back to indexJay McClelland: Neural Networks and the Emergence of Cognition | Lex Fridman Podcast #222
link |
The following is a conversation with Jay McClelland,
link |
a cognitive scientist at Stanford
link |
and one of the seminal figures
link |
in the history of artificial intelligence
link |
and specifically neural networks.
link |
Having written the parallel distributed processing book
link |
with David Rommelhart,
link |
who coauthored the backpropagation paper with Jeff Hinton.
link |
In their collaborations, they've paved the way
link |
for many of the ideas
link |
at the center of the neural network based
link |
machine learning revolution of the past 15 years.
link |
To support this podcast,
link |
please check out our sponsors in the description.
link |
This is the Lex Friedman podcast
link |
and here is my conversation with Jay McClelland.
link |
You are one of the seminal figures
link |
in the history of neural networks.
link |
At the intersection of cognitive psychology
link |
and computer science,
link |
what to you has over the decades emerged
link |
as the most beautiful aspect about neural networks?
link |
Both artificial and biological.
link |
The fundamental thing I think about with neural networks
link |
is how they allow us to link
link |
biology with the mysteries of thought.
link |
When I was first entering the field myself
link |
in the late 60s, early 70s,
link |
cognitive psychology had just become a field.
link |
There was a book published in 67 called Cognitive Psychology.
link |
And the author said that the study of the nervous system
link |
was only of peripheral interest.
link |
It wasn't going to tell us anything about the mind.
link |
And I didn't agree with that.
link |
I always felt, oh, look, I'm a physical being.
link |
From dust to dust, you know,
link |
ashes to ashes, and somehow I emerged from that.
link |
So that's really interesting.
link |
So there was a sense with cognitive psychology
link |
that in understanding the neuronal structure of things,
link |
you're not going to be able to understand the mind.
link |
And then your sense is if we study these neural networks,
link |
we might be able to get at least very close
link |
to understanding the fundamentals of the human mind.
link |
I used to think, or I used to talk about the idea
link |
of awakening from the Cartesian dream.
link |
So Descartes, you know, thought about these things, right?
link |
He was walking in the gardens of Versailles one day,
link |
and he stepped on a stone.
link |
And a statue moved.
link |
And he walked a little further,
link |
he stepped on another stone, and another statue moved.
link |
And he, like, why did the statue move
link |
when I stepped on the stone?
link |
And he went and talked to the gardeners,
link |
and he found out that they had a hydraulic system
link |
that allowed the physical contact with the stone
link |
to cause water to flow in various directions,
link |
which caused water to flow into the statue
link |
and move the statue.
link |
And he used this as the beginnings of a theory
link |
about how animals act.
link |
And he had this notion that these little fibers
link |
that people had identified that weren't carrying the blood,
link |
you know, were these little hydraulic tubes
link |
that if you touch something, there would be pressure,
link |
and it would send a signal of pressure
link |
to the other parts of the system,
link |
and that would cause action.
link |
So he had a mechanistic theory of animal behavior.
link |
And he thought that the human had this animal body,
link |
but that some divine something else
link |
had to have come down and been placed in him
link |
to give him the ability to think, right?
link |
So the physical world includes the body in action,
link |
but it doesn't include thought according to Descartes, right?
link |
And so the study of physiology at that time
link |
was the study of sensory systems and motor systems
link |
and things that you could directly measure
link |
when you stimulated neurons and stuff like that.
link |
And the study of cognition was something that, you know,
link |
was tied in with abstract computer algorithms
link |
and things like that.
link |
But when I was an undergraduate,
link |
I learned about the physiological mechanisms.
link |
And so when I'm studying cognitive psychology
link |
as a first year PhD student, I'm saying,
link |
wait a minute, the whole thing is biological, right?
link |
You had that intuition right away.
link |
That always seemed obvious to you.
link |
Isn't that magical, though,
link |
that from just a little bit of biology can emerge
link |
the full beauty of the human experience?
link |
Why is that so obvious to you?
link |
Well, obvious and not obvious at the same time.
link |
And I think about Darwin in this context, too,
link |
because Darwin knew very early on
link |
that none of the ideas that anybody had ever offered
link |
gave him a sense of understanding
link |
how evolution could have worked.
link |
But he wanted to figure out how it could have worked.
link |
That was his goal.
link |
And he spent a lot of time working on this idea
link |
and reading about things that gave him hints
link |
and thinking they were interesting but not knowing why
link |
and drawing more and more pictures of different birds
link |
that differ slightly from each other and so on, you know.
link |
And then he figured it out.
link |
But after he figured it out, he had nightmares about it.
link |
He would dream about the complexity of the eye
link |
and the arguments that people had given
link |
about how ridiculous it was to imagine
link |
that that could have ever emerged
link |
from some sort of, you know, unguided process, right?
link |
That it hadn't been the product of design.
link |
And so he didn't publish for a long time,
link |
in part because he was scared of his own ideas.
link |
He didn't think they could possibly be true.
link |
But then, you know, by the time
link |
the 20th century rolls around, we all,
link |
you know, we understand that,
link |
many people understand or believe
link |
that evolution produced, you know, the entire
link |
range of animals that there are.
link |
And, you know, Descartes's idea starts to seem
link |
a little wonky after a while, right?
link |
Like, well, wait a minute.
link |
There's the apes and the chimpanzees and the bonobos
link |
and, you know, like, they're pretty smart in some ways.
link |
You know, so what?
link |
Oh, you know, somebody comes up,
link |
oh, there's a certain part of the brain
link |
that's still different.
link |
They don't, you know, there's no hippocampus
link |
in the monkey brain.
link |
It's only in the human brain.
link |
Huxley had to do a surgery in front of many, many people
link |
in the late 19th century to show to them
link |
there's actually a hippocampus in the chimpanzee's brain.
link |
You know, so the continuity of the species
link |
is another element that, you know,
link |
contributes to this sort of, you know, idea
link |
that we are ourselves a total product of nature.
link |
And that, to me, is the magic and the mystery,
link |
how nature could actually, you know,
link |
give rise to organisms that have the capabilities
link |
So it's interesting because even the idea of evolution
link |
is hard for me to keep all together in my mind.
link |
So because we think of a human time scale,
link |
it's hard to imagine, like, the development
link |
of the human eye would give me nightmares too.
link |
Because you have to think across many, many, many
link |
generations, and it's very tempting to think about
link |
kind of a growth of a complicated object
link |
and it's like, how is it possible for such a thing
link |
Because also, me, from a robotics engineering perspective,
link |
it's very hard to build these systems.
link |
How can, through an undirected process,
link |
can a complex thing be designed?
link |
It seems not, it seems wrong.
link |
Yeah, so that's absolutely right.
link |
And I, you know, a slightly different career path
link |
that would have been equally interesting to me
link |
would have been to actually study the process
link |
of embryological development flowing on
link |
into brain development and the exquisite sort of laying
link |
down of pathways and so on that occurs in the brain.
link |
And I know the slightest bit about that is not my field,
link |
but there are, you know, fascinating aspects
link |
to this process that eventually result in the, you know,
link |
the complexity of various brains.
link |
At least, you know, one thing we're,
link |
in the field, I think people have felt for a long time,
link |
in the study of vision, the continuity between humans
link |
and nonhuman animals has been second nature
link |
I was having, I had this conversation with somebody
link |
who is a vision scientist and he was saying,
link |
oh, we don't have any problem with this.
link |
You know, the monkey's visual system
link |
and the human visual system, extremely similar
link |
up to certain levels, of course, they diverge after a while.
link |
But the first, the visual pathway from the eye
link |
to the brain and the first few layers of cortex
link |
or cortical areas, I guess one would say,
link |
are extremely similar.
link |
Yeah, so on the cognition side is where the leap
link |
seems to happen with humans,
link |
that it does seem we're kind of special.
link |
And that's a really interesting question
link |
when thinking about alien life
link |
or if there's other intelligent alien civilizations
link |
out there, is how special is this leap?
link |
So one special thing seems to be the origin of life itself.
link |
However you define that, there's a gray area.
link |
And the other leap, this is very biased perspective
link |
of a human, is the origin of intelligence.
link |
And again, from an engineer perspective,
link |
it's a difficult question to ask.
link |
An important one is how difficult is that leap?
link |
How special were humans?
link |
Did a monolith come down?
link |
Did aliens bring down a monolith
link |
and some apes had to touch a monolith to get it?
link |
That's a lot like Descartes idea, right?
link |
Exactly, but it just seems one heck of a leap
link |
to get to this level of intelligence.
link |
Yeah, and so Chomsky argued that some genetic fluke occurred
link |
100,000 years ago and just happened
link |
that some human, some hominin predecessor of current humans
link |
had this one genetic tweak that resulted in language.
link |
And language then provided this special thing that separates us
link |
from all other animals.
link |
I think there's a lot of truth to the value and importance
link |
of language, but I think it comes along
link |
with the evolution of a lot of other related things related
link |
to sociality and mutual engagement with others
link |
and establishment of, I don't know,
link |
rich mechanisms for organizing and understanding
link |
of the world, which language then plugs into.
link |
Right, so language is a tool that
link |
allows you to do this kind of collective intelligence.
link |
And whatever is at the core of the thing that
link |
allows for this collective intelligence is the main thing.
link |
And it's interesting to think about that one fluke, one
link |
mutation could lead to the first crack opening of the door
link |
to human intelligence.
link |
All it takes is one.
link |
Evolution just kind of opens the door a little bit,
link |
and then time and selection takes care of the rest.
link |
You know, there's so many fascinating aspects
link |
to these kinds of things.
link |
So we think of evolution as continuous, right?
link |
We think, oh, yes, OK, over 500 million years,
link |
there could have been this relatively continuous changes.
link |
And but that's not what anthropologists,
link |
evolutionary biologists found from the fossil record.
link |
They found hundreds of millions of years of stasis.
link |
And then suddenly a change occurs.
link |
Well, suddenly on that scale is a million years or something,
link |
or even 10 million years.
link |
But the concept of punctuated equilibrium
link |
was a very important concept in evolutionary biology.
link |
And that also feels somehow right about the stages
link |
of our mental abilities.
link |
We seem to have a certain kind of mindset at a certain age.
link |
And then at another age, we look at that four year old
link |
and say, oh, my god, how could they have thought that way?
link |
So Piaget was known for this kind of stage theory
link |
of child development, right?
link |
And you look at it closely, and suddenly those stages
link |
are so discreet and it transitions.
link |
But the difference between the four year old and the seven
link |
year old is profound.
link |
And that's another thing that's always interested me
link |
is how something happens over the course of several years
link |
of experience where at some point
link |
we reach the point where something
link |
like an insight or a transition or a new stage of development
link |
And these kinds of things can be understood
link |
in complex systems research.
link |
And so evolutionary biology, developmental biology,
link |
cognitive development are all things
link |
that have been approached in this kind of way.
link |
Just like you said, I find both fascinating
link |
those early years of human life, but also
link |
the early minutes, days from the embryonic development
link |
to how from embryos you get the brain.
link |
That development, again, from an engineer perspective,
link |
So the early, when you deploy the brain to the human world
link |
and it gets to explore that world and learn,
link |
that's fascinating.
link |
But just like the assembly of the mechanism
link |
that is capable of learning, that's amazing.
link |
The stuff they're doing with brain organoids
link |
where you can build many brains and study
link |
that self assembly of a mechanism from the DNA material,
link |
that's like, what the heck?
link |
You have literally biological programs
link |
that just generate a system, this mushy thing that's
link |
able to be robust and learn in a very unpredictable world
link |
and learn seemingly arbitrary things,
link |
or a very large number of things that enable survival.
