The following is a conversation with Matt Potmanick, Director of Neuroscience Research at DeepMind.
He's a brilliant cross-disciplinary mind navigating effortlessly between cognitive
psychology, computational neuroscience, and artificial intelligence.
Quick summary of the ads. Two sponsors. The Jordan Harbinger Show and Magic Spoon Serial.
Please consider supporting the podcast by going to jordanharbinger.com slash lex,
and also going to magicspoon.com slash lex and using code lex at checkout after you buy
all of their serial. Click the links, buy the stuff. It's the best way to support this podcast
and journey I'm on. If you enjoy this podcast, subscribe on YouTube, review it with 5,000
Apple Podcasts, follow on Spotify, support on Patreon, or connect with me on Twitter at Lex
Friedman, spelled surprisingly without the E, just F-R-I-D-M-A-N. As usual, I'll do a few minutes
of ads now and never any ads in the middle that can break the flow of the conversation.
This episode is supported by the Jordan Harbinger Show. Go to jordanharbinger.com slash lex.
It's how he knows I sent you. On that page, subscribe to his podcast,
on Apple Podcasts, Spotify, and you know where to look. I've been binging on his podcast.
Jordan is a great interviewer and even a better human being. I recently listened to his conversation
with Jack Barsky, former sleeper agent for the KGB in the 80s and author of Deep Under Cover,
which is a memoir that paints yet another interesting perspective on the Cold War era.
I've been reading a lot about the Stalin and then Gorbachev and Putin eras of Russia,
but this conversation made me realize that I need to do a deep dive into the Cold War era
to get a complete picture of Russia's recent history. Again, go to jordanharbinger.com slash
lex. Subscribe to his podcast. It's how he knows I sent you. It's awesome. You won't regret it.
This episode is also supported by Magic Spoon, low carb, keto friendly, super amazingly delicious
cereal. I've been on a keto or very low carb diet for a long time now. It helps with my
mental performance. It helps with my physical performance even doing this crazy push up pull
up challenge I'm doing, including the running. It just feels great. I used to love cereal.
Obviously I can't have it now because most cereals have a crazy amount of sugar,
which is terrible for you. So I quit it years ago, but Magic Spoon amazingly somehow is a
totally different thing. Zero sugar, 11 grams of protein and only three net grams of carbs.
It tastes delicious. It has a lot of flavors, two new ones, including peanut butter. But if you
know what's good for you, you'll go with cocoa, my favorite flavor and the flavor of champions.
Click the magicspoon.com slash lex link in the description and use code lex at checkout for free
shipping and to let them know I sent you. They have agreed to sponsor this podcast for a long time.
They're an amazing sponsor and an even better cereal. I highly recommend it. It's delicious.
It's good for you. You won't regret it. And now here's my conversation with Matt
Botkinick. How much of the human brain do you think we understand?
I think we're at a weird moment in the history of neuroscience in the sense that
I feel like we understand a lot about the brain at a very high level,
but a very coarse level.
When you say high level, what are you thinking? Are you thinking functional? Are you thinking
structurally? In other words, what is the brain for? What kinds of computation does the brain do?
What kinds of behaviors would we have to explain if we were going to look down
at the mechanistic level? And at that level, I feel like we understand much, much more about
the brain than we did when I was in high school. But it's almost like we're seeing it through a
fog. It's only at a very coarse level. We don't really understand what the neuronal mechanisms
are that underlie these computations. We've gotten better at saying what are the functions
that the brain is computing that we would have to understand if we were going to get down to
the neuronal level. At the other end of the spectrum, in the last few years, incredible
progress has been made in terms of technologies that allow us to see, actually literally see,
in some cases, what's going on at the single unit level, even the dendritic level,
and then there's this yawning gap in between.
Well, that's interesting. So at the high level, so that's almost a cognitive science level.
And then at the neuronal level, that's neurobiology and neuroscience, just studying single neurons,
the synaptic connections and all the dopamine, all the kind of neurotransmitters.
One blanket statement I should probably make is that as I've gotten older, I have become more
and more reluctant to make a distinction between psychology and neuroscience. To me, the point
of neuroscience is to study what the brain is for. If you're a nephrologist and you want to
learn about the kidney, you start by saying, what is this thing for? Well, it seems to be for
taking blood on one side that has metabolites in it that shouldn't be there, sucking them out of the
blood while leaving the good stuff behind, and then excreting that in the form of urine.
That's what the kidney is for. It's obvious. So the rest of the work is deciding how it does that.
And this, it seems to me, is the right approach to take to the brain. You say, well, what is the
brain for? The brain, as far as I can tell, is for producing behavior. It's for going from
perceptual inputs to behavioral outputs. And the behavioral outputs should be adaptive.
So that's what psychology is about. It's about understanding the structure of that function.
And then the rest of neuroscience is about figuring out how those operations are actually carried
out at a mechanistic level.
That's really interesting. But so unlike the kidney, the brain, the gap between the electrical
signal and behavior, you see, you truly see neuroscience as the science that touches behavior,
how the brain generates behavior, or how the brain converts raw visual information into
understanding. You basically see cognitive science, psychology, and neuroscience as all one science.
Yeah. It's a personal statement. Is that a hopeful or a realistic statement?
So certainly you will be correct in your feeling in some number of years,
but that number of years could be 200, 300 years from now.
Oh, well, well, there's a, is that aspirational or is that pragmatic engineering feeling that you
have? It's both in the sense that this is what I hope and expect will bear fruit over the coming
decades. But it's also pragmatic in the sense that I'm not sure what we're doing in either
psychology or neuroscience, if that's not the framing. I don't know what it means to understand
the brain if part of the enterprise is not about understanding the behavior that's being produced.
I mean, yeah, but I would compare it to maybe astronomers looking at the movement of the
planets and the stars and without any interest of the underlying physics, right? And I would argue
that at least in the early days, there's some value to just tracing the movement of the planets
and the stars without thinking about the physics too much, because it's such a big leap to start
thinking about the physics before you even understand even the basic structural elements of...
Oh, I agree with that. I agree.
But you're saying in the end, the goal should be to deeply understand.
Well, right. And I think... So I thought about this a lot when I was in grad school, because a lot
of what I studied in grad school was psychology. And I found myself a little bit confused about
what it meant to... It seems like what we were talking about a lot of the time were
virtual causal mechanisms. Like, oh, well, you know, attentional selection then selects some
object in the environment, and that is then passed on to the motor... Information about that is passed
on to the motor system. But these are virtual mechanisms. They're metaphors. There's no reduction
to... There's no reduction going on in that conversation to some physical mechanism that...
Which is really what it would take to fully understand how behavior is rising. But the
causal mechanisms are definitely neurons interacting. I'm willing to say that at this
point in history. So in psychology, at least for me personally, there was this strange insecurity
about trafficking in these metaphors, which we're supposed to explain the function of the mind.
If you can't ground them in physical mechanisms, then what is the explanatory validity of these
explanations? And I managed to soothe my own nerves by thinking about the history of
genetics research. So I'm very far from being an expert on the history of this field. But I know
enough to say that Mendelian genetics preceded Watson and Crick. And so there was a significant
period of time during which people were productively investigating the structure of inheritance
using what was essentially a metaphor, the notion of a gene. And genes do this and genes do that.
But we're the genes. They're sort of an explanatory thing that we made up. And we ascribed to them
these causal properties. So there's a dominant, there's a recessive, and then they recombine it.
And then later, there was a kind of blank there that was filled in with a physical mechanism.
That connection was made. But it was worth having that metaphor, because that gave us
a good sense of what kind of causal mechanism we were looking for.
Right. And the fundamental metaphor of cognition, you said, is the interaction of neurons.
Is that what is the metaphor? No, no, the metaphor, the metaphors we use in cognitive psychology are
things like attention, the way that memory works. I retrieve something from memory.
A memory retrieval occurs. What is that? That's not a physical mechanism that I can examine in
its own right. But it's still worth having that metaphorical level.
Yeah, I misunderstood, actually. So the higher level abstractions is the metaphor that's most
useful. But what about, how does that connect to the idea that that arises from interaction of
neurons? Is the interaction of neurons also not a metaphor to you? Or is it literally,
that's no longer a metaphor. That's already the lowest level of abstractions that could actually
be directly studied. Well, I'm hesitating because I think what I want to say could end up being
controversial. So what I want to say is, yes, the interactions of neurons, that's not
metaphorical. That's a physical fact. That's where the causal interactions actually occur.
Now, I suppose you couldn't say, well, even that is metaphorical relative to the quantum
events that underline. I don't want to go down that rabbit hole.
It's always turtles on top of turtles. It's all the way down.
There is a reduction that you can do. You can say these psychological phenomena
are can be explained through a very different kind of causal mechanism, which has to do with
neurotransmitter release. And so what we're really trying to do in neuroscience, large,
as I say, which for me includes psychology, is to take these psychological phenomena
and map them onto neural events. I think remaining forever at the level of
description that is natural for psychology, for me personally, would be disappointing.
