction.

This is the Artificial Intelligence Podcast.

If you enjoy it, subscribe on YouTube, give it 5 stars on Apple Podcast, support it on Patreon, or simply connect with me on Twitter, at Lex Friedman, spelled F-R-I-D-M-A-N.

As usual, I'll do 1 or 2 minutes of ads now and never any ads in the middle that can break the flow of the conversation.

I hope that works for you and doesn't hurt the listening experience.

This show is presented by Cash App, the number 1 finance app in the App Store.

Cash App lets you send money to friends, buy Bitcoin, and invest in the stock market with as little as $1.

Since Cash App does fractional share trading, Let me mention that the order execution algorithm that works behind the scenes to create the abstraction of fractional orders is to me an algorithmic marvel.

So big props to the Cash App engineers for solving a hard problem that in the end provides an easy interface that takes a step up the next layer of abstraction over the stock market, making trading more accessible for new investors and diversification much easier.

So again, if you get Cash App from the App Store or Google Play and use code LEXPODCAST, you'll get $10, and Cash App will also donate $10 to FIRST, 1 of my favorite organizations that is helping advance robotics and STEM education for young people around the world.

And Now, here's my conversation with John Hopfield.

What difference between biological neural networks and artificial neural networks is most captivating and profound to you?

At the higher philosophical level, Let's not get technical just yet.

1 of the things that very much intrigues me is the fact that neurons have all kinds of components, properties to them.

And evolutionary biology, if you have some little quirk in how a molecule works or how a cell works, and it can be made use of, evolution will sharpen it up and make it into a useful feature rather than a glitch.

And so you expect in neurobiology for evolution to have captured all kinds of possibilities of getting neurons, of how you get neurons to do things for you.

And that aspect has been completely suppressed in artificial neural networks.

Do the glitches become features in the biological neural network?

Look, let me take 1 of the things that I used to do research on.

If you take things which oscillate, have rhythms which are sort of close to each other.

Under some circumstances, these things will have a phase transition, and suddenly the rhythm will...

There was a marvelous physical example of that in the Millennium Bridge across the Thames River about...

And pedestrians walking across, pedestrians don't walk synchronized, they don't walk in lockstep.

But they're all walking about the same frequency.

And the bridge could sway at that frequency, and the slight sway made pedestrians tend a little bit to lock into step.

And after a while, the bridge was oscillating back and forth, and the pedestrians were walking in step to it.

And you could see it in the movies made out of the bridge.

And the engineers made a simple-minded mistake.

They assumed when you walk, it's step, step, step, and it's back and forth motion.

But when you walk, it's also right foot, left foot, side to side motion.

And it's the side to side motion for which the bridge was strong enough, but it wasn't stiff enough.

And as a result, you would feel the motion, and you'd fall into step with it.

And people were very uncomfortable with it.

They closed the bridge for 2 years while they built stiffening for it.

Now, nerves, look, nerve cells produce action potentials.

You have a bunch of cells which are loosely coupled together producing action potentials of the same rate.

There'll be some circumstances under which these things can lock together.

Well, if they fire together, you can be sure that other cells are gonna notice it.

So you can make a computational feature out of this in an evolving brain.

Most artificial neural networks don't even have action potentials, let alone have the possibility for synchronizing them.

And you mentioned the evolutionary process.

So the evolutionary process that builds on top of biological systems leverages that, the weird mess of it somehow.

So how do you make sense of that ability to leverage all the different kinds of complexities in the biological brain?

Well, look, at the biological molecule level, you have a piece of DNA which encode for a particular protein.

You could duplicate that piece of DNA and now 1 part of it can code for that protein but the other 1 could itself change a little bit and thus start coding for a molecule which is slightly different.

Now that molecule which is slightly different had a function which helped any old chemical reaction which was important to the cell.

It would go ahead and let that try, and evolution would slowly improve that function.

And so you have the possibility of duplicating, and then having things drift apart.

1 of them retain the old function, the other 1 do something new for you.

And there's evolutionary pressure to improve.

Look, there is in computers too, but it's improvement has to do with closing some companies and opening some others.

The evolutionary process looks a little different.

Companies close, yeah, go bankrupt and are born, yeah, shorter, but not much shorter.

I mean, if you think of companies as a single organism that builds and you all know, yeah, it's a fascinating dual correspondence there between biological and...

And companies have difficulty having a new product competing with an old product.

And when IBM built its first PC, you probably read the book, They made a little isolated internal unit to make the PC.

And for the first time in IBM's history, they didn't insist that you build it out of IBM components.

But they understood that they could get into this market, which is a very different thing by completely changing their culture.

And biology finds other markets in a more adaptive way.

So maybe you've already said it, but what to use the most beautiful aspect or mechanism of the human mind?

Is it the adaptive, the ability to adapt as you've described, or is there some other little quirk that you particularly like?

Adaptation is everything when you get down to it.

But the difference, there are differences between adaptation, where your learning goes on only over generations, over evolutionary time, where your learning goes on at the timescale of 1 individual who must learn from the environment during that individual's lifetime.