link |
Ultimately, that is a very important part
link |
of the whole process of understanding
link |
this emergence of mind from brain kind of thing.
link |
And the whole thing seems to be pretty continuous.
link |
So let me step back to neural networks
link |
for another brief minute.
link |
You wrote parallel distributed processing books
link |
that explored ideas of neural networks in the 1980s
link |
together with a few folks.
link |
But the books you wrote with David Romelhart,
link |
who is the first author on the back propagation
link |
paper with Jeff Hinton.
link |
So these are just some figures at the time
link |
that we're thinking about these big ideas.
link |
What are some memorable moments of discovery
link |
and beautiful ideas from those early days?
link |
I'm going to start sort of with my own process in the mid 70s
link |
and then into the late 70s when I met Jeff Hinton
link |
and he came to San Diego and we were all together.
link |
In my time in graduate schools, I've already described to you,
link |
I had this sort of feeling of, OK, I'm
link |
really interested in human cognition,
link |
but this disembodied sort of way of thinking about it
link |
that I'm getting from the current mode of thought about it
link |
isn't working fully for me.
link |
And when I got my assistant professorship,
link |
I went to UCSD and that was in 1974.
link |
Something amazing had just happened.
link |
Dave Romelhart had written a book together
link |
with another man named Don Norman
link |
and the book was called Explorations in Cognition.
link |
And it was a series of chapters exploring
link |
interesting questions about cognition,
link |
but in a completely sort of abstract, nonbiological kind
link |
And I'm saying, gee, this is amazing.
link |
I'm coming to this community where people can get together
link |
and feel like they've collectively exploring ideas.
link |
And it was a book that had a lot of, I don't know,
link |
And Don Norman, who was the more senior figure
link |
to Romelhart at that time who led that project,
link |
always created this spirit of playful exploration of ideas.
link |
And so I'm like, wow, this is great.
link |
But I was also still trying to get from the neurons
link |
And I realized at one point, I got this opportunity
link |
to go to a conference where I heard a talk by a man named
link |
James Anderson, who was an engineer,
link |
but by then a professor in a psychology department, who
link |
had used linear algebra to create neural network
link |
models of perception and categorization and memory.
link |
And it just blew me out of the water
link |
that one could create a model that was simulating neurons,
link |
not just engaged in a stepwise algorithmic process that
link |
was construed abstractly.
link |
But it was simulating remembering and recalling
link |
and recognizing the prior occurrence of a stimulus
link |
or something like that.
link |
So for me, this was a bridge between the mind and the brain.
link |
And I remember I was walking across campus one day in 1977,
link |
and I almost felt like St. Paul on the road to Damascus.
link |
I said to myself, if I think about the mind in terms
link |
of a neural network, it will help
link |
me answer the questions about the mind
link |
that I'm trying to answer.
link |
And that really excited me.
link |
So I think that a lot of people were
link |
becoming excited about that.
link |
And one of those people was Jim Anderson, who I had mentioned.
link |
Another one was Steve Grossberg, who
link |
had been writing about neural networks since the 60s.
link |
And Jeff Hinton was yet another.
link |
And his PhD dissertation showed up in an applicant pool
link |
to a postdoctoral training program
link |
that Dave and Don, the two men I mentioned before,
link |
Rumelhart and Norman, were administering.
link |
And Rumelhart got really excited about Hinton's PhD dissertation.
link |
And so Hinton was one of the first people
link |
who came and joined this group of postdoctoral scholars
link |
that was funded by this wonderful grant that they got.
link |
Another one who is also well known
link |
in neural network circles is Paul Smolenski.
link |
He was another one of that group.
link |
Anyway, Jeff and Jim Anderson organized a conference
link |
at UCSD where we were.
link |
And it was called Parallel Models of Associative Memory.
link |
And it brought all the people together
link |
who had been thinking about these kinds of ideas
link |
And this began to kind of really resonate
link |
with some of Rumelhart's own thinking,
link |
some of his reasons for wanting something
link |
other than the kinds of computation
link |
he'd been doing so far.
link |
So let me talk about Rumelhart now for a minute,
link |
OK, with that context.
link |
Well, let me also just pause because he
link |
said so many interesting things before we go to Rumelhart.
link |
So first of all, for people who are not familiar,
link |
neural networks are at the core of the machine learning,
link |
deep learning revolution of today.
link |
Geoffrey Hinton that we mentioned
link |
is one of the figures that were important in the history
link |
like yourself in the development of these neural networks,
link |
artificial neural networks that are then
link |
used for the machine learning application.
link |
Like I mentioned, the backpropagation paper
link |
is one of the optimization mechanisms
link |
by which these networks can learn.
link |
And the word parallel is really interesting.
link |
So it's almost like synonymous from a computational
link |
perspective how you thought at the time about neural networks
link |
as parallel computation.
link |
Would that be fair to say?
link |
Well, yeah, the parallel, the word parallel in this
link |
comes from the idea that each neuron is
link |
an independent computational unit, right?
link |
It gathers data from other neurons,
link |
it integrates it in a certain way,
link |
and then it produces a result. And it's
link |
a very simple little computational unit.
link |
But it's autonomous in the sense that it does its thing, right?
link |
It's in a biological medium where
link |
it's getting nutrients and various chemicals
link |
But you can think of it as almost like a little computer
link |
So the idea is that each our brains have, oh, look,
link |
100 or hundreds, almost a billion
link |
of these little neurons, right?
link |
And they're all capable of doing their work at the same time.
link |
So it's like instead of just a single central processor that's
link |
engaged in chug one step after another,
link |
we have a billion of these little computational units
link |
working at the same time.
link |
So at the time that's, I don't know, maybe you can comment,
link |
it seems to me, even still to me,
link |
quite a revolutionary way to think about computation
link |
relative to the development of theoretical computer science
link |
alongside of that where it's very much like sequential computer.
link |
You're analyzing algorithms that are running on a single computer.
link |
You're saying, wait a minute, why don't we
link |
take a really dumb, very simple computer
link |
and just have a lot of them interconnected together?
link |
And they're all operating in their own little world
link |
and they're communicating with each other
link |
and thinking of computation that way.
link |
And from that kind of computation,
link |
trying to understand how things like certain characteristics
link |
of the human mind can emerge.
link |
That's quite a revolutionary way of thinking, I would say.
link |
Well, yes, I agree with you.
link |
And there's still this sort of sense
link |
of not sort of knowing how we kind of get all the way there,
link |
And this very much remains at the core of the questions
link |
that everybody's asking about the capabilities
link |
of deep learning and all these kinds of things.
link |
But if I could just play this out a little bit,
link |
a convolutional neural network or a CNN,
link |
which many people may have heard of, is a set of,
link |
you could think of it biologically as a set of
link |
collections of neurons.
link |
Each collection has maybe 10,000 neurons in it.
link |
But there's many layers.
link |
Some of these things are hundreds or even
link |
1,000 layers deep.
link |
But others are closer to the biological brain
link |
and maybe they're like 20 layers deep or something like that.
link |
So within each layer, we have thousands of neurons
link |
or tens of thousands maybe.
link |
Well, in the brain, we probably have millions in each layer.
link |
But we're getting sort of similar in a certain way.
link |
And then we think, OK, at the bottom level,
link |
there's an array of things that are like the photoreceptors.
link |
In the eye, they respond to the amount
link |
of light of a certain wavelength at a certain location
link |
on the pixel array.
link |
So that's like the biological eye.
link |
And then there's several further stages going up,
link |
layers of these neuron like units.
link |
And you go from that raw input array of pixels
link |
to the classification, you've actually
link |
built a system that could do the same kind of thing
link |
that you and I do when we open our eyes and we look around
link |
and we see there's a cup, there's a cell phone,
link |
there's a water bottle.
link |
And these systems are doing that now, right?
link |
So they are, in terms of the parallel idea
link |
that we were talking about before,
link |
they are doing this massively parallel computation
link |
in the sense that each of the neurons in each
link |
of those layers is thought of as computing
link |
its little bit of something about the input
link |
simultaneously with all the other ones in the same layer.
link |
We get to the point of abstracting that away
link |
and thinking, oh, it's just one whole vector that's
link |
being computed, one activation pattern that's
link |
computed in a single step.
link |
And that abstraction is useful, but it's still that parallel.
link |
And distributed processing, right?
link |
Each one of these guys is just contributing
link |
a tiny bit to that whole thing.
link |
And that's the excitement that you felt,
link |
that from these simple things, you can emerge.
link |
When you add these level of abstractions on it,
link |
you can start getting all the beautiful things
link |
that we think about as cognition.
link |
And so, OK, so you have this conference.
link |
I forgot the name already, but it's
link |
Parallel and Something Associative Memory and so on.
link |
Very exciting, technical and exciting title.
link |
And you started talking about Dave Romerhart.
link |
So who is this person that was so,
link |
you've spoken very highly of him.
link |
Can you tell me about him, his ideas, his mind, who he was
link |
as a human being, as a scientist?
link |
So Dave came from a little tiny town in Western South Dakota.
link |
And his mother was the librarian,
link |
and his father was the editor of the newspaper.
link |
And I know one of his brothers pretty well.
link |
They grew up, there were four brothers,
link |
and they grew up together.
link |
And their father encouraged them to compete with each other
link |
They competed in sports, and they competed in mind games.
link |
I don't know, things like Sudoku and chess and various things
link |
And Dave was a standout undergraduate.
link |
He went at a younger age than most people
link |
do to college at the University of South Dakota
link |
and majored in mathematics.
link |
And I don't know how he got interested in psychology,
link |
but he applied to the mathematical psychology
link |
program at Stanford and was accepted as a PhD student
link |
to study mathematical psychology at Stanford.
link |
So mathematical psychology is the use of mathematics
link |
to model mental processes.
link |
So something that I think these days
link |
might be called cognitive modeling, that whole space.
link |
Yeah, it's mathematical in the sense
link |
that you say, if this is true and that is true,
link |
then I can derive that this should follow.
link |
And so you say, these are my stipulations
link |
about the fundamental principles,
link |
and this is my prediction about behavior.
link |
And it's all done with equations.
link |
It's not done with a computer simulation.
link |
So you solve the equation, and that tells you
link |
what the probability that the subject
link |
will be correct on the seventh trial or the experiment is
link |
or something like that.
link |
So it's a use of mathematics to descriptively characterize
link |
aspects of behavior.
link |
And Stanford at that time was the place
link |
where there were several really, really strong
link |
mathematical thinkers who were also connected with three
link |
or four others around the country who brought
link |
a lot of really exciting ideas onto the table.
link |
And it was a very, very prestigious part
link |
of the field of psychology at that time.
link |
So Rummelhart comes into this.
link |
He was a very strong student within that program.
link |
And he got this job at this brand new university
link |
in San Diego in 1967, where he's one of the first assistant
link |
professors in the Department of Psychology at UCSD.
link |
So I got there in 74, seven years later,
link |
and Rummelhart at that time was still
link |
doing mathematical modeling.
link |
But he had gotten interested in cognition.
link |
He'd gotten interested in understanding.
link |
And understanding, I think, remains,
link |
what does it mean to understand anyway?
link |
It's an interesting sort of curious,
link |
how would we know if we really understood something?
link |
But he was interested in building machines
link |
that would hear a couple of sentences
link |
and have an insight about what was going on.
link |
So for example, one of his favorite things at that time
link |
was, Margie was sitting on the front step
link |
when she heard the familiar jingle of the good humor man.