I want to understand how mental activity arises from neural activity. But the converse is also true.
Studying neural activity without any sense of what you're trying to explain,
to me, feels like at best roping around at random.
Now, you've talked about this bridging of the gap between psychology and neuroscience,
but do you think it's possible? I fell in love with psychology and psychiatry in general,
with Freud and when I was really young, and I hope to understand the mind. And for me,
understanding the mind at least at that young age before discovered AI and even neuroscience was to
is psychology. And do you think it's possible to understand the mind without getting into all
the messy details of neuroscience? Like you kind of mentioned, to you, it's appealing to try to
understand the mechanisms at the lowest level. But do you think that's needed? That's required
to understand how the mind works? That's an important part of the whole picture.
But I would be the last person on earth to suggest that that reality renders psychology in its own
right unproductive. I trained as a psychologist. I am fond of saying that I have learned much more
from psychology than I have from neuroscience. To me, psychology is a hugely important discipline.
And one thing that warms in my heart is that
ways of investigating behavior that have been native to cognitive psychology since it's
dawn in the 60s are starting to become interesting to AI researchers for a variety of reasons.
And that's been exciting for me to see.
Can you maybe talk a little bit about what you see as beautiful aspects of psychology,
maybe limiting aspects of psychology? I mean, maybe just started off as a science, as a field.
To me, it was when I understood what psychology is, analytical psychology, like the way it's
actually carried out, it's really disappointing to see two aspects. One is how small the N is,
how small the number of subjects is in the studies. And two, it was disappointing to see
how controlled the entire, how much it was in the lab. It wasn't studying humans in the wild.
There was no mechanism for studying humans in the wild. So that's where I became a little bit
disillusioned to psychology. And then the modern world of the internet is so exciting to me,
the Twitter data or YouTube data, like data of human behavior on the internet becomes exciting
because the N grows and then in the wild grows. But that's just my narrow sense. Do you have
a optimistic or pessimistic cynical view of psychology? How do you see the field broadly?
When I was in graduate school, it was early enough that there was still a thrill in seeing
that there were ways of doing experimental science that provided insight to the structure of the
mind. One thing that impressed me most when I was at that stage in my education was
neuropsychology, analyzing the behavior of populations who had brain damage of different
kinds and trying to understand what the specific deficits were that arose from
a lesion in a particular part of the brain and the kind of experimentation that was done and
that's still being done to get answers in that context was so creative and it was so deliberate.
It was good science. An experiment answered one question but raised another and somebody
would do an experiment that answered that question and you really felt like you were narrowing in on
some kind of approximate understanding of what this part of the brain was for.
Do you have an example from memory of what kind of aspects of the mind could be studied in this
kind of way? Oh, sure. The very detailed neuropsychological studies of language
function, looking at production and reception and the relationship between visual function,
you know, reading and auditory and semantic and still are these beautiful models that came
out of that kind of research that really made you feel like you understood something that you
hadn't understood before about how language processing is organized in the brain. But having
said all that, I agree with you that the cost of doing highly controlled experiments
is that you, by construction, miss out on the richness and complexity of the real world.
One thing that, so I was drawn into science by what in those days was called connectionism,
which is of course what we now call deep learning. And at that point in history,
neural networks were primarily being used in order to model human cognition.
They weren't yet really useful for industrial applications.
So you always found neural networks in biological form beautiful?
Oh, neural networks were very concretely the thing that drew me into science.
I was handed, are you familiar with the PDP books from the 80s?
I went to medical school before I went into science.
Really? Interesting. Wow.
I also did a graduate degree in art history, so I'm kind of exploring.
Well, art history, I understand. That's just a curious, creative mind. But medical school,
with the dream of what if we take that slight tangent,
what did you want to be a surgeon?
I actually was quite interested in surgery. I was interested in surgery and psychiatry,
and I thought that I must be the only person on the planet who was torn between those two fields.
And I said exactly that to my advisor in medical school, who turned out,
I found out later, to be a famous psychoanalyst. And he said to me, no, no,
it's actually not so uncommon to be interested in surgery and psychiatry.
And he conjectured that the reason that people develop these two interests is that
both fields are about going beneath the surface and kind of getting into the kind of secret.
I mean, maybe you understand this as someone who was interested in psychoanalysis.
There's a cliche phrase that people use now on NPR, the secret life of blanketed blank.
Right? And that was part of the thrill of surgery was seeing the secret activity
that's inside everybody's abdomen and thorax.
That's a very poetic way to connect it to disciplines that are very
practically speaking different from each other.
That's for sure.
That's for sure.
Yes.
So how do we get on to medical school?
So I was in medical school, and I was doing a psychiatry rotation,
and my kind of advisor in that rotation asked me what I was interested in.
And I said, well, maybe psychiatry.
He said, why?
And I said, well, I've always been interested in how the brain works.
I'm pretty sure that nobody's doing scientific research that addresses my interests,
which I didn't have a word for it then, but I would have said about cognition.
And he said, well, I'm not sure that's true.
You might be interested in these books.
And he pulled down the PDB books from his shelf, and they were still shrink wrapped.
He hadn't read them, but he handed them to me.
He said, you feel free to borrow these.
And that was, I went back to my dorm room and I just read them cover to cover.
And what's PDB?
Parallel distributed processing, which was one of the original names for deep learning.
And so, I apologize for the romanticized question, but what idea in the space of neural
size, in the space of the human brain, is to you the most beautiful, mysterious, surprising?
What had always fascinated me, even when I was a pretty young kid, I think,
was the paradox that lies in the fact that the brain is so mysterious and so seems so distant.
But at the same time, it's responsible for the full transparency of everyday life.
The brain is literally what makes everything obvious and familiar.
And there's always one in the room with you.
When I taught at Princeton, I used to teach a cognitive neuroscience course.
And the very last thing I would say to the students was,
when people think of scientific inspiration, the metaphor is often, well, look to the stars.
The stars will inspire you to wonder at the universe and think about your place in it
and how things work.
And I'm all for looking at the stars.
But I've always been much more inspired and my sense of wonder comes from the, not from the
distant, mysterious stars, but from the extremely, intimately close brain.
There's something just endlessly fascinating to me about that.
Like Jessica said, the one is close and yet distant in terms of our understanding of it.
Do you, are you also captivated by the fact that this very conversation is happening because
two brains are communicating?
Yes, exactly.
I guess what I mean is the subjective nature of the experience.
If we can take a small tangent into the mystical of it, the consciousness, or when you're saying
you're captivated by the idea of the brain, are you talking about specifically the mechanism
of cognition? Or are you also just, like, at least for me, it's almost like paralyzing
the beauty and the mystery of the fact that it creates the entirety of the experience,
not just the reasoning capability, but the experience?
Well, I definitely resonate with that latter thought.
And I often find discussions of artificial intelligence to be disappointingly narrow.
Speaking as someone who has always had an interest in art.
Right, I was just going to go there because it sounds like somebody who has an interest in art.
Yeah, I mean, there are many layers to full bore human experience.
And in some ways, it's not enough to say, oh, well, don't worry, we're talking about cognition,
but we'll add emotion.
There's an incredible scope to what humans go through in every moment.
And yes, so that's part of what fascinates me is that our brains are producing that.
But at the same time, it's so mysterious to us, how?
We literally, our brains are literally in our heads producing this experience.
Producing the experience.
And yet it's so mysterious to us, and the scientific challenge of getting at the
actual explanation for that is so overwhelming.
Certain people have fixations on particular questions, and that's always just always been mine.
Yeah, I would say the poetry that is fascinating.
And I'm really interested in natural language as well.
And when you look at artificial intelligence community, it always saddens me how much when
you try to create a benchmark for the community together around how much of the magic of language
is lost when you create that benchmark, that there's something we talk about experience,
the music of the language, the wit, something that makes a rich experience,
something that would be required to pass the spirit of the Turing test is lost in these
benchmarks. And I wonder how to get it back in because it's very difficult.
The moment you try to do real good rigorous science, you lose some of that magic.
When you try to study cognition in a rigorous scientific way, it feels like you're losing
some of the magic, the seeing cognition in a mechanistic way that AI folk at this stage
in our history look at. I agree with you. But at the same time, one thing that I found
really exciting about that first wave of deep learning models in cognition was
the fact that the people who were building these models were focused on
on the richness and complexity of human cognition. So an early debate in cognitive science,
which I sort of witnessed as a grad student, was about something that sounds very dry,
which is the formation of the past tense. But there were these two camps. One said,
well, the mind encodes certain rules. And it also has a list of exceptions,
because of course, the rule is add ED, but that's not always what you do. So you have to have a
list of exceptions. And then there were the connectionists who evolved into the deep learning
people who said, well, if you look carefully at the data, if you look at actually look at
corpora, like language corpora, it turns out to be very rich because, yes, there are most
there are most verbs that and, you know, you just tack on ED. And then there are exceptions,
but there are also there's also there are there are rules that in there's the exceptions aren't
just random. There are certain clues to which which which verbs should be exceptional. And then
there are exceptions to the exceptions. And there was a word that was kind of deployed in order
to capture this, which was quasi regular. In other words, there are rules, but it's it's messy.
