Transcript
00:00:00 The following is a conversation with Matt Botmanek,
00:00:03 Director of Neuroscience Research at DeepMind.
00:00:06 He’s a brilliant, cross disciplinary mind,
00:00:09 navigating effortlessly between cognitive psychology,
00:00:12 computational neuroscience, and artificial intelligence.
00:00:16 Quick summary of the ads.
00:00:18 Two sponsors, The Jordan Harbinger Show
00:00:21 and Magic Spoon Cereal.
00:00:23 Please consider supporting the podcast
00:00:25 by going to jordanharbinger.com slash lex
00:00:29 and also going to magicspoon.com slash lex
00:00:33 and using code lex at checkout
00:00:36 after you buy all of their cereal.
00:00:39 Click the links, buy the stuff.
00:00:40 It’s the best way to support this podcast
00:00:43 and the journey I’m on.
00:00:44 If you enjoy this podcast, subscribe on YouTube,
00:00:47 review it with five stars on Apple Podcast,
00:00:49 follow on Spotify, support on Patreon,
00:00:52 or connect with me on Twitter at lexfriedman,
00:00:55 spelled surprisingly without the E,
00:00:58 just F R I D M A N.
00:01:02 As usual, I’ll do a few minutes of ads now
00:01:03 and never any ads in the middle
00:01:05 that can break the flow of the conversation.
00:01:07 This episode is supported by The Jordan Harbinger Show.
00:01:11 Go to jordanharbinger.com slash lex.
00:01:15 It’s how he knows I sent you.
00:01:16 On that page, subscribe to his podcast
00:01:19 on Apple Podcast, Spotify, and you know where to look.
00:01:24 I’ve been binging on his podcast.
00:01:26 Jordan is a great interviewer
00:01:28 and even a better human being.
00:01:30 I recently listened to his conversation with Jack Barsky,
00:01:32 former sleeper agent for the KGB in the 80s
00:01:36 and author of Deep Undercover,
00:01:38 which is a memoir that paints yet another
00:01:40 interesting perspective on the Cold War era.
00:01:43 I’ve been reading a lot about the Stalin
00:01:46 and then Gorbachev and Putin eras of Russia,
00:01:49 but this conversation made me realize
00:01:50 that I need to do a deep dive into the Cold War era
00:01:53 to get a complete picture of Russia’s recent history.
00:01:57 Again, go to jordanharbinger.com slash lex.
00:02:01 Subscribe to his podcast.
00:02:02 It’s how he knows I sent you.
00:02:04 It’s awesome, you won’t regret it.
00:02:06 This episode is also supported by Magic Spoon,
00:02:10 low carb, keto friendly, super amazingly delicious cereal.
00:02:15 I’ve been on a keto or very low carb diet
00:02:18 for a long time now.
00:02:19 It helps with my mental performance.
00:02:21 It helps with my physical performance,
00:02:22 even during this crazy push up, pull up challenge I’m doing,
00:02:26 including the running, it just feels great.
00:02:29 I used to love cereal.
00:02:31 Obviously, I can’t have it now
00:02:33 because most cereals have crazy amounts of sugar,
00:02:36 which is terrible for you, so I quit it years ago.
00:02:40 But Magic Spoon, amazingly, somehow,
00:02:44 is a totally different thing.
00:02:45 Zero sugar, 11 grams of protein,
00:02:48 and only three net grams of carbs.
00:02:50 It tastes delicious.
00:02:53 It has a lot of flavors, two new ones,
00:02:55 including peanut butter.
00:02:56 But if you know what’s good for you,
00:02:58 you’ll go with cocoa, my favorite flavor,
00:03:01 and the flavor of champions.
00:03:04 Click the magicspoon.com slash lex link in the description
00:03:07 and use code lex at checkout for free shipping
00:03:11 and to let them know I sent you.
00:03:13 They have agreed to sponsor this podcast for a long time.
00:03:16 They’re an amazing sponsor and an even better cereal.
00:03:19 I highly recommend it.
00:03:21 It’s delicious, it’s good for you, you won’t regret it.
00:03:24 And now, here’s my conversation with Matt Botpenik.
00:03:29 How much of the human brain do you think we understand?
00:03:33 I think we’re at a weird moment
00:03:36 in the history of neuroscience in the sense that
00:03:45 I feel like we understand a lot about the brain
00:03:47 at a very high level, but a very coarse level.
00:03:52 When you say high level, what are you thinking?
00:03:54 Are you thinking functional?
00:03:55 Are you thinking structurally?
00:03:56 So in other words, what is the brain for?
00:04:00 What kinds of computation does the brain do?
00:04:05 What kinds of behaviors would we have to explain
00:04:12 if we were gonna look down at the mechanistic level?
00:04:16 And at that level, I feel like we understand
00:04:18 much, much more about the brain
00:04:19 than we did when I was in high school.
00:04:22 But it’s almost like we’re seeing it through a fog.
00:04:25 It’s only at a very coarse level.