link |
She remembered her birthday money and ran into the house.
link |
What is Margie doing?
link |
Well, there's a couple of ideas you could have,
link |
but the most natural one is that the good humor
link |
man brings ice cream.
link |
She likes ice cream.
link |
She knows she needs money to buy ice cream,
link |
so she's going to run into the house and get her money
link |
so she can buy herself an ice cream.
link |
It's a huge amount of inference that
link |
has to happen to get those things to link up
link |
And he was interested in how the hell that could happen.
link |
And he was trying to build good old fashioned AI style
link |
models of representation of language and content of things
link |
So like formal logic and knowledge bases,
link |
like that kind of stuff.
link |
So he was integrating that with his thinking about cognition.
link |
The mechanisms of cognition, how can they mechanistically
link |
be applied to build these knowledge,
link |
like to actually build something that
link |
looks like a web of knowledge and thereby from there emerges
link |
something like understanding, whatever the heck that is.
link |
Yeah, he was grappling.
link |
This was something that they grappled
link |
with at the end of that book that I was describing,
link |
Explorations in Cognition.
link |
But he was realizing that the paradigm of good old fashioned
link |
AI wasn't giving him the answers to these questions.
link |
By the way, that's called good old fashioned AI now.
link |
It wasn't called that at the time.
link |
It was beginning to be called that.
link |
Oh, because it was from the 60s.
link |
By the late 70s, it was kind of old fashioned,
link |
and it hadn't really panned out.
link |
And people were beginning to recognize that.
link |
And Rommelhardt was like, yeah, he's part of the recognition
link |
that this wasn't all working.
link |
Anyway, so he started thinking in terms of the idea
link |
that we needed systems that allowed us to integrate
link |
multiple simultaneous constraints in a way that would
link |
be mutually influencing each other.
link |
So he wrote a paper that just really, first time I read it,
link |
I said, oh, well, yeah, but is this important?
link |
But after a while, it just got under my skin.
link |
And it was called An Interactive Model of Reading.
link |
And in this paper, he laid out the idea
link |
that every aspect of our interpretation of what's
link |
coming off the page when we read at every level of analysis
link |
you can think of actually depends
link |
on all the other levels of analysis.
link |
So what are the actual pixels making up each letter?
link |
And what do those pixels signify about which letters they are?
link |
And what do those letters tell us about what words are there?
link |
And what do those words tell us about what ideas
link |
the author is trying to convey?
link |
And so he had this model where we
link |
have these little tiny elements that represent
link |
each of the pixels of each of the letters,
link |
and then other ones that represent the line segments
link |
in them, and other ones that represent the letters,
link |
and other ones that represent the words.
link |
And at that time, his idea was there's this set of experts.
link |
There's an expert about how to construct a line out of pixels,
link |
and another expert about which sets of lines
link |
go together to make which letters,
link |
and another one about which letters go together
link |
to make which words, and another one about what
link |
the meanings of the words are, and another one about how
link |
the meanings fit together, and things like that.
link |
And all these experts are looking at this data,
link |
and they're updating hypotheses at other levels.
link |
So the word expert can tell the letter expert,
link |
oh, I think there should be a T there,
link |
because I think there should be a word the here.
link |
And the bottom up sort of feature to letter expert
link |
could say, I think there should be a T there, too.
link |
And if they agree, then you see a T, right?
link |
And so there's a top down, bottom up interactive process,
link |
but it's going on at all layers simultaneously.
link |
So everything can filter all the way down from the top,
link |
as well as all the way up from the bottom.
link |
And it's a completely interactive, bidirectional,
link |
parallel distributed process.
link |
That is somehow, because of the abstractions, it's hierarchical.
link |
So there's different layers of responsibilities,
link |
different levels of responsibilities.
link |
First of all, it's fascinating to think about it
link |
in this kind of mechanistic way.
link |
So not thinking purely from the structure
link |
of a neural network or something like a neural network,
link |
but thinking about these little guys
link |
that work on letters, and then the letters come words
link |
and words become sentences.
link |
And that's a very interesting hypothesis
link |
that from that kind of hierarchical structure
link |
can emerge understanding.
link |
Yeah, so, but the thing is, though,
link |
I wanna just sort of relate this
link |
to the earlier part of the conversation.
link |
When Romelhart was first thinking about it,
link |
there were these experts on the side,
link |
one for the features and one for the letters
link |
and one for how the letters make the words and so on.
link |
And they would each be working,
link |
sort of evaluating various propositions about,
link |
you know, is this combination of features here
link |
going to be one that looks like the letter T and so on.
link |
And what he realized,
link |
kind of after reading Hinton's dissertation
link |
and hearing about Jim Anderson's
link |
linear algebra based neural network models
link |
that I was telling you about before
link |
was that he could replace those experts
link |
with neuron like processing units,
link |
which just would have their connection weights
link |
that would do this job.
link |
So what ended up happening was
link |
that Romelhart and I got together
link |
and we created a model
link |
called the interactive activation model of letter perception,
link |
which takes these little pixel level inputs,
link |
constructs line segment features, letters and words.
link |
But now we built it out of a set of neuron
link |
like processing units that are just connected
link |
to each other with connection weights.
link |
So the unit for the word time has a connection
link |
to the unit for the letter T in the first position
link |
and the letter I in the second position, so on.
link |
And because these connections are bi directional,
link |
if you have prior knowledge that it might be the word time
link |
that starts to prime the letters and the features.
link |
And if you don't, then it has to start bottom up.
link |
But the directionality just depends
link |
on where the information comes in first.
link |
And if you have context together
link |
with features at the same time,
link |
they can convergently result in an emergent perception.
link |
And that was the piece of work that we did together
link |
that sort of got us both completely convinced
link |
that this neural network way of thinking
link |
was going to be able to actually address the questions
link |
that we were interested in as cognitive psychologists.
link |
So the algorithmic side, the optimization side,
link |
those are all details like when you first start the idea
link |
that you can get far with this kind of way of thinking,
link |
that in itself is a profound idea.
link |
So do you like the term connectionism
link |
to describe this kind of set of ideas?
link |
I think it's useful.
link |
It highlights the notion that the knowledge
link |
that the system exploits is in the connections
link |
between the units, right?
link |
There isn't a separate dictionary.
link |
There's just the connections between the units.
link |
So I already sort of laid that on the table
link |
with the connections from the letter units
link |
to the unit for the word time, right?
link |
The unit for the word time isn't a unit for the word time
link |
for any other reason than it's got the connections
link |
to the letters that make up the word time.
link |
Those are the units on the input that excited
link |
when it's excited that it in a sense represents
link |
in the system that there's support for the hypothesis
link |
that the word time is present in the input.
link |
But it's not, the word time isn't written anywhere
link |
inside the bottle, it's only written there
link |
in the picture we drew of the model
link |
to say that's the unit for the word time, right?
link |
And if somebody wants to tell me,
link |
well, how do you spell that word?
link |
You have to use the connections from that out
link |
to then get those letters, for example.
link |
That's such a, that's a counterintuitive idea
link |
where humans want to think in this logic way.
link |
This idea of connectionism, it doesn't, it's weird.
link |
It's weird that this is how it all works.
link |
Yeah, but let's go back to that CNN, right?
link |
That CNN with all those layers of neuron
link |
like processing units that we were talking about before,
link |
it's gonna come out and say, this is a cat, that's a dog,
link |
but it has no idea why it said that.
link |
It's just got all these connections
link |
between all these layers of neurons,
link |
like from the very first layer to the,
link |
you know, like whatever these layers are,
link |
they just get numbered after a while
link |
because they, you know, they somehow further in you go,
link |
the more abstract the features are,
link |
but it's a graded and continuous sort of process
link |
of abstraction anyway.
link |
And, you know, it goes from very local,
link |
very specific to much more sort of global,
link |
but it's still, you know, another sort of pattern
link |
of activation over an array of units.
link |
And then at the output side, it says it's a cat
link |
And when I open my eyes and say, oh, that's Lex,
link |
or, oh, you know, there's my own dog
link |
and I recognize my dog,
link |
which is a member of the same species as many other dogs,
link |
but I know this one
link |
because of some slightly unique characteristics.
link |
I don't know how to describe what it is
link |
that makes me know that I'm looking at Lex
link |
or at my particular dog, right?
link |
Or even that I'm looking at a particular brand of car.
link |
Like I can say a few words about it,
link |
but I wrote you a paragraph about the car,
link |
you would have trouble figuring out
link |
which car is he talking about, right?
link |
So the idea that we have propositional knowledge
link |
of what it is that allows us to recognize
link |
that this is an actual instance
link |
of this particular natural kind
link |
has always been something that it never worked, right?
link |
You couldn't ever write down a set of propositions
link |
for visual recognition.
link |
And so in that space, it sort of always seemed very natural
link |
that something more implicit,
link |
you don't have access to what the details
link |
of the computation were in between,
link |
you just get the result.
link |
So that's the other part of connectionism,
link |
you cannot, you don't read the contents of the connections,
link |
the connections only cause outputs to occur
link |
Yeah, and for us that like final layer
link |
or some particular layer is very important,
link |
the one that tells us that it's our dog
link |
or like it's a cat or a dog,
link |
but each layer is probably equally as important
link |
in the grand scheme of things.
link |
Like there's no reason why the cat versus dog
link |
is more important than the lower level activations,
link |
it doesn't really matter.
link |
I mean, all of it is just this beautiful stacking
link |
on top of each other.
link |
And we humans live in this particular layers,
link |
for us it's useful to survive,
link |
to use those cat versus dog, predator versus prey,
link |
all those kinds of things.
link |
It's fascinating that it's all continuous,
link |
but then you then ask,
link |
the history of artificial intelligence, you ask,
link |
are we able to introspect and convert the very things
link |
that allow us to tell the difference between cat and dog
link |
into a logic, into formal logic?
link |
That's been the dream.
link |
I would say that's still part of the dream of symbolic AI.
link |
And I've recently talked to Doug Lenat who created Psych
link |
and that's a project that lasted for many decades
link |
and still carries a sort of dream in it, right?
link |
But we still don't know the answer, right?
link |
It seems like connectionism is really powerful,
link |
but it also seems like there's this building of knowledge.
link |
And so how do we, how do you square those two?
link |
Like, do you think the connections can contain
link |
the depth of human knowledge and the depth
link |
of what Dave Romahart was thinking about of understanding?
link |
Well, that remains the $64 question.
link |
With inflation, that number is higher.
link |
Maybe it's the $64 billion question now.
link |
You know, I think that from the emergentist side,
link |
which, you know, I placed myself on.
link |
So I used to sometimes tell people
link |
I was a radical, eliminative connectionist
link |
because I didn't want them to think
link |
that I wanted to build like anything into the machine.
link |
But I don't like the word eliminative anymore
link |
because it makes it seem like it's wrong to think
link |
that there is this emergent level of understanding.
link |
And I disagree with that.
link |
So I think, you know, I would call myself
link |
an a radical emergentist connectionist
link |
rather than eliminative connectionist, right?
link |
Because I want to acknowledge
link |
that these higher level kinds of aspects
link |
of our cognition are real, but they're not,
link |
they don't exist as such.
link |
And there was an example that Doug Hofstadter used to use
link |
that I thought was helpful in this respect.
link |
Just the idea that we can think about sand dunes
link |
as entities and talk about like how many there are even.
link |
But we also know that a sand dune is a very fluid thing.
link |
It's a pile of sand that is capable
link |
of moving around under the wind and reforming itself
link |
in somewhat different ways.