And there there's there's structure, even among the exceptions. And it would be, yeah, you could
try to write down, you could try to write down the structure in some sort of closed form. But
really, the right way to understand how the brain is handling all this, and by the way, producing
all of this is to build a deep neural network and train it on this data and see how it ends up
representing all of this richness. So the way that deep learning was deployed in cognitive
psychology was that was the spirit of it. It was about that richness. And that's something that I
always found very, very compelling, still do. Is it is there something, especially interesting
and profound to you, in terms of our current deep learning neural network, artificial neural
network approaches, and the whatever we do understand about the biological neural networks
in our brain, is there there's some there's quite a few differences. Are some of them to you
either interesting or perhaps profound in terms of in terms of the gap we might want to try to
close in trying to create a human level intelligence? What I would say here is something that a lot of
people are saying, which is that one seeming limitation of the systems that we're building now
is that they lack the kind of flexibility, the readiness to turn on a dime when the context
calls for it. That is so characteristic of human behavior. So is that connected to you to the
like which aspect of the neural networks in our brain is that connected to? Is that closer to the
cognitive science level of now again, see like my natural inclination is to separate into three
disciplines of neuroscience, cognitive science and psychology. And you've already kind of shut
that down by saying you're kind of see them as separate. But just to look at those layers, I
guess where is there something about the lowest layer of the way the neurons interact that is
profound to you in terms of its difference to the artificial neural networks? Or is all the
key differences at a higher level of abstraction? One thing I often think about is that if you
take an introductory computer science course and they are introducing you to the notion of
Turing machines, one way of articulating what the significance of a Turing machine is,
is that it's a machine emulator. It can emulate any other machine. And that way of looking at
that a Turing machine really sticks with me. I think of humans as maybe sharing in some of that
character. We're capacity limited. We're not Turing machines, obviously. But we have the
ability to adapt behaviors that are very much unlike anything we've done before. But there's
some basic mechanism that's implemented in our brain that allows us to run software.
But just on that point, you mentioned a Turing machine. But nevertheless, it's fundamentally
our brains are just computational devices in your view. Is that what you're getting at?
It was a little bit unclear to this line you drew. Is there any magic in there? Or is it just basic
computation? I'm happy to think of it as just basic computation. But mind you, I won't be
satisfied until somebody explains to me what the basic computations are that are leading to
the full richness of human cognition. It's not going to be enough for me to understand what the
computations are that allow people to do arithmetic or play chess. I want the whole thing.
And a small tangent because you kind of mentioned coronavirus. There's group behavior.
Oh, sure. Is there something interesting to your search of understanding the human mind
where behavior of large groups or just behavior of groups is interesting? Seeing that as a
collective mind, as a collective intelligence, perhaps seeing the groups of people as a single
intelligent organism, especially looking at the reinforcement learning work you've done recently?
Well, yeah, I have the honor of working with a lot of incredibly smart people. And I won't
want to take any credit for leading the way on the multi-agent work that's come out of my group
or deep mind lately. But I do find it fascinating. And I think it can't be debated. Human behavior
arises within communities. That just seems to me self-evident.
But to me, it is self-evident, but that seems to be a profound aspect of something that created.
And that was like, if you look at 2001 Space Odyssey, when the monkeys touched the...
Like, that's the magical moment, I think Yovar Harari argues that the ability of our
large numbers of humans to hold an idea, to converge towards idea together, like he said,
shaking hands versus bumping elbows, somehow converge without being in a room altogether,
just this distributed convergence towards an idea over a particular period of time,
seems to be fundamental to just every aspect of our cognition, of our intelligence,
because humans will talk about reward, but it seems like we don't really have a clear objective
function under which we operate. But we all kind of converge towards one somehow. And that to me
has always been a mystery that I think is somehow productive for also understanding AI systems.
But I guess that's the next step. The first step is try to understand the mind.
Well, I don't know. I mean, I think there's something to the argument that
that kind of bottom... Strictly bottom-up approach is wrongheaded. In other words,
there are basic aspects of human intelligence that can only be understood in the context of
groups. I'm perfectly open to that. I've never been particularly
convinced by the notion that we should consider intelligence to adhere at the level of communities.
I don't know why. I'm sort of stuck on the notion that the basic unit that we want to
understand is individual humans. And if we have to understand that in the context of other humans,
fine. But for me, intelligence is just... I stubbornly define it as something that is an
aspect of an individual human. That's just my... I'm with you, but that could be the reductionist
dream of a scientist because you can understand a single human. It also is very possible that
intelligence can only arise when there's multiple intelligences. When there's multiple...
It's a sad thing if that's true because it's very difficult to study. But if it's just one human,
that one human would not be... Homo sapiens would not become that intelligent. That's a
possibility. One thing I will say along these lines is that I think a serious effort to understand
human intelligence and maybe to build human-like intelligence needs to pay just as much attention
to the structure of the environment as to the structure of the cognizing system,
whether it's a brain or an AI system. That's one thing I took away actually from my early
studies with the pioneers of neural network research, people like Jay McClelland and John Cohen.
The structure of cognition is really... It's only partly a function of the
architecture of the brain and the learning algorithms that it implements. What really
shapes it is the interaction of those things with the structure of the world in which those things
are embedded. That's especially important for... That's made most clear in reinforcement learning
where the simulated environment is... You can only learn as much as you can simulate. That's what
DeepMind made very clear with the other aspect of the environment, which is the self-play mechanism
of the other agent of the competitive behavior, which the other agent becomes the environment,
essentially. That's one of the most exciting ideas in AI is the self-play mechanism that's able to
learn successfully. There you go. There's a thing where competition is essential for learning,
at least in that context. If we can step back into another beautiful world, which is the actual
mechanics, the dirty mess of it of the human brain, is there something for people who might
not know? Is there something you can comment on or describe the key parts of the brain that are
important for intelligence or just in general, what are the different parts of the brain that
you're curious about that you've studied and that are just good to know about when you're
thinking about cognition? Well, my area of expertise, if I have one, is prefrontal cortex.
What's that? It depends on who you ask. The technical definition is anatomical. There are
parts of your brain that are responsible for motor behavior, and they're very easy to identify.
The region of your cerebral cortex, the outer crust of your brain, that lies in front of those
is defined as the prefrontal cortex. When you say anatomical, sorry to interrupt,
so that's referring to the geographic region as opposed to some kind of functional definition?
Exactly. This is the coward's way out. I'm telling you what the prefrontal cortex is,
just in terms of what part of the real estate it occupies.
It's the thing in the front of the brain.
Yeah, exactly. In fact, the early history of the neuroscientific investigation of what this
front part of the brain does is funny to read. It was really World War I that started people
down this road of trying to figure out what different parts of the human brain do in the
sense that there were a lot of people with brain damage who came back from the war with
brain damage. That provided, as tragic as that was, an opportunity for scientists to try to
identify the functions of different brain regions. That was actually incredibly productive,
but one of the frustrations that neuropsychologists faced was they couldn't really identify exactly
what the deficit was that arose from damage to these most frontal parts of the brain.
It was just a very difficult thing to pin down. There were a couple of neuropsychologists who
identified, through a large amount of clinical experience and close observation, they started to
put their finger on a syndrome that was associated with frontal damage. Actually,
one of them was a Russian neuropsychologist named Luria, who students of cognitive psychology
still read. What he started to figure out was that the frontal cortex was somehow involved in
flexibility, in guiding behaviors that required someone to override a habit,
or to do something unusual, or to change what they were doing in every flexible way from one
moment to another. Focused on new experiences. The way your brain processes and acts in new
experiences. Yeah. What later helped bring this function into better focus was a distinction between
controlled and automatic behavior. In other literatures, this is referred to as habitual
behavior versus goal-directed behavior. It's very clear that the human brain has pathways that are
dedicated to habits, to things that you do all the time. They need to be automatized so that
they don't require you to concentrate too much. That leaves your cognitive capacity free to do
other things. Just think about the difference between driving when you're learning to drive
versus driving after you're a fairly expert. There are brain pathways that slowly absorb
those frequently performed behaviors so that they can be habits, so that they can be automatic.
That's the purest form of learning. I guess it's happening there. This is jumping ahead,
which is why that perhaps is the most useful for us to focus on and try to see how
artificial intelligence systems can learn. Is that the way you think?
It's interesting. I do think about this distinction between controlled and automatic,
or goal-directed and habitual behavior a lot in thinking about where we are in AI research.
Just to finish the dissertation here, the role of the prefrontal cortex is generally
understood these days in contradistinction to that habitual domain. In other words,
the prefrontal cortex is what helps you override those habits. It's what allows you to say,
well, what I usually do in this situation is X, but given the context, I probably should do Y.