00:04:26 We don’t really understand what the neuronal mechanisms are
00:04:30 that underlie these computations.
00:04:32 We’ve gotten better at saying,
00:04:34 what are the functions that the brain is computing
00:04:36 that we would have to understand
00:04:38 if we were gonna get down to the neuronal level?
00:04:40 And at the other end of the spectrum,
00:04:45 in the last few years, incredible progress has been made
00:04:49 in terms of technologies that allow us to see,
00:04:54 actually literally see, in some cases,
00:04:57 what’s going on at the single unit level,
00:05:01 even the dendritic level.
00:05:02 And then there’s this yawning gap in between.
00:05:05 Well, that’s interesting.
00:05:06 So at the high level,
00:05:07 so that’s almost a cognitive science level.
00:05:09 And then at the neuronal level,
00:05:11 that’s neurobiology and neuroscience,
00:05:14 just studying single neurons,
00:05:16 the synaptic connections and all the dopamine,
00:05:19 all the kind of neurotransmitters.
00:05:21 One blanket statement I should probably make
00:05:23 is that as I’ve gotten older,
00:05:27 I have become more and more reluctant
00:05:30 to make a distinction between psychology and neuroscience.
00:05:33 To me, the point of neuroscience
00:05:37 is to study what the brain is for.
00:05:41 If you’re a nephrologist
00:05:44 and you wanna learn about the kidney,
00:05:46 you start by saying, what is this thing for?
00:05:50 Well, it seems to be for taking blood on one side
00:05:55 that has metabolites in it that shouldn’t be there,
00:06:01 sucking them out of the blood
00:06:03 while leaving the good stuff behind,
00:06:05 and then excreting that in the form of urine.
00:06:07 That’s what the kidney is for.
00:06:08 It’s like obvious.
00:06:10 So the rest of the work is deciding how it does that.
00:06:13 And this, it seems to me,
00:06:14 is the right approach to take to the brain.
00:06:17 You say, well, what is the brain for?
00:06:19 The brain, as far as I can tell, is for producing behavior.
00:06:22 It’s for going from perceptual inputs to behavioral outputs,
00:06:27 and the behavioral outputs should be adaptive.
00:06:31 So that’s what psychology is about.
00:06:33 It’s about understanding the structure of that function.
00:06:35 And then the rest of neuroscience is about figuring out
00:06:38 how those operations are actually carried out
00:06:41 at a mechanistic level.
00:06:44 That’s really interesting, but so unlike the kidney,
00:06:47 the brain, the gap between the electrical signal
00:06:52 and behavior, so you truly see neuroscience
00:06:57 as the science that touches behavior,
00:07:01 how the brain generates behavior,
00:07:03 or how the brain converts raw visual information
00:07:07 into understanding.
00:07:08 Like, you basically see cognitive science,
00:07:12 psychology, and neuroscience as all one science.
00:07:15 Yeah, it’s a personal statement.
00:07:19 Is that a hopeful or a realistic statement?
00:07:22 So certainly you will be correct in your feeling
00:07:26 in some number of years, but that number of years
00:07:29 could be 200, 300 years from now.
00:07:31 Oh, well, there’s a…
00:07:33 Is that aspirational or is that pragmatic engineering
00:07:37 feeling that you have?
00:07:39 It’s both in the sense that this is what I hope
00:07:46 and expect will bear fruit over the coming decades,
00:07:53 but it’s also pragmatic in the sense that I’m not sure
00:07:57 what we’re doing in either psychology or neuroscience
00:08:02 if that’s not the framing.
00:08:04 I don’t know what it means to understand the brain
00:08:09 if there’s no, if part of the enterprise
00:08:14 is not about understanding the behavior
00:08:18 that’s being produced.
00:08:20 I mean, yeah, but I would compare it
00:08:23 to maybe astronomers looking at the movement
00:08:25 of the planets and the stars without any interest
00:08:30 of the underlying physics, right?
00:08:32 And I would argue that at least in the early days,
00:08:35 there is some value to just tracing the movement
00:08:37 of the planets and the stars without thinking
00:08:41 about the physics too much because it’s such a big leap
00:08:44 to start thinking about the physics
00:08:45 before you even understand even the basic structural
00:08:48 elements of…
00:08:49 Oh, I agree with that.
00:08:50 I agree.
00:08:51 But you’re saying in the end, the goal should be
00:08:53 to deeply understand.
00:08:54 Well, right, and I think…
00:08:57 So I thought about this a lot when I was in grad school
00:08:59 because a lot of what I studied in grad school
00:09:00 was psychology and I found myself a little bit confused
00:09:06 about what it meant to…
00:09:08 It seems like what we were talking about a lot of the time
00:09:11 were virtual causal mechanisms.
00:09:14 Like, oh, well, you know, attentional selection
00:09:18 then selects some object in the environment
00:09:22 and that is then passed on to the motor, you know,
00:09:25 information about that is passed on to the motor system.
00:09:27 But these are virtual mechanisms.