link |
And if we think about our thoughts as like sand dunes,
link |
as being things that emerge from just the way
link |
all the lower level elements sort of work together
link |
and are constrained by external forces,
link |
then we can say, yes, they exist as such,
link |
but they also, we shouldn't treat them
link |
as completely monolithic entities that we can understand
link |
without understanding sort of all of the stuff
link |
that allows them to change in the ways that they do.
link |
And that's where I think the connectionist
link |
feeds into the cognitive.
link |
It's like, okay, so if the substrate
link |
is parallel distributed connectionist, then it doesn't mean
link |
that the contents of thought isn't like abstract
link |
and symbolic, but it's more fluid maybe
link |
than it's easier to capture
link |
with a set of logical expressions.
link |
Yeah, that's a heck of a sort of thing
link |
to put at the top of a resume,
link |
radical, emergentist, connectionist.
link |
So there is, just like you said, a beautiful dance
link |
between that, between the machinery of intelligence,
link |
like the neural network side of it,
link |
and the stuff that emerges.
link |
I mean, the stuff that emerges seems to be,
link |
I don't know, I don't know what that is,
link |
that it seems like maybe all of reality is emergent.
link |
What I think about, this is made most distinctly rich to me
link |
when I look at cellular automata, look at game of life,
link |
that from very, very simple things,
link |
very rich, complex things emerge
link |
that start looking very quickly like organisms
link |
that you forget how the actual thing operates.
link |
They start looking like they're moving around,
link |
they're eating each other,
link |
some of them are generating offspring.
link |
You forget very quickly.
link |
And it seems like maybe it's something
link |
about the human mind that wants to operate
link |
in some layer of the emergent,
link |
and forget about the mechanism
link |
of how that emergence happens.
link |
So it, just like you are in your radicalness,
link |
I'm also, it seems like unfair
link |
to eliminate the magic of that emergent,
link |
like eliminate the fact that that emergent is real.
link |
Yeah, no, I agree.
link |
I'm not, that's why I got rid of eliminative, right?
link |
Eliminative, yeah.
link |
Yeah, because it seemed like that was trying to say
link |
that it's all completely like.
link |
An illusion of some kind, it's not.
link |
Well, who knows whether there isn't,
link |
there aren't some illusory characteristics there.
link |
And I think that philosophically many people
link |
have confronted that possibility over time,
link |
but it's still important to accept it as magic, right?
link |
So, I think of Fellini in this context,
link |
I think of others who have appreciated the role of magic,
link |
the role of magic, of actual trickery
link |
in creating illusions that move us.
link |
And Plato was on to this too.
link |
It's like somehow or other these shadows
link |
give rise to something much deeper than that.
link |
And that's, so we won't try to figure out what it is.
link |
We'll just accept it as given that that occurs.
link |
And, you know, but he was still onto the magic of it.
link |
Yeah, yeah, we won't try to really, really,
link |
really deeply understand how it works.
link |
We'll just enjoy the fact that it's kind of fun.
link |
Okay, but you worked closely with Dave Romo Hart.
link |
He passed away as a human being.
link |
What do you remember about him?
link |
Do you miss the guy?
link |
Absolutely, you know, he passed away 15ish years ago now.
link |
And his demise was actually one of the most poignant
link |
and, you know, like relevant tragedies, relevant to our conversation.
link |
He started to undergo a progressive neurological condition
link |
that isn't far from what we're used to.
link |
A neurological condition that isn't fully understood.
link |
That is to say his particular course isn't fully understood
link |
because, you know, brain scans weren't done at certain stages
link |
and no autopsy was done or anything like that.
link |
The wishes of the family.
link |
We don't know as much about the underlying pathology as we might,
link |
but I had begun to get interested in this neurological condition
link |
that might have been the very one that he was succumbing to
link |
as my own efforts to understand another aspect of this mystery
link |
that we've been discussing while he was beginning
link |
to get progressively more and more affected.
link |
So I'm going to talk about the disorder
link |
and not about Rumelhart for a second, okay?
link |
The disorder is something my colleagues and collaborators
link |
have chosen to call semantic dementia.
link |
So it's a specific form of loss of mind
link |
related to meaning, semantic dementia.
link |
And it's progressive in the sense that the patient loses the ability
link |
to appreciate the meaning of the experiences that they have,
link |
either from touch, from sight, from sound, from language.
link |
They, I hear sounds, but I don't know what they mean kind of thing.
link |
So as this illness progresses, it starts with the patient
link |
being unable to differentiate like similar breeds of dog
link |
or remember the lower frequency unfamiliar categories
link |
that they used to be able to remember.
link |
But as it progresses, it becomes more and more striking
link |
and the patient loses the ability to recognize things like
link |
pigs and goats and sheep and calls all middle sized animals dogs
link |
and can't recognize rabbits and rodents anymore.
link |
They call all the little ones cats
link |
and they can't recognize hippopotamuses and cows anymore.
link |
They call them all horses.
link |
So there was this one patient who went through this progression
link |
where at a certain point, any four legged animal,
link |
he would call it either a horse or a dog or a cat.
link |
And if it was big, he would tend to call it a horse.
link |
If it was small, he'd tend to call it a cat.
link |
Middle sized ones, he called dogs.
link |
This is just a part of the syndrome though.
link |
The patient loses the ability to relate concepts to each other.
link |
So my collaborator in this work, Carolyn Patterson,
link |
developed a test called the pyramids and palm trees test.
link |
So you give the patient a picture of pyramids
link |
and they have a choice which goes with the pyramids,
link |
palm trees or pine trees.
link |
And she showed that this wasn't just a matter of language
link |
because the patient's loss of this ability shows up
link |
whether you present the material with words or with pictures.
link |
The pictures, they can't put the pictures together
link |
with each other properly anymore.
link |
They can't relate the pictures to the words either.
link |
They can't do word picture matching.
link |
But they've lost the conceptual grounding
link |
from either modality of input.
link |
And so that's why it's called semantic dementia.
link |
The very semantics is disintegrating.
link |
And we understand this in terms of our idea
link |
that distributed representation, a pattern of activation,
link |
represents the concepts, really similar ones.
link |
As you degrade them, they start being,
link |
you lose the differences.
link |
So the difference between the dog and the goat
link |
is no longer part of the pattern anymore.
link |
And since dog is really familiar,
link |
that's the thing that remains.
link |
And we understand that in the way the models work and learn.
link |
But Rumelhart underwent this condition.
link |
So on the one hand, it's a fascinating aspect
link |
of parallel distributed processing to be.
link |
It reveals this sort of texture of distributed representation
link |
in a very nice way, I've always felt.
link |
But at the same time, it was extremely poignant
link |
because this is exactly the condition
link |
that Rumelhart was undergoing.
link |
And there was a period of time when he was this man
link |
who had been the most focused, goal directed, competitive,
link |
thoughtful person who was willing to work for years
link |
to solve a hard problem, he starts to disappear.
link |
And there was a period of time when it was hard for any of us
link |
to really appreciate that he was sort of, in some sense,
link |
not fully there anymore.
link |
Do you know if he was able to introspect
link |
the solution of the understanding mind?
link |
I mean, this is one of the big scientists that thinks about this.
link |
Was he able to look at himself and understand the fading mind?
link |
You know, we can contrast Hawking and Rumelhart in this way.
link |
And I like to do that to honor Rumelhart
link |
because I think Rumelhart is sort of like the Hawking
link |
of cognitive science to me in some ways.
link |
Both of them suffered from a degenerative condition.
link |
In Hawking's case, it affected the motor system.
link |
In Rumelhart's case, it's affecting the semantics.
link |
And not just the pure object semantics,
link |
but maybe the self semantics as well.
link |
And we don't understand that.
link |
So I would say he didn't.
link |
And this was part of what, from the outside,
link |
was a profound tragedy.
link |
But on the other hand, at some level, he sort of did
link |
because there was a period of time when it finally was realized
link |
that he had really become profoundly impaired.
link |
This was clearly a biological condition.
link |
It wasn't just like he was distracted that day or something like that.
link |
So he retired from his professorship at Stanford
link |
and he became, he lived with his brother for a couple years
link |
and then he moved into a facility for people with cognitive impairments.
link |
One that many elderly people end up in when they have cognitive impairments.
link |
And I would spend time with him during that period.
link |
This was like in the late 90s, around 2000 even.
link |
And we would go bowling and he could still bowl.
link |
And after bowling, I took him to lunch and I said,
link |
where would you like to go?
link |
You want to go to Wendy's?
link |
And I said, okay, well, where do you want to go?
link |
And he just pointed.
link |
He said, turn here.
link |
So he still had a certain amount of spatial cognition
link |
and he could get me to the restaurant.
link |
And then when we got to the restaurant, I said,
link |
what do you want to order?
link |
And he couldn't come up with any of the words,
link |
but he knew where on the menu the thing was that he wanted.
link |
So it's, you know, and he couldn't say what it was,
link |
but he knew that that's what he wanted to eat.
link |
And so it's like it isn't monolithic at all.
link |
Our cognition is, you know, first of all, graded in certain kinds of ways,
link |
but also multipartite and there's many elements to it and things,
link |
certain sort of partial competencies still exist
link |
in the absence of other aspects of these competencies.
link |
So this is what always fascinated me about what used to be called
link |
cognitive neuropsychology, you know,
link |
the effects of brain damage on cognition.
link |
But in particular, this gradual disintegration part.
link |
You know, I'm a big believer that the loss of a human being that you value
link |
is as powerful as, you know, first falling in love with that human being.
link |
I think it's all a celebration of the human being.
link |
So the disintegration itself too is a celebration in a way.
link |
But just to say something more about the scientist
link |
and the backpropagation idea that you mentioned.
link |
So in 1982, Hinton had been there as a postdoc and organized that conference.
link |
He'd actually gone away and gotten an assistant professorship
link |
and then there was this opportunity to bring him back.
link |
So Jeff Hinton was back on a sabbatical.
link |
And Rommelhard and I had decided we wanted to do this, you know,
link |
we thought it was really exciting and the papers on the interactive activation model
link |
that I was telling you about had just been published
link |
and we both sort of saw a huge potential for this work and Jeff was there.
link |
And so the three of us started a research group,
link |
which we called the PDP Research Group.
link |
And several other people came.
link |
Francis Crick, who was at the Salk Institute, heard about it from Jeff
link |
because Jeff was known among Brits to be brilliant
link |
and Francis was well connected with his British friends.
link |
So Francis Crick came.
link |
That's a heck of a group of people, wow.
link |
And Paul Spolensky was one of the other postdocs.
link |
He was still there as a postdoc.
link |
And a few other people.
link |
But anyway, Jeff talked to us about learning
link |
and how we should think about how, you know, learning occurs in a neural network.
link |
And he said, the problem with the way you guys have been approaching this
link |
is that you've been looking for inspiration from biology
link |
to tell you what the rules should be for how the synapses should change
link |
the strengths of their connections, how the connections should form.
link |
He said, that's the wrong way to go about it.
link |
What you should do is you should think in terms of
link |
how you can adjust connection weights to solve a problem.
link |
So you define your problem and then you figure out
link |
how the adjustment of the connection weights will solve the problem.
link |
And Rumelhart heard that and said to himself, okay,
link |
so I'm going to start thinking about it that way.
link |
I'm going to essentially imagine that I have some objective function,
link |
some goal of the computation.
link |
I want my machine to correctly classify all of these images.
link |
And I can score that.
link |
I can measure how well they're doing on each image.
link |
And I get some measure of error or loss, it's typically called in deep learning.