The elbow bump is a great example. Reaching out and shaking hands is probably habitual behavior,
and it's the prefrontal cortex that allows us to bear in mind that there's something unusual
going on right now, and in this situation, I need to not do the usual thing. The kind of behaviors
that Luria reported, and he built tests for detecting these kinds of things, were exactly
like this. In other words, when I stick out my hand, I want you instead to present your elbow.
A patient with frontal damage would have a great deal of trouble with that.
Somebody proffering their hand would elicit a handshake. The prefrontal cortex is what
allows us to say, hold on, that's the usual thing, but I have the ability to bear in mind
even very unusual contexts and to reason about what behavior is appropriate there.
Just to get a sense, are us humans special in the presence of the prefrontal cortex?
Do mice have a prefrontal cortex? Do other mammals that we can study? If no,
then how do they integrate new experiences? That's a really tricky question and a very
timely question because we have revolutionary new technologies for monitoring, measuring,
and also causally influencing neural behavior in mice and fruit flies.
These techniques are not fully available even for studying brain function in monkeys,
let alone humans. For me, at least, it's a very urgent question whether the kinds of
things that we want to understand about human intelligence can be pursued in these other
organisms. To put it briefly, there's disagreement. People who study fruit flies will often tell you,
hey, fruit flies are smarter than you think. They'll point to experiments where fruit flies were able
to learn new behaviors, were able to generalize from one stimulus to another in a way that suggests
that they have abstractions that guide their generalization. I've had many conversations in
which I will start by recounting some observation about mouse behavior, where it seemed like
mice were taking an awfully long time to learn a task that for a human would be
profoundly trivial. I will conclude from that that mice really don't have the cognitive flexibility
that we want to explain and that a mouse researcher will say to me, well, you know, hold on.
That experiment may not have worked because you asked a mouse to deal with stimuli and behaviors
that were very unnatural for the mouse. If instead, you kept the logic of the experiment
the same, but presented the information in a way that aligns with what mice are used to
dealing with in their natural habitats, you might find that a mouse actually has more intelligence
than you think. Then they'll go on to show you videos of mice doing things in their natural
habitat, which seem strikingly intelligent, dealing with physical problems. I have to drag
this piece of food back to my lair, but there's something in my way, and how do I get rid of
that thing? I think these are open questions to put it to sum that up.
Then taking a small step back related to that, as you mentioned, we're taking a little shortcut by
saying it's a geographic part of the prefrontal cortex is the region of the brain. What's your
sense in a bigger philosophical view, prefrontal cortex and the brain in general? Do you have a
sense that it's a set of subsystems in the way we've kind of implied that they're pretty distinct?
To what degree is it that, or to what degree is it a giant interconnected mess where everything
kind of does everything and it's impossible to disentangle them?
I think there's overwhelming evidence that there's functional differentiation,
that it's clearly not the case that all parts of the brain are doing the same thing.
This follows immediately from the kinds of studies of brain damage that we were chatting
about before. It's obvious from what you see if you stick an electrode in the brain and measure
what's going on at the level of neural activity. Having said that, there are two other things to
add, which kind of maybe tug in the other direction. One is that when you look carefully
at functional differentiation in the brain, what you usually end up concluding, at least this is
my observation of the literature, is that the differences between regions are graded
rather than being discrete. It doesn't seem like it's easy to divide the brain up into
true modules that have clear boundaries and that have clear channels of communication between them.
And this applies to the prefrontal cortex?
Yeah, the prefrontal cortex is made up of a bunch of different subregions,
the functions of which are not clearly defined and the borders of which seem to be quite vague.
Then there's another thing that's popping up in very recent research, which
involves application of these new techniques. There are a number of studies that suggest that
parts of the brain that we would have previously thought were quite focused in their
function are actually carrying signals that we wouldn't have thought would be there. For example,
looking in the primary visual cortex, which is classically thought of as basically the
first cortical waystation for processing visual information. Basically, what it should care about
is where are the edges in this scene that I'm viewing? It turns out that if you have enough data,
you can recover information from primary visual cortex about all sorts of things,
like what behavior the animal is engaged in right now and how much reward is on offer
in the task that it's pursuing. It's clear that even regions whose function is pretty well defined
at a core screen are nonetheless carrying some information about information from very different
domains. The history of neuroscience is this oscillation between the two views that you
articulated, the modular view and then the big mush view. I guess we're going to end up somewhere
in the middle, which is unfortunate for our understanding because there's something about
our conceptual system that finds it's easy to think about a modularized system and easy to
think about a completely undifferentiated system. Something that lies in between is confusing,
but we're going to have to get used to it. Unless we can understand deeply the lower
level mechanism of neuronal communication. On that topic, you mentioned information.
Just to get a sense, I imagine something that there's still mystery and disagreement on
is how does the brain carry information and signal? What in your sense is the basic
mechanism of communication in the brain? Well, I guess I'm old fashioned in that
I consider the networks that we use in deep learning research to be a reasonable approximation
to the mechanisms that carry information in the brain. The usual way of articulating that is to
say, what really matters is a rate code. What matters is how quickly is an individual neuron
spiking? What's the frequency at which it's spiking? Is it the timing of the spiking?
Yeah. Is it firing fast or slow? Let's put a number on that, and that number is enough to
capture what neurons are doing. There's still uncertainty about whether that's an
adequate description of how information is transmitted within the brain. There are
studies that suggest that the precise timing of spikes matters. There are studies that suggest
that there are computations that go on within the dendritic tree, within a neuron, that are
quite rich and structured and that really don't equate to anything that we're doing in our artificial
neural networks. Having said that, I feel like we're getting somewhere by sticking to this
high level of abstraction. By the way, we're talking about the electrical signal. I remember
reading some vague paper somewhere recently where the mechanical signal, like the vibrations or
something of the neurons also communicate information. I haven't seen that. There's
somebody was arguing that the electrical signal, this is in nature paper, something like that,
where the electrical signal is actually a side effect of the mechanical signal. I don't think
they changed the story, but it's almost an interesting idea that there could be a deeper.
It's always in physics with quantum mechanics, there's always a deeper story that could be
underlying the whole thing. You think it's basically the rate of spiking that gets us,
that's the lowest hanging fruit that can get us really far.
This is a classical view. The only way in which this stance would be controversial is
in the sense that there are members of the neuroscience community who are interested
in alternatives, but this is really a very mainstream view. The way that neurons communicate
is that neurotransmitters arrive, they wash up on a neuron. The neuron has receptors for
those transmitters. The meeting of the transmitter with these receptors changes the voltage of the
neuron. If enough voltage change occurs, then a spike occurs, one of these discrete events.
It's that spike that is conducted down the axon and leads to neurotransmitter release.
This is just like neuroscience 101. This is the way the brain is supposed to work.
What we do when we build artificial neural networks of the kind that are now popular in the AI
community is that we don't worry about those individual spikes, we just worry about the
frequency at which those spikes are being generated. People talk about that as the
activity of a neuron. The activity of units in a deep learning system is broadly analogous to
the spike rate of a neuron. There are people who believe that there are other forms of
communication in the brain. In fact, I've been involved in some research recently that suggests
that the voltage fluctuations that occur in populations of neurons that are below the
level of spike production may be important for communication, but I'm still pretty old school
in the sense that I think that the things that we're building in AI research constitute reasonable
models of how a brain would work. Let me ask just for fun a crazy question,
because I can. Do you think it's possible we're completely wrong about the way this basic
mechanism of neuronal communication, that the information is stored in some very different
kind of way in the brain? Heck yes. I wouldn't be a scientist if I didn't think there was any
chance we were wrong, but if you look at the history of deep learning research as it's been
applied to neuroscience, of course, the vast majority of deep learning research these days
isn't about neuroscience, but if you go back to the 1980s, there's an unbroken chain of research
in which a particular strategy is taken, which is, hey, let's train a deep learning system.
Let's train a multi-layered neural network on this task that we trained our rat on or our monkey on
or this human being on. Then let's look at what the units deep in the system are doing.
Let's ask whether what they're doing resembles what we know about what neurons deep in the
brain are doing. Over and over and over and over, that strategy works in the sense that
the learning algorithms that we have access to, which typically center on back propagation,
they give rise to patterns of activity, patterns of response,
patterns of neuronal behavior in these artificial models that look hauntingly similar
to what you see in the brain. Is that a coincidence?
At a certain point, it starts looking like such coincidence is unlikely to not be deeply
meaningful. Yeah, the circumstantial evidence is overwhelmed.
But you're always open to a total flipping of the table. Hey, of course. You have co-authored
several recent papers that weave beautifully between the world of neuroscience and artificial
intelligence. If we could just try to dance around and talk about some of them, maybe try
to pick out interesting ideas that jump to your mind from memory. Maybe looking at,
we were talking about the prefrontal cortex, the 2018, I believe, paper called the prefrontal
cortex is a matter of reinforcement learning system. Is there a key idea that you can speak to
from that paper? Yeah. The key idea is about meta-learning.