And biology has both kinds of learning in it.

And the thing which makes neurobiology hard is that it's a mathematical system as it were built on this other kind of evolutionary system.

Well when you talk to a computer scientist about neural networks, it's all math.

The fact that biology actually came about from evolution and the fact that Biology is about a system which you can build in 3 dimensions.

If you look at computer chips, computer chips are basically two-dimensional structures, maybe 2.1 dimensions, but they really have difficulty doing three-dimensional wiring.

Biology is, neocortex is actually also sheet-like, and it sits on top of the white matter, which is about 10 times the volume of the gray matter and contains all of what you might call the wires.

But there's a huge, the effect of computer structure on what is easy and what is hard is immense.

And biology does, it makes some things easy that are very difficult to understand how to do computationally.

On the other hand, you can't do simple floating point arithmetic, so it's awfully stupid.

Yeah, and you're saying this kind of 3 dimensional complicated structure makes, it's still math.

The kind of math it's doing enables you to solve problems of a very different kind.

The evolutionary adaptation and the adaptation or learning at the scale of a single human life, which do you, which is particularly beautiful to you and interesting from a research and from just a human perspective?

I find things most interesting that I begin to see how to get into the edges of them and tease them apart a little bit to see how they work.

And since I can't see the evolutionary process going on, I'm in awe of it, but I find it just a black hole as far as trying to understand what to do.

And so in a certain sense, I'm in awe of it, but I couldn't be interested in working on it.

The human life's time scale is however thing you can tease apart and study.

Yeah, you can do, there's developmental neurobiology which understands how the connections and how the structure evolves from a combination of what the genetics is like and the real, the fact that you're building a system in 3 dimensions.

In just days and months, those early, early days of a human life are really interesting.

They are and of course, there are times of immense cell multiplication.

There are also times of the greatest cell death in the brain is during infancy.

So what is not effective, what is not wired well enough to use at the moment, throw it out.

From, let me ask, from what field do you think the biggest breakthroughs in understanding the mind will come in the next decades?

Is it neuroscience, computer science, neurobiology, psychology, physics, maybe math, maybe literature?

Well, of course, I see the world always through a lens of physics.

And the way I pick problems is very characteristic of physics and of an intellectual background which is not psychology, which is not chemistry and so on and so on.

Both of my parents were physicists and the real thing I got out of that was a feeling that the world is an understandable place.

And if you do enough experiments and think about what they mean and structure things that you can do the mathematics of the relevant to the experiments.

You also be able to understand how things work.

Did you change your mind at all through many decades of trying to understand the mind, of studying it different kinds of ways, not even the mind, just biological systems.

You still have hope that physics, that you can understand.

There's a question of what do you mean by understand?

When I taught freshman physics, I used to say I wanted to give physics to understand the subject, to understand Newton's laws.

I didn't want them simply to memorize a set of examples to which they knew the equations to write down to generate the answers, I had this nebulous idea of understanding.

So that if you looked at a situation, you could say, Oh, I expect the ball to make that trajectory.

I expect some intuitive notion of understanding.

And I don't know how to express that very well.

And you run smack up against it when you look at these simple neural nets, feed forward neural nets, which do amazing things and yet you know contain nothing of the essence of what I would have felt was understanding.

Understanding is more than just an enormous lookup table.

So, I mean, ask another way, these feed forward neural networks, do you think they'll ever understand?

I think if you look at real systems, feedback is an essential aspect of how these real systems compute.

On the other hand, if I have a mathematical system with feedback, I know I can unlayer this and do it.

But I have an exponential expansion in the amount of stuff I have to build if I'm gonna solve the problem that way.

So we can talk even about recurrent neural networks, so recurrence.

But do you think all the pieces are there to achieve understanding through these simple mechanisms?

Like, back to our original question, what is the fundamental, is there a fundamental difference between artificial neural networks and biological?

Suppose you ask a neurosurgeon, when is somebody dead?

They'll probably go back to saying, well, I can look at the brain rhythms and tell you this is a brain which is never going to function again.

This 1 is, this other 1 is 1 which if we treat it well is still recoverable.

And then just do that by some electrodes looking at simple electrical patterns which don't look in any detail at all at what individual neurons are doing.

These rhythms are utterly absent from anything which goes on at Google.

So, well that's like comparing, okay, I'll tell you.

It's like you're comparing the greatest classical musician in the world to a child first learning to play.

The question I'm at, but they're still both playing the piano.

I'm asking, will it ever go on at Google?

Because you're 1 of the seminal figures in both launching both disciplines, both sides of the river.

I think it's going to go on generation after generation, the way it has, where what you might call the AI computer science community says, let's take the following.

This is our model of neurobiology at the moment.

Let's pretend it's good enough and do everything we can with it." And it does interesting things, and after a while, it sort of grinds into the sand, and you say, ah, something else is needed for neurobiology, and some other grand thing comes in and enables you to go a lot further.

I think there could be generations of this evolution.

I don't know how many of them, and each 1 is going to get you further into what our brain does, and in some sense, pass the Turing test longer and for more broad aspects.