link |
And I'm going to figure out how to adjust the connection weights
link |
so as to minimize my loss or reduce the error.
link |
And that's called, you know, gradient descent.
link |
And engineers were already familiar with the concept of gradient descent.
link |
And in fact, there was an algorithm called the delta rule
link |
that had been invented by a professor in the electrical engineering department
link |
at Stanford, Bernie Widrow and a collaborator named Hoff.
link |
So gradient descent in continuous neural networks
link |
with multiple neuron like processing units was already understood
link |
for a single layer of connection weights.
link |
We have some inputs over a set of neurons.
link |
We want the output to produce a certain pattern.
link |
We can define the difference between our target
link |
and what the neural network is producing.
link |
And we can figure out how to change the connection weights to reduce that error.
link |
So what Romilhar did was to generalize that
link |
so as to be able to change the connections from earlier layers of units
link |
to the ones at a hidden layer between the input and the output.
link |
And so he first called the algorithm the generalized delta rule
link |
because it's just an extension of the gradient descent idea.
link |
And interestingly enough, Hinton was thinking that this wasn't going to work very well.
link |
So Hinton had his own alternative algorithm at the time
link |
based on the concept of the Boltzmann machine that he was pursuing.
link |
So the paper on the Boltzmann machine came out in,
link |
learning in Boltzmann machines came out in 1985.
link |
But it turned out that back prop worked better than the Boltzmann machine learning algorithm.
link |
So this generalized delta algorithm ended up being called back propagation, as you say, back prop.
link |
Yeah. And probably that name is opaque to me.
link |
What does that mean?
link |
What it meant was that in order to figure out what the changes you needed to make
link |
to the connections from the input to the hidden layer,
link |
you had to back propagate the error signals from the output layer
link |
through the connections from the hidden layer to the output
link |
to get the signals that would be the error signals for the hidden layer.
link |
And that's how Rumelhart formulated it.
link |
It was like, well, we know what the error signals are at the output layer.
link |
Let's see if we can get a signal at the hidden layer
link |
that tells each hidden unit what its error signal is essentially.
link |
So it's back propagating through the connections
link |
from the hidden to the output to get the signals to tell the hidden units
link |
how to change their weights from the input.
link |
And that's why it's called back prop.
link |
Yeah. But so it came from Hinton having introduced the concept of, you know,
link |
define your objective function, figure out how to take the derivative
link |
so that you can adjust the connections so that they make progress towards your goal.
link |
So stop thinking about biology for a second
link |
and let's start to think about optimization and computation a little bit more.
link |
So what about Jeff Hinton?
link |
You've gotten a chance to work with him in that little thing.
link |
The set of people involved there is quite incredible.
link |
The small set of people under the PDP flag,
link |
it's just given the amount of impact those ideas have had over the years,
link |
it's kind of incredible to think about.
link |
But, you know, just like you said, like yourself,
link |
Jeffrey Hinton is seen as one of the, not just like a seminal figure in AI,
link |
but just a brilliant person,
link |
just like the horsepower of the mind is pretty high up there for him
link |
because he's just a great thinker.
link |
So what kind of ideas have you learned from him?
link |
Have you influenced each other on?
link |
Have you debated over what stands out to you in the full space of ideas here
link |
at the intersection of computation and cognition?
link |
Well, so Jeff has said many things to me that had a profound impact on my thinking.
link |
And he's written several articles which were way ahead of their time.
link |
He had two papers in 1981, just to give one example,
link |
one of which was essentially the idea of transformers
link |
and another of which was an early paper on semantic cognition
link |
which inspired him and Rumelhart and me throughout the 80s
link |
and, you know, still I think sort of grounds my own thinking
link |
about the semantic aspects of cognition.
link |
He also, in a small paper that was never published that he wrote in 1977,
link |
you know, before he actually arrived at UCSD or maybe a couple years even before that,
link |
I don't know, when he was a PhD student,
link |
he described how a neural network could do recursive computation.
link |
And it was a very clever idea that he's continued to explore over time,
link |
which was sort of the idea that when you call a subroutine,
link |
you need to save the state that you had when you called it
link |
so you can get back to where you were when you're finished with the subroutine.
link |
And the idea was that you would save the state of the calling routine
link |
by making fast changes to connection weights.
link |
And then when you finished with the subroutine call,
link |
those fast changes in the connection weights would allow you to go back
link |
to where you had been before and reinstate the previous context
link |
so that you could continue on with the top level of the computation.
link |
Anyway, that was part of the idea.
link |
And I always thought, okay, that's really, you know,
link |
he had extremely creative ideas that were quite a lot ahead of his time
link |
and many of them in the 1970s and early 1980s.
link |
So another thing about Geoff Hinton's way of thinking,
link |
which has profoundly influenced my effort to understand
link |
human mathematical cognition, is that he doesn't write too many equations.
link |
And people tell stories like, oh, in the Hinton Lab meetings,
link |
you don't get up at the board and write equations
link |
like you do in everybody else's machine learning lab.
link |
What you do is you draw a picture.
link |
And, you know, he explains aspects of the way deep learning works
link |
by putting his hands together and showing you the shape of a ravine
link |
and using that as a geometrical metaphor for what's happening
link |
as this gradient descent process.
link |
You're coming down the wall of a ravine.
link |
If you take too big a jump, you're going to jump to the other side.
link |
And so that's why we have to turn down the learning rate, for example.
link |
And it speaks to me of the fundamentally intuitive character of deep insight
link |
together with a commitment to really understanding
link |
in a way that's absolutely ultimately explicit and clear, but also intuitive.
link |
Yeah, there's certain people like that.
link |
Here's an example, some kind of weird mix of visual and intuitive
link |
and all those kinds of things.
link |
Feynman is another example, different style of thinking, but very unique.
link |
And when you're around those people, for me in the engineering realm,
link |
there's a guy named Jim Keller who's a chip designer, engineer.
link |
Every time I talk to him, it doesn't matter what we're talking about.
link |
Just having experienced that unique way of thinking transforms you
link |
and makes your work much better.
link |
And that's the magic.
link |
You look at Daniel Kahneman, you look at the great collaborations
link |
throughout the history of science.
link |
That's the magic of that.
link |
It's not always the exact ideas that you talk about,
link |
but it's the process of generating those ideas.
link |
Being around that, spending time with that human being,
link |
you can come up with some brilliant work,
link |
especially when it's cross disciplinary as it was a little bit in your case with Jeff.
link |
Jeff is a descendant of the logician Boole.
link |
He comes from a long line of English academics.
link |
And together with the deeply intuitive thinking ability that he has,
link |
he also has, it's been clear, he's described this to me,
link |
and I think he's mentioned it from time to time in other interviews
link |
that he's had with people.
link |
He's wanted to be able to sort of think of himself as contributing
link |
to the understanding of reasoning itself, not just human reasoning.
link |
Like Boole is about logic, right?
link |
It's about what can we conclude from what else and how do we formalize that.
link |
And as a computer scientist, logician, philosopher,
link |
the goal is to understand how we derive truths from other,
link |
from givens and things like this.
link |
And the work that Jeff was doing in the early to mid 80s
link |
on something called the Bolton machine was his way of connecting
link |
with that Boolean tradition and bringing it into the more continuous,
link |
probabilistic graded constraint satisfaction realm.
link |
And it was a beautiful set of ideas linked with theoretical physics
link |
as well as with logic.
link |
And it's always been, I mean, I've always been inspired
link |
by the Bolton machine too.
link |
It's like, well, if the neurons are probabilistic rather than deterministic
link |
in their computations, then maybe this somehow is part of the serendipity
link |
or adventitiousness of the moment of insight, right?
link |
It might not have occurred at that particular instant.
link |
It might be sort of partially the result of a stochastic process.
link |
And that too is part of the magic of the emergence of some of these things.
link |
Well, you're right with the Boolean lineage and the dream of computer science
link |
is somehow, I mean, I certainly think of humans this way,
link |
that humans are one particular manifestation of intelligence,
link |
that there's something bigger going on and you're hoping to figure that out.
link |
The mechanisms of intelligence, the mechanisms of cognition
link |
are much bigger than just humans.
link |
Yeah. So I think of, I started using the phrase computational intelligence
link |
at some point as to characterize the field that I thought, you know,
link |
people like Geoff Hinton and many of the people I know at DeepMind
link |
are working in and where I feel like I'm, you know,
link |
I'm a kind of a human oriented computational intelligence researcher
link |
in that I'm actually kind of interested in the human solution.
link |
But at the same time, I feel like that's where a huge amount
link |
of the excitement of deep learning actually lies is in the idea that,
link |
you know, we may be able to even go beyond what we can achieve
link |
with our own nervous systems when we build computational intelligences
link |
that are, you know, not limited in the ways that we are by our own biology.
link |
Perhaps allowing us to scale the very mechanisms of human intelligence
link |
just increases power through scale.
link |
Yes. And I think that that, you know, obviously that's the,
link |
that's being played out massively at Google Brain, at OpenAI
link |
and to some extent at DeepMind as well.
link |
I guess I shouldn't say to some extent.
link |
Just the massive scale of the computations that are used to succeed
link |
at games like Go or to solve the protein folding problems
link |
that they've been solving and so on.
link |
Still not as many synapses and neurons as the human brain.
link |
So we still got, we're still beating them on that.
link |
We humans are beating the AIs, but they're catching up pretty quickly.
link |
You write about modeling of mathematical cognition.
link |
So let me first ask about mathematics in general.
link |
There's a paper titled Parallel Distributed Processing
link |
Approach to Mathematical Cognition where in the introduction
link |
there's some beautiful discussion of mathematics.
link |
And you referenced there Tristan Needham who criticizes a narrow
link |
form of view of mathematics by liking the studying of mathematics
link |
as symbol manipulation to studying music without ever hearing a note.
link |
So from that perspective, what do you think is mathematics?
link |
What is this world of mathematics like?
link |
Well, I think of mathematics as a set of tools for exploring
link |
idealized worlds that often turn out to be extremely relevant
link |
to the real world but need not.
link |
But they're worlds in which objects exist with idealized properties
link |
and in which the relationships among them can be characterized
link |
with precision so as to allow the implications of certain facts
link |
to then allow you to derive other facts with certainty.
link |
So if you have two triangles and you know that there is an angle
link |
in the first one that has the same measure as an angle in the second one
link |
and you know that the lengths of the sides adjacent to that angle
link |
in each of the two triangles, the corresponding sides adjacent
link |
to that angle also have the same measure, then you can then conclude
link |
that the triangles are congruent.
link |
That is to say they have all of their properties in common.
link |
And that is something about triangles.
link |
It's not a matter of formulas.
link |
These are idealized objects.
link |
In fact, we built bridges out of triangles and we understand
link |
how to measure the height of something we can't climb by extending
link |
these ideas about triangles a little further.
link |
And all of the ability to get a tiny speck of matter launched
link |
from the planet Earth to intersect with some tiny, tiny little body
link |
way out in way beyond Pluto somewhere at exactly a predicted time
link |
and date is something that depends on these ideas.
link |
And it's actually happening in the real physical world that these ideas
link |
make contact with it in those kinds of instances.
link |
But there are these idealized objects, these triangles or these distances
link |
or these points, whatever they are, that allow for this set of tools
link |
to be created that then gives human beings this incredible leverage
link |
that they didn't have without these concepts.
link |
And I think this is actually already true when we think about just,
link |
you know, the natural numbers.
link |
I always like to include zero, so I'm going to say the nonnegative integers,
link |
but that's a place where some people prefer not to include zero.