What is meta-learning? Meta-learning is, by definition,
a situation in which you have a learning algorithm and the learning algorithm operates
in such a way that it gives rise to another learning algorithm. In the earliest applications
of this idea, you had one learning algorithm sort of adjusting the parameters on another
learning algorithm. But the case that we're interested in this paper is one where you start with
just one learning algorithm and then another learning algorithm kind of emerges out of thin air.
I can say more about what I mean by that. I don't mean to be
scurrent. But that's the idea of meta-learning. It relates to the old idea in psychology of
learning to learn, situations where you have experiences that make you better at learning
something new. A familiar example would be learning a foreign language. The first time
you learn a foreign language, it may be quite laborious and disorienting and novel. But let's
say you've learned two foreign languages, the third foreign language obviously is going to be much
easier to pick up. Why? Because you've learned how to learn. You know how this goes. You know,
okay, I'm going to have to learn how to conjugate. I'm going to have to... That's a simple form of
meta-learning in the sense that there's some slow learning mechanism that's helping you update your
fast learning mechanism. Does that make sense? That makes sense. From our understanding, from the
psychology world, from neuroscience, our understanding how meta-learning works might work
in the human brain, what lessons can we draw from that that we can bring into the artificial
intelligence world? Well, yeah. The origin of that paper was in AI work that we were doing in my
group. We were looking at what happens when you train a recurrent neural network using standard
reinforcement learning algorithms. But you train that network not just in one task, but you train
it in a bunch of interrelated tasks. Then you ask what happens when you give it yet another task in
that line of interrelated tasks. What we started to realize is that a form of meta-learning spontaneously
happens in recurrent neural networks. The simplest way to explain it is to say a recurrent neural
network has a memory in its activation patterns. It's recurrent by definition in the sense that
you have units that connect to other units that connect to other units. You have loops of connectivity,
which allows activity to stick around and be updated over time. In psychology, in neuroscience,
we call this working memory. It's like actively holding something in mind.
That memory gives the recurrent neural network a dynamics. The way that the activity pattern
evolves over time is inherent to the connectivity of the recurrent neural network.
That's idea number one. Now, the dynamics of that network are shaped by the connectivity,
by the synaptic weights. Those synaptic weights are being shaped by this reinforcement learning
algorithm that you're training the network with. The punchline is, if you train a recurrent neural
network with a reinforcement learning algorithm that's adjusting its weights and you do that for
long enough, the activation dynamics will become very interesting. Imagine I give you a task where
you have to press one button or another, left button or right button. There's some probability
that I'm going to give you an M&M if you press the left button. There's some probability I'll
give you an M&M if you press the other button. You have to figure out what those probabilities are
just by trying things out. As I said before, instead of just giving you one of these tasks,
I give you a whole sequence. I give you two buttons and you figure out which one's best.
I go, good job. Here's a new box. Two new buttons. You have to figure out which one's best. Good
job. Here's a new box. Every box has its own probabilities and you have to figure it out.
If you train a recurrent neural network on that sequence of tasks, what happens? It seemed
almost magical to us when we first started realizing what was going on. The slow learning
algorithm that's adjusting the synaptic weights, those slow synaptic changes give rise to a network
dynamics that themselves turn into a learning algorithm. In other words, you can tell this
is happening by just freezing the synaptic weights saying, okay, no more learning. You're done.
Here's a new box. Figure out which button is best. The recurrent neural network will do this
just fine. It figures out which button is best. It transitions from exploring the two buttons
to just pressing the one that it likes best in a very rational way. How is that happening? It's
happening because the activity dynamics of the network have been shaped by this slow learning
process that's occurred over many, many boxes. What's happened is that this slow learning algorithm
that's slowly adjusting the weights is changing the dynamics of the network, the activity dynamics,
into its own learning algorithm. As we were realizing that this is a thing,
it just so happened that the group that was working on this included a bunch of neuroscientists.
It started ringing a bell for us, which is to say that we thought, this sounds a lot like
the distinction between synaptic learning and activity, synaptic memory and activity-based
memory in the brain. It also reminded us of recurrent connectivity that's very characteristic of
prefrontal function. This is why it's good to have people working on AI that know a little bit
about neuroscience and vice versa, because we started thinking about whether we could apply
this principle to neuroscience. That's where the paper came from. The kind of principle of
the recurrence they can see in the prefrontal cortex, then you start to realize that it's
possible for something like an idea of emerging from this learning process as long as you keep
varying the environment sufficiently. Exactly. The kind of metaphorical transition we made to
neuroscience was to think, okay, well, we know that the prefrontal cortex is highly recurrent.
We know that it's an important locus for working memory, for activation-based memory.
Maybe the prefrontal cortex supports reinforcement learning. In other words,
what is reinforcement learning? You take an action, you see how much reward you got,
you update your policy of behavior. Maybe the prefrontal cortex is doing that sort of thing
strictly in its activation patterns. It's keeping around a memory in its activity patterns of what
you did, how much reward you got, and it's using that activity-based memory as a basis for updating
behavior. Then the question is, well, how did the prefrontal cortex get so smart? In other words,
how did it, where did these activity dynamics come from? How did that program that's implemented in
the recurrent dynamics of the prefrontal cortex arise? One answer that became evident in this work
was, well, maybe the mechanisms that operate on the synaptic level, which we believe are mediated
by dopamine, are responsible for shaping those dynamics. This may be a silly question, but
because this kind of several temporal classes of learning are happening and the learning to
learn emerges, can you keep building stacks of learning to learn to learn, learning to learn
to learn to learn to learn? Because it keeps, I mean, basically abstractions of more powerful
abilities to generalize of learning complex rules. Is this over-stretching this kind of mechanism?
Well, one of the people in AI who started thinking about meta-learning from very early on,
Juergen and Schmidt-Huber, cheekily suggested, I think it may have been in his PhD thesis,
that we should think about meta-meta-meta-meta-meta-meta-meta-learning. That's really what's
going to get us to true intelligence. Certainly, there's a poetic aspect to it, and it seems
interesting and correct that that kind of level of abstraction would be powerful. But
is that something you see in the brain? Is it useful to think of learning in these meta-meta-meta
way, or is it just meta-learning? Well, one thing that really fascinated me about this
mechanism that we were starting to look at, and other groups started talking about very similar
things at the same time, and then a kind of explosion of interest in meta-learning happened
in the AI community shortly after that. I don't know if we had anything to do with that, but
I was gratified to see that a lot of people started talking about meta-learning. One of the
things that I liked about the kind of flavor of meta-learning that we were studying was that
it didn't require anything special. It was just, if you took a system that had some form of memory
that the function of which could be shaped by pick your RL algorithm, then this would just
happen. I mean, there are a lot of forms of meta-learning algorithms that have been proposed
since then that are fascinating and effective in their domains of application. But they're
engineered. There are things that somebody had to say, well, gee, if we wanted meta-learning to
happen, how would we do that? Here's an algorithm that would... But there's something about the
kind of meta-learning that we were studying that seemed to me special in the sense that
it wasn't an algorithm. It was just something that automatically happened if you had a system
that had memory and it was trained with a reinforcement learning algorithm. In that sense,
it can be as meta as it wants to be. There's no limit on how
abstract the meta-learning can get because it's not reliant on a human engineering a particular
meta-learning algorithm to get there. I also, I don't know, I guess I hope that that's relevant
in the brain. I think there's a kind of beauty in the ability of this emergent...
The emergent aspect of it. Yeah, it's something engineered.
Exactly. It's something that just happens in a sense. In a sense, you can't avoid this happening.
If you have a system that has memory and the function of that memory is shaped by reinforcement
learning and this system is trained in a series of interrelated tasks, this is going to happen.
You can't stop it. As long as you have certain properties, maybe like a recurrent structure to...
You have to have memory. It actually doesn't have to be a recurrent neural network. A paper
that I was honored to be involved with even earlier used a kind of slot-based memory.
Do you remember the title? Just before people watched it.
It was memory augmented neural networks. I think the title was meta-learning in memory
augmented neural networks. It was the same exact story. If you have a system with memory,
here it was a different kind of memory, but the function of that memory is shaped by reinforcement
learning. Here it was the reads and writes that occurred on this slot-based memory.
This will just happen. This brings us back to something I was saying earlier about
the importance of the environment. This will happen if the system is being trained in a setting
where there's a sequence of tasks that all share some abstract structure, sometimes talk
about task distributions. That's something that's very obviously true of the world that humans
inhabit. If you just think about what you do every day, you never do exactly the same thing that
you did the day before, but everything that you do has a family resemblance. It shares
structure with something that you did before. The real world is saturated with this property.
It's endless variety with endless redundancy. That's the setting in which this meta-learning
happens. It does seem like we're just so good at finding, just like in this emergent phenomena
you described, we're really good at finding that redundancy, finding those similarities,
the family resemblance. Some people call it, what is it? Melanie Mitchell was talking about
analogies, so we're able to connect concepts together in this same kind of automated emergent
way, which there's so many echoes here of psychology and neuroscience and obviously now with reinforcement
learning with recurrent neural networks at the core. If we could talk a little bit about dopamine,
you're a part of co-authoring really exciting recent paper, very recent in terms of release
on dopamine and temporal difference learning. Can you describe the key ideas of that paper?