And how many of these are there are going to have to be before you say, I've made something, I've made a human, I don't know.

And going back to my brain waves as it were.

From the AI point of view, they would say, ah, maybe these are an heavy phenomenon and not important at all.

Dodge, go above 45 miles an hour and the wheels would shimmy.

But in biology, if it were useful to know, when are you going more than 45 miles an hour, you just capture that, and you wouldn't worry about where it came from.

It's going to be a long time before that kind of thing, which can take place in large, complex networks of things, is actually used in the computation.

Look, the, How many transistors are there in your laptop these days?

And all the transistors are somewhat similar.

And most physical systems, with that many parts, all of which are similar, have collective properties.

Sound waves in air, earthquakes, what have you, have collective properties.

There are no collective properties used in artificial neural networks in AI.

If biology uses them, it's going to take us to more generations of things for people to actually dig in and see how they are used and what they mean.

We might have to return several times to neurobiology and try to make our transistors more messy.

At the same time, the simple ones will conquer big aspects.

And I think 1 of the most biggest surprises to me was how well learning systems, which are manifestly non-biological, how important they can be actually, and how important and how useful they can be in AI.

So if we can just take a stroll to some of your work that is incredibly surprising that it works as well as it does that launched a lot of the recent work with neural networks.

If we go to what are now called Hopfield networks, can you tell me what is associative memory in the mind for the human side?

Okay, what you mean by associative memory is, ah, you have a memory of each of your friends.

Your friend has all kinds of properties, from what they look like, to what their voice sounds like, to where they went to college, where you met them, go on and on, what science papers they've written.

If I start talking about a 5 foot 10, wire-aided cognitive scientist who's got a very bad back, it doesn't take very long for you to say, oh, he's talking about Jeff Hinton.

I never mentioned the name or anything very particular.

But somehow a few facts that are associated with this, with a particular person enables you to get a hold of the rest of the facts.

Or not the rest of them, another subset of them.

And it's this ability to link things together, link experiences together, which it goes under the general name of associative memory.

And a large part of intelligent behavior is actually just large associative memories at work, as far as I can see.

What do you think is the mechanism of how it works in the mind?

Do you have inklings of how this essential thing for cognition works?

years ago was of course a crude physics model to show the kind, actually enable you to understand, my old sense of understanding as a physicist, because you could say, ah, I understand why this goes to stable states, it's like things going downhill.

And that gives you something with which to think in physical terms rather than only in mathematical terms.

So you've created these associative artificial networks.

And now if you look at what I did, I didn't at all describe a system which gracefully learns.

I described a system in which you could understand how learning could link things together, how very crudely it might learn.

1 of the things which intrigues me as I reinvestigate that system now to some extent is, look, I'll see you every second for the next hour or what have you.

Each look at you is a little bit different.

I don't store all those second-by-second images, I don't store 3,000 images, I somehow compact this information.

So I now have a view of you, which I can use.

It doesn't slavishly remember anything in particular, but it compacts the information into useful chunks, which are somehow, it's these chunks, which are not just activities of neurons, bigger things than that, which are the real entities which are useful to you.

Useful to you to describe, to compress this information.

And you have to compress it in such a way that if I get, if the information comes in just like this again, I don't bother to rewrite it.

Or efforts to rewrite it simply do not yield anything because those things are already written.

And that needs to be not, look this up, have I written this, have I started somewhere already.

It's gotta be something which is much more automatic in the machine hardware.

Right, so in the human mind, how complicated is that process, do you think?

So you've created, it feels weird to be sitting with John Hopfield calling him Hopfield Networks, but.

Yeah, but nevertheless, that's what everyone calls him, so here we are.

You and Richard Feynman sat down and talked about associative memory.

Now, if you look at the mind, well, you can't quite simplify it so perfectly.

Well, let me backtrack just a little bit.

You can ask, if you want to model biology, if you want to model neurobiology, what is the time scale?

There's a dynamical system in which, of a fairly fast time scale, in which you could say, the synapses don't change much during this computation, so I'll think of the synapses fixed and just do the dynamics of the activity.

Or you can say, the synapses are changing fast enough that I have to have the synaptic dynamics working at the same time as the system dynamics in order to understand the biology.

Most, if you look at the feedforward artificial neural nets, they're all done as learning is, First of all, I spend some time learning, not performing, and I turn off learning and I perform.

And so, as I look more deeply at neurobiology, even as an associate of memory, I've got to face the fact that the dynamics of a synapse change is going on all the time.

And I can't just get by by saying, I'll do the dynamics of activity with fixed synapses.

So the synaptic, the dynamics of the synapses is actually fundamental to the whole system.

And there's nothing necessarily separating the time scales.

When the time scales can be separated, it's neat from the physicist's or the mathematician's point of view, but it's not necessarily true in neurobiology.

So You're kind of dancing beautifully between showing a lot of respect to physics and then also saying that physics cannot quite reach the complexity of biology.

Or do you continuously dance between the 2?

See all Lex Fridman transcripts on Youtube

John Hopfield: Physics View of the Mind and Neurobiology | Lex Fridman Podcast #76