link |
We like zero here, natural numbers, zero, one, two, three, four, five,
link |
six, seven, and so on.
link |
Yeah. And because they give you the ability to be exact about
link |
how many sheep you have.
link |
I sent you out this morning, there were 23 sheep.
link |
You came back with only 22. What happened?
link |
The fundamental problem of physics, how many sheep you have.
link |
It's a fundamental problem of human society that you damn well better
link |
bring back the same number of sheep as you started with.
link |
And it allows commerce, it allows contracts, it allows the establishment
link |
of records and so on to have systems that allow these things to be notated.
link |
But they have an inherent aboutness to them that's one in the same time sort of
link |
abstract and idealized and generalizable, while on the other hand,
link |
potentially very, very grounded and concrete.
link |
And one of the things that makes for the incredible achievements of the human mind
link |
is the fact that humans invented these idealized systems that leverage
link |
the power of human thought in such a way as to allow all this kind of thing to happen.
link |
And so that's what mathematics to me is the development of systems for thinking about
link |
the properties and relations among sets of idealized objects and
link |
the mathematical notation system that we unfortunately focus way too much on
link |
is just our way of expressing propositions about these properties.
link |
It's just like we're talking with Chomsky in language.
link |
It's the thing we've invented for the communication of those ideas.
link |
They're not necessarily the deep representation of those ideas.
link |
So what's a good way to model such powerful mathematical reasoning, would you say?
link |
What are some ideas you have for capturing this in a model?
link |
The insights that human mathematicians have had is a combination of the kind of the
link |
intuitive kind of connectionist like knowledge that makes it so that something is just like
link |
obviously true so that you don't have to think about why it's true.
link |
That then makes it possible to then take the next step and ponder and reason and
link |
figure out something that you previously didn't have that intuition about.
link |
It then ultimately becomes a part of the intuition that the next generation of
link |
mathematical thinkers have to ground their own thinking on so that they can extend the ideas even further.
link |
I came across this quotation from Henri Poincare while I was walking in the woods with my wife
link |
in a state park in Northern California late last summer.
link |
And what it said on the bench was it is by logic that we prove but by intuition that we discover.
link |
And so what for me the essence of the project is to understand how to bring the intuitive
link |
connectionist resources to bear on letting the intuitive discovery arise from engagement in
link |
thinking with this formal system.
link |
So I think of the ability of somebody like Hinton or Newton or Einstein or Rumelhart or
link |
Poincare to Archimedes is another example.
link |
So suddenly a flash of insight occurs. It's like the constellation of all of these
link |
simultaneous constraints that somehow or other causes the mind to settle into a novel state that
link |
it never did before and give rise to a new idea that then you can say, okay, well, now how can I
link |
prove this? How do I write down the steps of that theorem that allow me to make it rigorous and certain?
link |
And so I feel like the kinds of things that we're beginning to see deep learning systems do of
link |
their own accord kind of gives me this feeling of hope or encouragement that ultimately it'll all happen.
link |
So in particular as many people now have become really interested in thinking about, you know,
link |
neural networks that have been trained with massive amounts of text can be given a prompt and they
link |
can then sort of generate some really interesting, fanciful, creative story from that prompt.
link |
And there's kind of like a sense that they've somehow synthesized something like novel out of
link |
the, you know, all of the particulars of all of the billions and billions of experiences that went
link |
into the training data that gives rise to something like this sort of intuitive sense of what would
link |
be a fun and interesting little story to tell or something like that. It just sort of wells up out
link |
of the letting the thing play out its own imagining of what somebody might say given this prompt as
link |
an input to get it to start to generate its own thoughts. And to me that sort of represents the
link |
potential of capturing the intuitive side of this.
link |
And there's other examples, I don't know if you find them as captivating is, you know, on the
link |
DeepMind side with AlphaZero, if you study chess, the kind of solutions that has come up in terms
link |
of chess, it is, there's novel ideas there. It feels very like there's brilliant moments of insight.
link |
And the mechanism they use, if you think of search as maybe more towards good old fashioned AI and
link |
then there's the connection is the neural network that has the intuition of looking at a board,
link |
looking at a set of patterns and saying, how good is this set of positions? And the next few
link |
positions, how good are those? And that's it. That's just an intuition. Grandmasters have this
link |
and understanding positionally, tactically, how good the situation is, how can it be improved
link |
without doing this full, like deep search. And then maybe doing a little bit of what human chess
link |
players call calculation, which is the search, taking a particular set of steps down the line to
link |
see how they unroll. But there is moments of genius in those systems too. So that's another hopeful
link |
illustration that from neural networks can emerge this novel creation of an idea.
link |
Yes. And I think that, you know, I think Demis Hassabis is, you know, he's spoken about those
link |
things. I heard him describe a move that was made in one of the go matches against Lisa Dahl in a
link |
very similar way. And it caused me to become really excited to kind of collaborate with some of those
link |
people and analyze it at DeepMind. So I think though that what I like to really emphasize here
link |
is one part of what I like to emphasize about mathematical cognition at least is that philosophers
link |
and logicians going back three or even a little more than 3000 years ago began to develop these
link |
formal systems and gradually the whole idea about thinking formally got constructed. And, you know,
link |
it's preceded Euclid, certainly present in the work of Thales and others. And I'm not the world's
link |
leading expert in all the details of that history, but Euclid's elements were the kind of the touch
link |
point of a coherent document that sort of laid out this idea of an actual formal system within which
link |
these objects were characterized and the system of inference that allowed new truths to be derived
link |
from others was sort of like established as a paradigm. And what I find interesting is the
link |
idea that the ability to become a person who is capable of thinking in this abstract formal way
link |
is a result of the same kind of immersion in experience thinking in that way that we now
link |
begin to think of our understanding of language as being, right? So, we immerse ourselves in a
link |
particular language, in a particular world of objects and their relationships and we learn
link |
to talk about that and we develop intuitive understanding of the real world. In a similar
link |
way, we can think that what academia has created for us, what those early philosophers and their
link |
academies in Athens and Alexandria and other places allowed was the development of these
link |
schools of thought, modes of thought that then become deeply ingrained and it becomes what it
link |
is that makes it so that somebody like Jerry Fodor would think that systematic thought is
link |
the essential characteristic of the human mind as opposed to a derived and an acquired characteristic
link |
that results from acculturation in a certain mode that's been invented by humans.
link |
Would you say it's more fundamental than like language? If we start dancing, if we bring
link |
Chomsky back into the conversation, first of all, is it unfair to draw a line between mathematical
link |
cognition and language, linguistic cognition?
link |
I think that's a very interesting question and I think it's one of the ones that I'm actually very
link |
interested in right now, but I think the answer is in important ways, it is important to draw that
link |
line, but then to come back and look at it again and see some of the subtleties and interesting
link |
aspects of the difference. So if we think about Chomsky himself, he was born into an academic
link |
family. His father was a professor of rabbinical studies at a small rabbinical college in
link |
Philadelphia. He was deeply enculturated in a culture of thought and reason and brought to the
link |
effort to understand natural language, this profound engagement with these formal systems. I
link |
think that there was tremendous power in that and that Chomsky had some amazing insights into the
link |
structure of natural language, but that, I'm going to use the word but there, the actual intuitive
link |
knowledge of these things only goes so far and does not go as far as it does in people like
link |
Chomsky himself. And this was something that was discovered in the PhD dissertation of Lyla
link |
Gleitman, who was actually trained in the same linguistics department with Chomsky. So what Lyla
link |
discovered was that the intuitions that linguists had about even the meaning of a phrase, not just
link |
about its grammar, but about what they thought a phrase must mean were very different from the
link |
intuitions of an ordinary person who wasn't a formally trained thinker. And well, it recently
link |
has become much more salient. I happened to have learned about this when I myself was a PhD student
link |
at the University of Pennsylvania, but I never knew how to put it together with all of my other
link |
thinking about these things. So I actually currently have the hypothesis that formally
link |
trained linguists and other formally trained academics, whether it be linguistics, philosophy,
link |
cognitive science, computer science, machine learning, mathematics,
link |
have a mode of engagement with experience that is intuitively deeply structured to be more
link |
organized around the systematicity and ability to be conformant with the principles of a system
link |
than is actually true of the natural human mind without that immersion.
link |
That's fascinating. So the different fields and approaches with which you start to study the mind
link |
actually take you away from the natural operation of the mind. So it makes it very difficult for you
link |
to be somebody who introspects.
link |
Yes. And this is where things about human belief and so called knowledge that we consider
link |
private, not our business to manipulate in others. We are not entitled to tell somebody else what to
link |
believe about certain kinds of things. What are those beliefs? Well, they are the product of this
link |
sort of immersion and enculturation. That is what I believe.
link |
And that's limiting.
link |
It's something to be aware of.
link |
Does that limit you from having a good model of cognition?
link |
So when you look at mathematical or linguistics, I mean, what is that line then? So is Chomsky
link |
unable to sneak up to the full picture of cognition? Are you, when you're focusing on
link |
mathematical thinking, are you also unable to do so?
link |
I think you're right. I think that's a great way of characterizing it. And
link |
I also think that it's related to the concept of beginner's mind and another concept called the
link |
expert blind spot. So the expert blind spot is much more prosaic seeming than this point that
link |
you were just making. But it's something that plagues experts when they try to communicate
link |
their understanding to non experts. And that is that things are self evident to them that
link |
they can't begin to even think about how they could explain it to somebody else.
link |
Because it's just like so patently obvious that it must be true. And
link |
when Kronacker said, God made the natural numbers, all else is the work of man,
link |
he was expressing that intuition that somehow or other, the basic fundamentals of discrete
link |
quantities being countable and innumerable and indefinite in number was not something that
link |
had to be discovered. But he was wrong. It turns out that many cognitive scientists
link |
agreed with him for a time. There was a long period of time where the natural
link |
numbers were considered to be a part of the innate endowment of core knowledge or to use
link |
the kind of phrases that Spelke and Kerry used to talk about what they believe are
link |
the innate primitives of the human mind. And they no longer believe that. It's actually
link |
been more or less accepted by almost everyone that the natural numbers are actually a cultural
link |
construction. And it's so interesting to go back and study those few people who still exist who
link |
don't have those systems. So this is just an example to me where a certain mode of thinking
link |
about language itself or a certain mode of thinking about geometry and those kinds of
link |
relations. So it becomes so second nature that you don't know what it is that you need to teach. And
link |
in fact, we don't really teach it all that explicitly anyway. You take a math class,
link |
the professor sort of teaches it to you the way they understand it. Some of the students in the
link |
class sort of like they get it. They start to get the way of thinking and they can actually do the
link |
problems that get put on the homework that the professor thinks are interesting and challenging
link |
ones. But most of the students who don't kind of engage as deeply don't ever get. And we think,
link |
oh, that man must be brilliant. He must have this special insight. But he must have some
link |
some biological sort of bit that's different, that makes him so that he or she could have
link |
that insight. Although I don't want to dismiss biological individual differences completely,
link |
I find it much more interesting to think about the possibility that it was that difference in the
link |
dinner table conversation at the Chomsky house when he was growing up that made it so that he
link |
had that cast of mind. Yeah. And there's a few topics we talked about that kind of interconnect
link |
because I wonder the better I get at certain things, we humans, the deeper we understand
link |
something, what are you starting to then miss about the rest of the world? We talked about David
link |
and his degenerative mind. And, you know, when you look in the mirror and wonder how different
link |
am I am I cognitively from the man I was a month ago, from the man I was a year ago, like what,
link |
you know, if I can, having thought about language of Chomsky for 10, 20 years, what am I no longer
link |
able to see? What is in my blind spot? And how big is that? And then to somehow be able to leap back
link |
out of your deep, like structure that you form for yourself about thinking about the world,
link |
leap back and look at the big picture again, or jump out of the your current way of thinking.