Sure. One thing I want to pause to do is acknowledge my co-authors on actually both of
the papers we're talking about. I'll certainly post all their names.
Okay, wonderful. I'm sort of a bashed to be the spokesperson for these papers when
I had such amazing collaborators on both, so it's a comfort to me to know that you'll
acknowledge them. Yeah, this is an incredible team there, but yeah.
Oh yeah, it's such a, it's so much fun. In the case of the dopamine paper, we also
collaborated with Naoichita at Harvard, who obviously a paper simply wouldn't have happened
without him. You were asking for like a thumbnail sketch of?
Yeah, thumbnail sketch or key ideas or things, the insights that continue on our kind of
discussion here between neuroscience and AI. Yeah, I mean, this was another,
a lot of the work that we've done so far is taking ideas that have bubbled up in AI and
asking the question of whether the brain might be doing something related, which
I think on the surface sounds like something that's really mainly of use to neuroscience.
We see it also as a way of validating what we're doing on the AI side. If we can gain
some evidence that the brain is using some technique that we've been trying out in our AI work,
that gives us confidence that it may be a good idea that it'll scale to rich complex tasks,
that it'll interface well with other mechanisms.
So you see it as a two-way road?
Yeah, for sure.
Just because a particular paper is a little bit focused on from one to the, from AI,
from neural networks to neuroscience, ultimately the discussion, the thinking,
the productive long-term aspect of it is the two-way road nature of the whole.
Yeah, I mean, we've talked about the notion of a virtuous circle between AI and neuroscience.
The way I see it, that's always been there since the two fields jointly existed.
There have been some phases in that history when AI was ahead, there are some phases when
neuroscience was ahead. I feel like, given the burst of innovation that's happened recently on
the AI side, AI is ahead in the sense that there are all of these ideas for which it's exciting
to consider that there might be neural analogs. Neuroscience in a sense has been focusing on
approaches to studying behavior that come from, that are derived from this earlier era of cognitive
psychology. So in some ways, fail to connect with some of the issues that we're grappling with in
AI, like how do we deal with large, complex environments. But I think it's inevitable that
this circle will keep turning and there will be a moment in the not too different distant future
when neuroscience is pelting AI researchers with insights that may change the direction of our work.
Just a quick human question. You have parts of your brain, this is very meta, but they're able
to both think about neuroscience and AI. I don't often meet people like that. Let me ask a meta
plasticity question. Do you think a human being can be both good at AI and neuroscience? On the
team at DeepMind, what kind of human can occupy these two realms? Is that something you see everybody
should be doing, can be doing, or is that a very special few can jump? Just like we talked about
art history, I would think it's a special person that can major in art history and also consider
being a surgeon. Otherwise known as a dilettante. Yeah, easily distracted. I think it does take
a special kind of person to be truly world-class at both AI and neuroscience and I am not on that
list. I happen to be someone who's interested in neuroscience and psychology, involved using the
kinds of modeling techniques that are now very central in AI. That bought me a ticket to be
involved in all of the amazing things that are going on in AI research right now. I do know
a few people who I would consider pretty expert on both fronts, and I won't embarrass them by
naming them, but there are exceptional people out there who are like this. The one thing that I
find is a barrier to being truly world-class on both fronts is the complexity of the technology
that's involved in both disciplines now. The engineering expertise that it takes to do
truly frontline hands-on AI research is really, really considerable. The learning curve of the
tools, just like the specifics of just whether it's programming or the kind of tools necessary to
collect the data, to manage the data, to distribute, to compute all that kind of stuff.
Yeah. On the neuroscience side, there'll be all different sets of tools.
Exactly, especially with the recent explosion in neuroscience methods.
Having said all that, I think the best scenario for both neuroscience and AI is to have people
who, interacting, who live at every point on this spectrum from exclusively focused on
neuroscience to exclusively focused on the engineering side of AI, but to have those people
inhabiting a community where they're talking to people who live elsewhere on the spectrum.
I may be someone who's very close to the center in the sense that I have one foot in the neuroscience
world and one foot in the AI world. That central position, I will admit, prevents me, at least
someone with my limited cognitive capacity, from having true technical expertise in either domain.
But at the same time, I at least hope that it's worthwhile having people around who can see the
connections. The emergent intelligence of the community when it's nicely distributed is useful.
Exactly. I've seen that work out well at DeepMind.
Even if you just focus on the AI work that happens at DeepMind, it's been a good thing
to have some people around doing that work whose PhDs are in neuroscience or psychology. Every
academic discipline has its blind spots and unfortunate obsessions and its metaphors
and its reference points. Having some intellectual diversity is really healthy.
People get each other unstuck, I think. I see it all the time at DeepMind. I like to think that the
people who bring some neuroscience background to the table are helping with that.
One of my probably the deepest passion for me, what I would say, maybe you kind of spoke off
Mike a little bit about it, but that I think is a blind spot for at least robotics and AI folks,
is human-robot interaction, human-agent interaction. Maybe do you have thoughts about
how we reduce the size of that blind spot? Do you also share the feeling that not enough
folks are studying this aspect of interaction? I'm actually pretty intensively interested
in this issue now. There are people in my group who've actually pivoted pretty hard over the last
few years from doing more traditional cognitive psychology and cognitive neuroscience to doing
experimental work on human-agent interaction. There are a couple reasons that I'm pretty passionately
interested in this. One is the outcome of having thought for a few years now about
what we're up to. What are we doing? What is this AI research for? What does it mean to make the world
the better place? I think I'm pretty sure that means making life better for humans. How do you
make life better for humans? That's a proposition that when you look at it carefully and honestly,
is rather horrendously complicated, especially when the AI systems that you're
building are learning systems. You're not programming something that you then introduce
to the world and it just works as programmed, like Google Maps or something. We're building
systems that learn from experience. That typically leads to AI safety questions. How do we keep these
things from getting out of control? How do we keep them from doing things that harm humans?
I hasten to say, I consider those hugely important issues. There are large sectors of the research
community at DeepMind and, of course, elsewhere who are dedicated to thinking
hard all day every day about that. I guess I would say a positive side to this too, which is to say,
what would it mean to make human life better? How can we imagine learning systems doing that?
In talking to my colleagues about that, we reached the initial conclusion that
it's not sufficient to philosophize about that. You actually have to take into account how humans
actually work and what humans want, and the difficulties of knowing what humans want,
and the difficulties that arise when humans want different things.
Human-agent interaction has become a quite intensive focus of my group lately.
For no other reason that, in order to really address that issue in an adequate way,
psychology becomes part of the picture.
There's a few elements there. If you focus on the robotics problem,
let's say AGI, without humans in the picture, you're missing, fundamentally, the final step.
When you do want to help human civilization, you eventually have to interact with humans.
When you create a learning system, just as you said, that will eventually have to interact with
humans, the interaction itself has to become part of the learning process.
Right. You can't just watch, well, my sense is,
it sounds like your sense is, you can't just watch humans to learn about humans.
You have to also be part of the human world. You have to interact with humans.
Yeah, exactly. Then questions arise that start imperceptibly but inevitably to slip beyond
the realm of engineering. Questions like, if you have an agent that can do something
that you can't do, under what conditions do you want that agent to do it?
I have a robot that can play Beethoven sonatas better than any human in the sense that
the sensitivity, the expression is just beyond what any human, do I want to listen to that?
Do I want to go to a concert and hear a robot play? These aren't engineering questions.
These are questions about human preference and human culture.
Psychology, bordering on philosophy.
Yeah. Then you start asking, well, even if we knew the answer to that,
is it our place as AI engineers to build that into these agents? Probably the agents should
interact with humans beyond the population of AI engineers and figure out what those
humans want. Then when you start, I referred this the moment ago, but even that becomes
complicated. What if two humans want different things and you have only one agent that's able
to interact with them and try to satisfy their preferences? Then you're into the realm of
economics and social choice theory and even politics. There's a sense in which if you
follow what we're doing to its logical conclusion, then it goes beyond questions of engineering
and technology and starts to shade in perceptibly into questions about what kind of society do you
want? Actually, once that dawned on me, I actually felt, I don't know what the right word is, quite
refreshed in my involvement in AI research. Building this kind of stuff is going to lead
us back to asking really fundamental questions about what's the good life and who gets to decide
it and bringing in viewpoints from multiple sub-communities to help us shape the way that
we live. There's something, it started making me feel like doing AI research in a fully responsible
way could potentially lead to a kind of cultural renewal.
Yeah, it's the way to understand human beings at the individual, the societal level and maybe
become a way to answer all the silly human questions of the meaning of life and all those
kinds of things. Even if it doesn't give us a way of answering those questions, it may
force us back to thinking about them. It might restore a certain, I don't know,
a certain depth or even, dare I say, spirituality to the way to the world.
I don't know. Maybe that's too grandiose.