link |
And to be able to introspect, like what are the limitations of your mind? How is your mind less
link |
powerful than it used to be or more powerful or different, powerful in different ways? So that
link |
seems to be a difficult thing to do because we're living, we're looking at the world through the
link |
lens of our mind, right? To step outside and introspect is difficult, but it seems necessary
link |
if you want to make progress. You know, one of the threads of psychological research that's always
link |
been very, I don't know, important to me to be aware of is the idea that our explanations of our
link |
own behavior aren't necessarily actually part of the causal process that caused that behavior to
link |
occur, or even valid observations of the set of constraints that led to the outcome, but they are
link |
post hoc rationalizations that we can give based on information at our disposal about what might
link |
have contributed to the result that we came to when asked. And so this is an idea that was
link |
introduced in a very important paper by Nisbet and Wilson about, you know, the limits on our ability
link |
to be aware of the factors that cause us to make the choices that we make. And, you know, I think
link |
it's something that we really ought to be much more cognizant of, in general, as human beings,
link |
is that our own insight into exactly why we hold the beliefs that we do and we hold the attitudes
link |
and make the choices and feel the feelings that we do is not something that we totally control
link |
or totally observe. And it's subject to, you know, our culturally transmitted understanding of what
link |
it is that is the mode that we give to explain these things when asked to do so as much as it is
link |
about anything else. And so even our ability to introspect and think we have access to our own
link |
thoughts is a product of culture and belief, you know, practice.
link |
So let me ask you the big question of advice. So you've lived an incredible life in terms of the
link |
ideas you've put out into the world, in terms of the trajectory you've taken through your career,
link |
through your life. What advice would you give to young people today, in high school, in college,
link |
about how to have a career or how to have a life they can be proud of?
link |
Finding the thing that you are intrinsically motivated to engage with and then celebrating
link |
that discovery is what it's all about. When I was in college, I struggled with that. I had thought
link |
I wanted to be a psychiatrist because I think I was interested in human psychology in high school.
link |
And at that time, the only sort of information I had that had anything to do with the psyche was,
link |
you know, Freud and Erich Fromm and sort of popular psychiatry kinds of things.
link |
And so, well, they were psychiatrists, right? So I had to be a psychiatrist.
link |
And that meant I had to go to medical school. And I got to college and I find myself taking,
link |
you know, the first semester of a three quarter physics class and it was mechanics. And this was
link |
so far from what it was I was interested in, but it was also too early in the morning in the winter
link |
court semester. So I never made it to the physics class. But I wondered about the rest of my
link |
freshman year and most of my sophomore year until I found myself in the midst of this situation where
link |
around me there was this big revolution happening. I was at Columbia University in 1968 and
link |
the Vietnam War is going on. Columbia is building a gym in Morningside Heights, which is part of
link |
Harlem. And people are thinking, oh, the big bad rich guys are stealing the parkland that
link |
belongs to the people of Harlem. And, you know, they're part of the military industrial complex,
link |
which is enslaving us and sending us all off to war in Vietnam. And so there was a big revolution
link |
that involved a confluence of black activism and, you know, SDS and social justice and the whole
link |
university blew up and got shut down. And I got a chance to sort of think about
link |
why people were behaving the way they were in this context. And I, you know, I happened to have
link |
taken mathematical statistics. I happened to have been taking psychology that quarter at just cycle
link |
one. And somehow things in that space all ran together in my mind and got me really excited
link |
about asking questions about why people, what made certain people go into the buildings and not
link |
others and things like that. And so suddenly I had a path forward and I had just been wandering
link |
around aimlessly. And at the different points in my career, you know, and I think, okay,
link |
well, should I take this class or should I just read that book about some idea that I want to
link |
understand better, you know, or should I pursue the thing that excites me and interests me or
link |
should I, you know, meet some requirement? You know, that's, I always did the latter.
link |
So I ended up, my professors in psychology were, thought I was great. They wanted me to go to
link |
graduate school. They nominated me for Phi Beta Kappa. And I went to the Phi Beta Kappa ceremony
link |
and this guy came up and he said, oh, are you Magna Arsuma? And I wasn't even getting honors
link |
based on my grades. They just happened to have thought I was interested enough in ideas to
link |
belong to Phi Beta Kappa. So. I mean, would it be fair to say you kind of stumbled around a little
link |
bit through accidents of too early morning of classes in physics and so on until you discovered
link |
intrinsic motivation, as you mentioned, and then that's it. It hooked you. And then you celebrate
link |
the fact that this happens to human beings. Yeah. And what is it that made what I did intrinsically
link |
motivating to me? Well, that's interesting and I don't know all the answers to it. And I don't
link |
think I want anybody to think that you should be sort of in any way, I don't know, sanctimonious or
link |
anything about it. You know, it's like, I really enjoyed doing statistical analysis of data. I
link |
really enjoyed running my own experiment, which was what I got a chance to do in the psychology
link |
department that chemistry and physics had never, I never imagined that mere mortals would ever do
link |
an experiment in those sciences, except one that was in the textbook that you were told to do in
link |
lab class. But in psychology, we were already like, even when I was taking psych one, it turned out
link |
we had our own rat and we got to, after two set experiments, we got to, okay, do something you
link |
think of with your rat. So it's the opportunity to do it myself and to bring together a certain
link |
set of things that engaged me intrinsically. And I think it has something to do with why
link |
certain people turn out to be profoundly amazing musical geniuses, right? They get immersed in it
link |
at an early enough point and it just sort of gets into the fabric. So my little brother had intrinsic
link |
motivation for music as we witnessed when he discovered how to put records on the phonograph
link |
when he was like 13 months old and recognize which one he wanted to play, not because he could read
link |
the labels, because he could sort of see which ones had which scratches, which were the different,
link |
you know, oh, that's rapidi espanol. And that's, you know, and, and, and,
link |
And he enjoyed that, that connected with him somehow.
link |
Yeah. And, and there was something that it fed into and it, you're extremely lucky if you have
link |
that and if you can nurture it and can let it grow and let it be, be an important part of your life.
link |
Yeah. Those are, those are the two things is like, be attentive enough to,
link |
to feel it when it comes, like this is something special. I mean, I don't know. For example,
link |
I really like tabular data, like Excel sheets. Like it brings me a deep joy. I don't know how
link |
useful that is for anything. That's part of what I'm talking about.
link |
Exactly. So there's like a million, not a million, but there's a lot of things
link |
like that. For me, you have to hear that for yourself, like be, like realize this is really
link |
joyful. But then the other part that you're mentioning, which is the nurture is take time
link |
and stay with it, stay with it a while and see where that takes you in life.
link |
Yeah. And I think, I think the, the, the motivational engagement results in the
link |
immersion that then creates the opportunity to obtain the expertise. So, you know, we could call
link |
it the Mozart effect, right? I mean, when I think about Mozart, I think about, you know,
link |
the person who was born as the fourth member of the family string quartet, right? And, and they
link |
handed him the violin when he was six weeks old. All right, start playing, you know, it's like,
link |
and so the, the level of immersion there was, was amazingly profound, but hopefully he also had,
link |
you know, some, something, maybe this is where the more sort of the genetic part comes in.
link |
Sometimes I think, you know, something in him resonated to the music so that that,
link |
the synergy of the combination of that was so powerful. So, so that's what I really considered
link |
to be the Mozart effect. It's sort of the, the synergy of something with, with experience that,
link |
that then results in the unique flowering of a particular, you know, mind.
link |
And so I know my siblings and I are all very different from each other. We've all gone in
link |
our own different directions. And, you know, I mentioned my younger brother who was very musical.
link |
I had my other younger brother was like this amazing, like intuitive engineer.
link |
And my sister, one of my sisters was passionate about, in, you know, water conservation well
link |
before it was, you know, such a hugely important issue that it is today. So we all sort of somehow
link |
these find a different thing. And I don't, I don't mean to say it isn't tied in with something about,
link |
about us biologically, but, but it's also when that happens, where you can find that, then,
link |
you know, you can do your thing and you can be excited about it. So people can be excited about
link |
fitting people on bicycles, as well as excited about making neural networks, achieve insights
link |
into human cognition, right? Yeah. Like for me personally, I've always been excited about
link |
love and friendship between humans. And just like the actual experience of it,
link |
since I was a child, just observing people around me and also been excited about robots.
link |
And there's something in me that thinks I really would love to explore how those two things
link |
combine. And it doesn't make any sense. A lot of it is also timing, just to think of your own career
link |
and your own life. You found yourself in certain pieces, places that happened to involve some of
link |
the greatest thinkers of our time. And so it just worked out that like, you guys developed those
link |
ideas. And there may be a lot of other people similar to you, and they were brilliant, and
link |
they never found that right connection and place to where they, their ideas could flourish. So
link |
it's timing, it's place, it's people. And ultimately the whole ride, you know, it's undirected.
link |
Can I ask you about something you mentioned in terms of psychiatry when you were younger?
link |
Because I had a similar experience of, you know, reading Freud and Carl Jung and just,
link |
you know, those kind of popular psychiatry ideas. And that was a dream for me early on in high
link |
school too. Like I hoped to understand the human mind by, somehow psychiatry felt like
link |
the right discipline for that. Does that make you sad? That psychiatry is not
link |
the mechanism by which you are able to explore the human mind. So for me, I was a little bit
link |
disillusioned because of how much prescription medication and biochemistry is involved in the
link |
discipline of psychiatry, as opposed to the dream of the Freud like, use the mechanisms of language
link |
to explore the human mind. So that was a little disappointing. And that's why I kind of went to
link |
computer science and thinking like, maybe you can explore the human mind by trying to build the
link |
thing. Yes. I wasn't exposed to the sort of the biomedical slash pharmacological aspects of
link |
psychiatry at that point because I dropped out of that whole idea of premed that I never even
link |
found out about that until much later. But you're absolutely right. So I was actually a member of the
link |
National Advisory Mental Health Council. That is to say the board of scientists who advise the
link |
director of the National Institute of Mental Health. And that was around the year 2000. And
link |
in fact, at that time, the man who came in as the new director, I had been on this board for a year
link |
when he came in, said, okay, schizophrenia is a biological illness. It's a lot like cancer.
link |
We've made huge strides in curing cancer. And that's what we're going to do with schizophrenia.
link |
We're going to find the medications that are going to cure this disease. And we're not going
link |
to listen to anybody's grandmother anymore. And good old behavioral psychology is not something
link |
we're going to support any further. And he completely alienated me from the Institute
link |
and from all of its prior policies, which had been much more holistic, I think, really at some level.
link |
And the other people on the board were like psychiatrists, very biological psychiatrists.
link |
It didn't pan out that nothing has changed in our ability to help people with mental illness.
link |
And so 20 years later, that particular path was a dead end, as far as I can tell.
link |
Well, there's some aspect to, and sorry to romanticize the whole philosophical conversation
link |
about the human mind. But to me, psychiatrists, for a time, held the flag of we're the deep thinkers.