Well, I'm with you. AI will be the philosophy of the 21st century, the way which will open
the door. I think a lot of AI researchers are afraid to open that door of exploring the beautiful
richness of the human-agent interaction, human-AI interaction. I'm really happy that somebody
like you have opened that door.
One thing I often think about is the usual schema for thinking about human-agent interaction is
this kind of dystopian, oh, our robot overlords. Again, I hasten to say AI safety is hugely
important and I'm not saying we shouldn't be thinking about those risks. Totally on board for
that. Having said that, what often follows for me is the thought that there's another
kind of narrative that might be relevant, which is when we think of humans gaining more and more
information about human life, the narrative there is usually that they gain more and more wisdom
and they get closer to enlightenment and they become more benevolent. The Buddha is like
that's a totally different narrative. Why isn't it the case that we imagine that the AI systems
that we're creating, they're going to figure out more and more about the way the world works
and the way that humans interact and they'll become beneficent. I'm not saying that will
happen. I don't honestly expect that to happen without some careful setting things up very
carefully, but it's another way things could go. I would even push back on that. I personally
believe that the most trajectories, natural human trajectories will lead us towards progress.
For me, there is a kind of sense that most trajectories in AI development will lead us
into trouble. To me, and we over focus on the worst case, it's like in computer science,
theoretical computer science has been this focus on worst case analysis. There's something appealing
to our human mind at some lowest level. We don't want to be eaten by the tiger, I guess.
Yes. We want to do the worst case analysis, but the reality is that shouldn't stop us from
actually building out all the other trajectories which are potentially leading to all the positive
worlds, all the enlightenment. There's a book, Enlightenment Now with Stephen Panker and so on.
This is looking generally at human progress. There's so many ways that human progress can
happen with AI. I think you have to do that research. You have to do that work. You have to do
the not just the AI safety work of the one worst case analysis, how do we prevent that, but the
actual tools and the glue and the mechanisms of human AI interaction that would lead to
all the positive interactions that can go. It's a super exciting area, right?
We should be spending a lot of our time saying what can go wrong. I think it's harder to see
that there's work to be done to bring into focus the question of what it would look like for things
to go right. That's not obvious. We wouldn't be doing this if we didn't have the sense there was
huge potential. We're not doing this for no reason. We have a sense that AGI would be
a major boom to humanity. I think it's worth starting now, even when our technology is quite
primitive, asking exactly what would that mean? We can start now with applications that are already
going to make the world a better place like solving protein folding. I think this deep
mind has gotten heavy into science applications lately, which I think is a wonderful, wonderful
move for us to be making. When we think about AGI, when we think about building fully intelligent
agents that are going to be able to, in a sense, do whatever they want, we should start thinking
about what do we want them to want? What kind of world do we want to live in? That's not an easy
question. I think we just need to start working on it. Even on the path to AGI, it doesn't have to
be AGI. We're just intelligent agents that interact with us and help us enrich our own
existence on social networks, for example, and recommend our systems of various intelligence.
There's so much interesting interaction that's yet to be understood and studied.
Twitter is struggling with this very idea of how do you create AI systems that increase the
quality and the health of a conversation? For sure. That's a beautiful human psychology question.
How do you do that without deception being involved, without manipulation being involved,
maximizing human autonomy? How do you make these choices in a democratic way?
Again, I'm speaking for myself here. How do we face the fact that it's a small group of
people who have the skillset to build these kinds of systems, but what it means to make the world
a better place is something that we all have to be talking about. The world that we're trying to
make a better place includes a huge variety of different kinds of people. How do we cope with
that? This is a problem that has been discussed in gory extensive detail in social choice theory.
One thing I'm really enjoying about the recent direction work has taken in some parts of my
team is that we're reading the AI literature, we're reading the neuroscience literature,
but we've also started reading economics and, as I mentioned, social choice theory,
even some political theory, because it turns out that it all becomes relevant. At the same time,
we've been trying not to write philosophy papers. We've been trying not to write position papers.
We're trying to figure out ways of doing actual empirical research that take the first small
steps to thinking about what it really means for humans with all of their complexity and
contradiction and paradox to be brought into contact with these AI systems in a way that
really makes the world a better place. Often reinforcement learning frameworks actually
allow you to do that machine learning. That's the exciting thing about AI is it allows you to
reduce the unsolvable problem, philosophical problem into something more
concrete that you can get a hold of. Yeah, and it allows you to define the problem in some way that
allows for growth in the system that you're not responsible for the details. You say,
this is generally what I want you to do, and then learning takes care of the rest.
Of course, the safety issues arise in that context, but I think also some of these positive
issues arise in that context. What would it mean for an AI system to really come to understand
what humans want with all of the subtleties of that? Humans want help with certain things,
but they don't want everything done for them. Part of the satisfaction that humans get from
life is in accomplishing things. If there were devices around that did everything for you, I
often think of the movie Wally. That's like dystopian in a totally different way. It's like,
the machines are doing everything for us. That's not what we wanted. Anyway, I find this
opens up a whole landscape of research that feels affirmative and exciting.
To me, it's one of the most exciting and it's wide open. We have to, because it's a cool paper,
talk about dopamine. I was going to give you a quick summary.
It's a quick summary of what's the title of the paper.
I think we called it a distributional code for value in dopamine-based reinforcement learning.
Yes. That's another project that grew out of pure AI research. A number of people at DeepMind
and a few other places had started working on a new version of reinforcement learning,
which was defined by taking something in traditional reinforcement learning and just
tweaking it. The thing that they took from traditional reinforcement learning was
a value signal. At the center of reinforcement learning, at least most algorithms, is some
representation of how well things are going. You're expected cumulative future reward.
That's usually represented as a single number. If you imagine a gambler in a casino and the
gambler is thinking, well, I have this probability of winning such and such an amount of money and
I have this probability of losing such and such an amount of money, that situation would be
represented as a single number, which is the expected weighted average of all those outcomes.
This new form of reinforcement learning said, well, what if we generalize that to
distributional representation? Now we think of the gambler as literally thinking, well,
there's this probability that I'll win this amount of money and there's this probability
that I'll lose that amount of money. We don't reduce that to a single number.
It had been observed through experiments, through just trying this out, that kind of
distributional representation really accelerated reinforcement learning and led to better
policies. What's your intuition about? We're talking about rewards. What's your intuition?
Why that is? Why does it depend? It's kind of a surprising historical note. At least
surprised me when I learned it, that this had been tried out in a kind of heuristic way. People
thought, well, gee, what would happen if we tried and then it had this empirically, it had this
striking effect. It was only then that people started thinking, well, gee, why is this working?
That's led to a series of studies just trying to figure out why it works, which is ongoing.
But one thing that's already clear from that research is that one reason that it helps is that
it drives richer representation learning. If you imagine two situations that have the same
expected value, the same kind of weighted average value, standard deep reinforcement
learning algorithms are going to take those two situations and in terms of the way they're
represented internally, squeeze them together. Because the thing that you're trying to
represent, which is their expected value, is the same. So all the way through the system,
things are going to be mushed together. But what if those two situations actually have
different value distributions? They have the same average value, but they have different
distributions of value. In that situation, distributional learning will maintain the
distinction between these two things. To make a long story short, distributional learning
can keep things separate in the internal representation that might otherwise be conflated
or squished together. Maintaining those distinctions can be useful when the system is
now faced with some other task where the distinction is important.
If we look at optimistic and pessimistic dopamine neurons, so first of all, what is dopamine?
Why is this at all useful to think about in the artificial intelligence sense,
but what do we know about dopamine in the human brain? What is it? Why is it useful?
Why is it interesting? What does it have to do with the prefrontal cortex and learning in general?
Yeah. Well, this is also a case where there's a huge amount of detail and debate. But one
currently prevailing idea is that the function of this neurotransmitter dopamine
resembles a particular component of standard reinforcement learning algorithms, which is
called the reward prediction error. I was talking a moment ago about these value representations.
How do you learn them? How do you update them based on experience? Well, if you made some
prediction about a future reward and then you get more reward than you were expecting,
then probably retrospectively, you want to go back and increase the value representation that
you attached to that earlier situation. If you got less reward than you were expecting,
you should probably decrement that estimate. That's the process of temporal difference learning.
Exactly. This is the central mechanism of temporal difference learning, which is one of
several backbone of our armamentarium in RL. This connection between the reward prediction error
and dopamine was made in the 1990s. There's been a huge amount of research that seems to back it
up. Dopamine may be doing other things, but this is clearly at least roughly one of the things that
it's doing. But the usual idea was that dopamine was representing these reward prediction errors
again in this single number way, representing your surprise with a single number. In distributional
reinforcement learning, this new elaboration of the standard approach, it's not only the value,
the value function that's represented as a single number, it's also the reward prediction error.