link |
In the same way that physicists are the deep thinkers about the nature of reality,
link |
psychiatrists are the deep thinkers about the nature of the human mind. And I think that flag
link |
has been taken from them and carried by people like you. It's like, it's more in the cognitive
link |
psychology, especially when you have a foot in the computational view of the world, because you can
link |
both build it, you can like, intuit about the functioning of the mind by building little models
link |
and be able to see mathematical things and then deploying those models, especially in computers,
link |
to say, does this actually work? They do like experiments. And then some combination of
link |
neuroscience, where you're starting to actually be able to observe, do certain experiments on
link |
human beings and observe how the brain is actually functioning. And there, using intuition, you can
link |
start being the philosopher. Like Richard Feynman is the philosopher, cognitive psychologists can
link |
become the philosopher, and psychiatrists become much more like doctors. They're like very medical.
link |
They help people with medication, biochemistry, and so on. But they are no longer the book writers
link |
and the philosophers, which of course I admire. I admire the Richard Feynman ability to do
link |
great low level mathematics and physics and the high level philosophy.
link |
Yeah, I think it was Fromm and Jung more than Freud that was sort of initially kind of like
link |
made me feel like, oh, this is really amazing and interesting and I want to explore it further.
link |
I actually, when I got to college and I lost that thread, I found more of it in sociology
link |
and literature than I did in any place else. So I took quite a lot of both of those
link |
disciplines as an undergraduate. And I was actually deeply ambivalent about
link |
the psychology because I was doing experiments after the initial flurry of interest in
link |
why people would occupy buildings during an insurrection and consider
link |
being so overcommitted to their beliefs. But I ended up in the psychology laboratory running
link |
experiments on pigeons. And so I had these profound dissonance between the kinds of issues
link |
that would be explored when I was thinking about what I read about in modern British literature
link |
versus what I could study with my pigeons in the laboratory. That got resolved when I went
link |
to graduate school and I discovered cognitive psychology. And so for me, that was the path
link |
out of this sort of like extremely sort of ambivalent divergence between the interest
link |
in the human condition and the desire to do actual mechanistically oriented thinking about it. And I
link |
think we've come a long way in that regard and that you're absolutely right that nowadays this
link |
is something that's accessible to people through the pathway in through computer science or the
link |
pathway in through neuroscience. You can get derailed in neuroscience down to the bottom of
link |
the system where you might find the cures of various conditions, but you don't get a chance
link |
to think about the higher level stuff. So it's in the systems and cognitive neuroscience and
link |
computational intelligence, miasma up there at the top that I think these opportunities are most
link |
are richest right now. And so yes, I am indeed blessed by having had the opportunity to fall
link |
into that space. So you mentioned the human condition, speaking which you happen to be a
link |
human being who's unfortunately not immortal. That seems to be a fundamental part of the human
link |
condition that this ride ends. Do you think about the fact that you're going to die one day? Are you
link |
afraid of death? I would say that I am not as much afraid of death as I am of degeneration. And
link |
I say that in part for reasons of having, you know, seen some tragic degenerative situations
link |
unfold. It's exciting when you can continue to participate and feel like you're near the place
link |
where the wave is breaking on the shore, if you like. And I think about my own future potential.
link |
If I were to begin to suffer from Alzheimer's disease or semantic dementia or some other
link |
condition, you know, I would sort of gradually lose the thread of that ability. And so one can
link |
live on for a decade after, you know, sort of having to retire because one no longer has
link |
these kinds of abilities to engage. And I think that's the thing that I fear the most.
link |
SL. The losing of that, like the breaking of the wave, the flourishing of the mind,
link |
where you have these ideas and they're swimming around and you're able to play with them.
link |
RL. Yeah. And collaborate with other people who, you know, are themselves
link |
really helping to push these ideas forward. So, yeah.
link |
SL. What about the edge of the cliff? The end? I mean, the mystery of it. I mean...
link |
RL. The migrated sort of conception of mind and, you know, sort of continuous sort of way of
link |
thinking about most things makes it so that, to me, the discreteness of that transition is less
link |
apparent than it seems to be to most people.
link |
SL. I see. I see. Yeah. Yeah. I wonder, so I don't know if you know the work of Ernest Becker
link |
and so on. I wonder what role mortality and our ability to be cognizant of it
link |
and anticipate it and perhaps be afraid of it, what role that plays in our reasoning of the world.
link |
RL. I think that it can be motivating to people to think they have a limited period left.
link |
SL. I think in my own case, you know, it's like seven or eight years ago now that I was
link |
sitting around doing experiments on decision making that were
link |
satisfying in a certain way because I could really get closure on whether the model fit the data
link |
perfectly or not. And I could see how one could test, you know, the predictions in monkeys as well
link |
as humans and really see what the neurons were doing. But I just realized, hey, wait a minute,
link |
you know, I may only have about 10 or 15 years left here. And I don't feel like I'm getting
link |
towards the answers to the really interesting questions while I'm doing this particular level
link |
of work. And that's when I said to myself, okay, let's pick something that's hard. So that's when
link |
I started working on mathematical cognition. And I think it was more in terms of, well,
link |
I got 15 more years possibly of useful life left. Let's imagine that it's only 10.
link |
I'm actually getting close to the end of that now, maybe three or four more years.
link |
But I'm beginning to feel like, well, I probably have another five after that. So, okay, I'll give
link |
myself another six or eight. But a deadline is looming and therefore. It's not going to go on
link |
forever. And so, yeah, I got to keep thinking about the questions that I think are the interesting and
link |
important ones for sure. What do you hope your legacy is? You've done some incredible work in
link |
your life as a man, as a scientist, when the aliens and the human civilization is long gone
link |
and the aliens are reading the encyclopedia about the human species. What do you hope is the
link |
paragraph written about you? I would want it to sort of highlight
link |
a couple things that I was able to see one path that was more exciting to me than the one that
link |
seemed already to be there for a cognitive psychologist, but not for any super special
link |
reason other than that I'd had the right context prior to that, but that I had gone ahead and
link |
followed that lead. And then I forget the exact wording, but I said in this preface that
link |
the joy of science is the moment in which a partially formed thought in the mind of one person
link |
gets crystallized a little better in the discourse and becomes the foundation
link |
of some exciting concrete piece of actual scientific progress. And I feel like that
link |
moment happened when Rumelhart and I were doing the interactive activation model and when
link |
Rumelhart heard Hinton talk about gradient descent and having the objective function to guide the
link |
learning process. And it happened a lot in that period and I sort of seek that kind of
link |
thing in my collaborations with my students. So the idea that this is a person who contributed
link |
to science by finding exciting collaborative opportunities to engage with other people
link |
through is something that I certainly hope is part of the paragraph.
link |
And like you said, taking a step maybe in directions that are non obvious. So it's the
link |
old Robert Frost road less taken. So maybe because you said like this incomplete initial idea,
link |
that step you take is a little bit off the beaten path.
link |
If I could just say one more thing here. This was something that really contributed
link |
to energizing me in a way that I feel it would be useful to share. My PhD dissertation project
link |
was completely empirical experimental project. And I wrote a paper based on the two main
link |
experiments that were the core of my dissertation and I submitted it to a journal. And at the end
link |
of the paper, I had a little section where I laid out the beginnings of my theory about what I
link |
thought was going on that would explain the data that I had collected. And I had submitted the
link |
paper to the Journal of Experimental Psychology. So I got back a letter from the editor saying,
link |
thank you very much. These are great experiments and we'd love to publish them in the journal.
link |
But what we'd like you to do is to leave the theorizing to the theorists and take that part
link |
out of the paper. And so I did, I took that part out of the paper. But I almost found myself labeled
link |
as a non theorist by this. And I could have succumbed to that and said, okay, well, I guess
link |
my job is to just go on and do experiments, right? But that's not what I wanted to do. And so when I
link |
got to my assistant professorship, although I continued to do experiments because I knew I had
link |
to get some papers out, I also at the end of my first year submitted my first article to
link |
Psychological Review, which was the theoretical journal where I took that section and elaborated
link |
it and wrote it up and submitted it to them. And they didn't accept that either, but they said,
link |
oh, this is interesting. You should keep thinking about it this time. And then that was what got me
link |
going to think, okay, you know, so it's not a superhuman thing to contribute to the development
link |
of theory. You know, you don't have to be, you can do it as a mere mortal.
link |
LB And the broader, I think, lesson is don't succumb to the labels of a particular reviewer.
link |
RL Yeah, that's for sure. Or anybody labeling you, right?
link |
LB Yeah, exactly. I mean that, yeah, exactly. And especially as you become successful,
link |
your labels get assigned to you for that you're successful for that thing.
link |
RL Connectionist or cognitive scientist and not a neuroscientist.
link |
LB And then you can, you can completely, that's just, that's the stories of the past. You're
link |
today a new person that can completely revolutionize in totally new areas. So don't
link |
let those labels hold you back. Well, let me ask the big question. When you look at into the,
link |
you said it started with Columbia trying to observe these humans and they're doing
link |
weird stuff and you want to know why are they doing this stuff. So Zuma even bigger.
link |
LB At the hundred plus billion people who've ever lived on earth. Why do you think we're all
link |
doing what we're doing? What do you think is the meaning of it all? The big why question.
link |
We seem to be very busy doing a bunch of stuff and we seem to be kind of directed towards somewhere.
link |
RL Well, I myself think that we make meaning for ourselves and that we find inspiration
link |
in the meaning that other people have made in the past. You know, and the great religious thinkers
link |
of the first millennium BC and, you know, few that came in the early part of the second millennium,
link |
you know, laid down some important foundations for us.
link |
But I do believe that, you know, we are an emergent result of a process that happened
link |
naturally without guidance and that meaning is what we make of it and that the creation of
link |
efforts to reify meaning in like religious traditions and so on is just a part of the
link |
expression of that goal that we have to, you know, not find out what the meaning is, but to
link |
make it ourselves. And so, to me, it's something that's very personal. It's very individual. It's
link |
like meaning will come for you through the particular combination of synergistic elements
link |
that are your fabric and your experience and your context and, you know, you should...
link |
It's all made in a certain kind of a local context though, right? Here I am at UCSD with this brilliant
link |
man, Rommelhart, who's having, you know, these doubts about symbolic artificial intelligence
link |
that resonate with my desire to see it grounded in the biology and let's make the most of that,
link |
you know? Yeah. And so, from that like little pocket, there's some kind of peculiar little
link |
emergent process that then, which is basically each one of us, each one of us humans is a kind of,
link |
you know, you think cells and they come together and it's an emergent process that then tells fancy
link |
stories about itself and then gets, just like you said, just enjoys the beauty of the stories
link |
we tell about ourselves. It's an emergent process that lives for a time, is defined by its local
link |
pocket and context in time and space and then tells pretty stories and we write those stories
link |
down and then we celebrate how nice the stories are and then it continues because we build stories
link |
on top of each other and eventually we'll colonize hopefully other planets, other solar systems,
link |
other galaxies and we'll tell even better stories. But it all starts here on Earth. Jay, you're
link |
speaking of peculiar emergent processes that lived one heck of a story. You're one of the
link |
the great scientists of cognitive science, of psychology, of computation. It's a huge honor
link |
you would talk to me today that you spend your very valuable time. I really enjoyed talking with
link |
you and thank you for all the work you've done. I can't wait to see what you do next.
link |
JL Well, thank you so much and this has been an amazing opportunity for me to let ideas that I've
link |
never fully expressed before come out because you asked such a wide range of the deeper questions
link |
that we've all been thinking about for so long. So thank you very much for that.
link |
RL Thank you. Thanks for listening to this conversation with Jay McClelland.
link |
To support this podcast, please check out our sponsors in the description.
link |
And now, let me leave you with some words from Jeffrey Hinton. In the long run,
link |
curiosity driven research works best. Real breakthroughs come from people focusing
link |
on what they're excited about. Thanks for listening and hope to see you next time.