What happened was that Will Dabney, one of my collaborators who was one of the first people
to work on distributional temporal difference learning, talked to a guy in my group, Zeb
Kurt Nelson, who's a computational neuroscientist, and said, gee, is it possible that dopamine
might be doing something like this distributional coding thing? They started looking at what was
in the literature, and then they brought me in, and we started talking to Nao Uchida,
and we came up with some specific predictions about if the brain is using this kind of distributional
coding, then in the tasks that now has studied, you should see this, this, this, and this,
and that's where the paper came from. We enumerated a set of predictions,
all of which ended up being fairly clearly confirmed, and all of which leads to at least
some initial indication that the brain might be doing something like this distributional coding,
that dopamine might be representing surprise signals in a way that is not just collapsing
everything to a single number, but instead is kind of respecting the variety of future outcomes,
if that makes sense. So yeah, so that's showing, suggesting possibly that dopamine has a really
interesting representation scheme in the human brain for its reward signal.
Exactly. That's fascinating. That's just, that's another beautiful example of AI revealing
something nice about neuroscience, potentially suggesting possibilities.
Well, you never know. So the minute you publish paper like that, the next thing you think is,
I hope that replicates. I hope we see that same thing in other data sets, but of course,
several labs now are doing the follow-up experiments, so we'll know soon. But it has been,
it has been a lot of fun for us to take these ideas from AI and kind of bring them into neuroscience,
and, you know, see how far we can get. So we kind of talked about it a little bit,
but where do you see the field of neuroscience and artificial intelligence heading broadly?
Like what are the possible exciting areas that you can see breakthroughs in the next,
let's get crazy, not just three or five years, but next 10, 20, 30 years.
That would make you excited, and perhaps you'd be part of.
On the neuroscience side, there's a great deal of interest now in what's going on in AI.
And at the same time, I feel like, so neuroscience, especially the part of neuroscience that's
focused on circuits and systems, you know, kind of like really mechanism focused,
there's been this explosion in new technology. And up until recently, the experiments that have
exploited this technology have not involved a lot of interesting behavior. And this is for a
variety of reasons, you know, one of which is in order to employ some of these technologies,
you actually have to, if you're studying a mouse, you have to head fix the mouse. In other words,
you have to immobilize the mouse. And so it's been tricky to come up with ways of
eliciting interesting behavior from a mouse that's restrained in this way. But people have
begun to create very interesting solutions to this, like virtual reality environments where
the animal can move a trackball. And as people have begun to explore what you can do with these
technologies, I feel like more and more people are asking, well, let's try to bring behavior
into the picture. Let's try to like reintroduce behavior, which was supposed to be what this
whole thing was about. And I'm hoping that those two trends, the kind of growing interest in behavior
and the widespread interest in what's going on in AI will come together to kind of open a new
chapter in neuroscience research where there's a kind of a rebirth of interest in the structure
of behavior and its underlying substrates, but that that research is being informed by
computational mechanisms that we're coming to understand in AI. You know, if we can do that,
then we might be taking a step closer to this utopian future that we were talking about earlier
where there's really no distinction between psychology and neuroscience. Neuroscience is
about studying the mechanisms that underlie whatever it is the brain is for and what is
the brain for? It's for behavior. I feel like we could maybe take a step toward that now
if people are motivated in the right way. You also asked about AI. So that was a neuroscience
question. You said neuroscience. That's right. And especially places like DeepMind are interested
in both branches. What about the engineering of intelligence systems?
I think one of the key challenges that a lot of people are seeing now in AI is to build systems
that have the kind of flexibility and the kind of flexibility that humans have in two senses.
One is that humans can be good at many things. They're not just expert at one thing. And they're
also flexible in the sense that they can switch between things very easily and they can pick up
new things very quickly because they very able to see what a new task has in common
with other things that they've done. And that's something that our AI systems just
blatantly do not have. There are some people who like to argue that deep learning and deep RL
are simply wrong for getting that kind of flexibility. I don't share that belief. But
the simpler fact of the matter is we're not building things yet that do have that kind of
flexibility. And I think the attention of a large part of the AI community is starting to pivot
to that question. How do we get that? That's going to lead to a focus on abstraction. It's
going to lead to a focus on what in psychology we call cognitive control, which is the ability
to switch between tasks, the ability to quickly put together a program of behavior that you've
never executed before, but you know makes sense for a particular set of demands. It's very closely
related to what the prefrontal cortex does on the neuroscience side. So I think it's
going to be an interesting new chapter. So that's the reasoning side and cognition side,
but let me ask the over romanticized question. Do you think we'll ever engineer an AGI system
that we humans would be able to love and that would love us back? So have that level and depth
of connection? I love that question. And it relates closely to things that I've been thinking about
a lot lately in the context of this human AI research. There's social psychology research
in particular by Susan Fisk at Princeton in the department where I used to work,
where she dissects human attitudes toward other humans into a two-dimensional scheme.
One dimension is about ability. How able, how capable is this other person?
But the other dimension is warmth. So you can imagine another person who's very skilled
and capable, but it's very cold. And you wouldn't really like highly, you might have some reservations
about that other person. But there's also a kind of reservation that we might have about
another person who elicits in us or displays a lot of human warmth, but is not good at getting
things done. The greatest esteem that we, we reserve our greatest esteem really for people who
are both highly capable and also quite warm. That's the best of the best. This isn't a
normative statement I'm making. This is just an empirical statement. These are the two dimensions
that people seem to kind of like, along which people size other people up. And in AI research,
we really focus on this capability thing. We want our agents to be able to do stuff. This thing
can play go at a super human level. That's awesome. But that's only one dimension. What's the,
what about the other dimension? What would it mean for an AI system to be warm? And I don't know,
maybe there are easy solutions here like we can put a face on our AI systems. It's cute. It has big
ears. I mean, that's probably part of it. But I think it also has to do with a pattern of behavior,
a pattern of, you know, what would it mean for an AI system to display caring, compassionate
behavior in a way that actually made us feel like it was for real, that we didn't feel like it was
simulated. We didn't feel like we were being duped. To me, that, you know, people talk about the
Turing test or some, some descendant of it. I feel like that's the ultimate Turing test.
You know, is there, is there an AI system that can not only convince us that it knows how to
reason and it knows how to interpret language, but that we're comfortable saying, yeah, that AI
system is a good guy. You know, like, I mean, that on the warmth scale, whatever warmth is,
we kind of intuitively understand it, but we also want to be able to, yeah, we don't understand it
explicitly enough yet to be able to engineer it. Exactly. And that's, and that's an open
scientific question. You kind of alluded to it several times in the human AI interaction.
That's the question that should be studied. And probably one of the most important questions as
we move to AI. And humans, we humans are, are so good at it. Yeah. You know, it's not just weird.
It's not just that we're born warm, you know, like, I suppose some people are, are warmer than
others given, you know, whatever genes they manage to inherit. But there's also, there's also,
there are also learned skills involved, right? I mean, there are ways of communicating to other
people that you care that they matter to you, that you're enjoying interacting with them,
right? And we learn these skills from one another. And it's not out of the question
that we could build engineered systems. I think it's hopeless, as you say, that we could somehow
hand design these sorts of, these sorts of behaviors. But it's not out of the question
that we could build systems that kind of, we, we, we instill in them something that
sets them out in the right direction. So that they, they end up learning what it is to interact
with humans in a way that's gratifying to humans. I mean, honestly, if that's not where we're headed,
I want out. I think it's exciting as a scientific problem, just as you described.
I honestly don't see a better way to end it than talking about warmth and love. And Matt,
I don't think I've ever had such a wonderful conversation where my questions were so bad
and your answers were so beautiful. So I deeply appreciate it. I really enjoyed it.
It's been very fun. I, you know, as you can probably tell, I, I really, you know, I, there's
something I like about kind of thinking outside the box and like, so it's good having fun with
that. Awesome. Thanks so much for doing it. Thanks for listening to this conversation with Matt
Boffinick. And thank you to our sponsors, The Jordan Harbinger Show and Magic Spoon Low Carb
Keto Cereal. Please consider supporting this podcast by going to jordanharbinger.com slash
lex and also go into magic spoon.com slash lex and using code lex at checkout. Click the links,
buy all the stuff. It's the best way to support this podcast and the journey I'm on
in my research and the startup. If you enjoy this thing, subscribe on YouTube, review it
with the five stars in a podcast, support on Patreon, follow on Spotify or connect with me
on Twitter at Lex Friedman. Again, spelled miraculously without the E, just F-R-I-D-M-A-N.
And now let me leave you with some words from neurologist V.S. Sarmachandran. How can a three
pound mass of jelly that you can hold in your palm, imagine angels, contemplate the meaning of an
affinity, even question its own place in cosmos, especially awe inspiring is the fact that any
single brain, including yours, is made up of atoms that were forged in the hearts of countless,
far flung stars billions of years ago. These particles drifted for eons and light years
until gravity and change brought them together here now. These atoms now form a conglomerate,
your brain, that can not only ponder the very stars they gave at birth, but can also think about
its own ability to think and wander about its own ability to wander. With the arrival of humans,
it has been said, the universe has suddenly become conscious of itself. This, truly,
is the greatest mystery of all. Thank you for listening and hope to see you next time.