00:09:29 These are, you know, they’re metaphors.
00:09:31 They’re, you know, there’s no reduction going on
00:09:37 in that conversation to some physical mechanism that,
00:09:40 you know, which is really what it would take
00:09:43 to fully understand, you know, how behavior is rising.
00:09:47 But the causal mechanisms are definitely neurons interacting.
00:09:50 I’m willing to say that at this point in history.
00:09:53 So in psychology, at least for me personally,
00:09:56 there was this strange insecurity about trafficking
00:10:00 in these metaphors, you know,
00:10:02 which were supposed to explain the function of the mind.
00:10:07 If you can’t ground them in physical mechanisms,
00:10:09 then what is the explanatory validity of these explanations?
00:10:16 And I managed to soothe my own nerves
00:10:21 by thinking about the history of genetics research.
00:10:29 So I’m very far from being an expert
00:10:32 on the history of this field.
00:10:34 But I know enough to say that, you know,
00:10:38 Mendelian genetics preceded, you know, Watson and Crick.
00:10:42 And so there was a significant period of time
00:10:45 during which people were, you know,
00:10:49 productively investigating the structure of inheritance
00:10:54 using what was essentially a metaphor,
00:10:56 the notion of a gene, you know.
00:10:58 Oh, genes do this and genes do that.
00:11:00 But, you know, where are the genes?
00:11:02 They’re sort of an explanatory thing that we made up.
00:11:06 And we ascribed to them these causal properties.
00:11:08 Oh, there’s a dominant, there’s a recessive,
00:11:10 and then they recombine it.
00:11:12 And then later, there was a kind of blank there
00:11:17 that was filled in with a physical mechanism.
00:11:21 That connection was made.
00:11:24 But it was worth having that metaphor
00:11:26 because that gave us a good sense
00:11:29 of what kind of causal mechanism we were looking for.
00:11:34 And the fundamental metaphor of cognition, you said,
00:11:38 is the interaction of neurons.
00:11:40 Is that, what is the metaphor?
00:11:42 No, no, the metaphor,
00:11:44 the metaphors we use in cognitive psychology
00:11:47 are things like attention, the way that memory works.
00:11:56 I retrieve something from memory, right?
00:11:59 A memory retrieval occurs.
00:12:01 What is that?
00:12:02 You know, that’s not a physical mechanism
00:12:06 that I can examine in its own right.
00:12:08 But it’s still worth having, that metaphorical level.
00:12:13 Yeah, so yeah, I misunderstood actually.
00:12:16 So the higher level of abstractions
00:12:17 is the metaphor that’s most useful.
00:12:19 Yes.
00:12:20 But what about, so how does that connect
00:12:24 to the idea that that arises from interaction of neurons?
00:12:33 Well, even, is the interaction of neurons
00:12:35 also not a metaphor to you?
00:12:38 Or is it literally, like that’s no longer a metaphor.
00:12:42 That’s already the lowest level of abstractions
00:12:46 that could actually be directly studied.
00:12:50 Well, I’m hesitating because I think
00:12:53 what I want to say could end up being controversial.
00:12:57 So what I want to say is, yes,
00:12:59 the interactions of neurons, that’s not metaphorical.
00:13:03 That’s a physical fact.
00:13:04 That’s where the causal interactions actually occur.
00:13:08 Now, I suppose you could say,
00:13:09 well, even that is metaphorical relative
00:13:12 to the quantum events that underlie.
00:13:15 I don’t want to go down that rabbit hole.
00:13:17 It’s always turtles on top of turtles.
00:13:18 Yeah, there’s turtles all the way down.
00:13:21 There’s a reduction that you can do.
00:13:22 You can say these psychological phenomena
00:13:25 can be explained through a very different
00:13:28 kind of causal mechanism,
00:13:29 which has to do with neurotransmitter release.
00:13:31 And so what we’re really trying to do
00:13:33 in neuroscience writ large, as I say,
00:13:37 which for me includes psychology,
00:13:39 is to take these psychological phenomena
00:13:44 and map them onto neural events.
00:13:49 I think remaining forever at the level of description
00:13:57 that is natural for psychology,
00:14:00 for me personally, would be disappointing.
00:14:02 I want to understand how mental activity
00:14:05 arises from neural activity.
00:14:10 But the converse is also true.
00:14:13 Studying neural activity without any sense
00:14:15 of what you’re trying to explain,
00:14:19 to me feels like at best groping around at random.
00:14:27 Now, you’ve kind of talked about this bridging
00:14:30 of the gap between psychology and neuroscience,
00:14:32 but do you think it’s possible,
00:14:34 like my love is, like I fell in love with psychology
00:14:38 and psychiatry in general with Freud
00:14:40 and when I was really young,
00:14:41 and I hoped to understand the mind.
00:14:43 And for me, understanding the mind,
00:14:45 at least at that young age before I discovered AI
00:14:48 and even neuroscience was to, is psychology.
00:14:52 And do you think it’s possible to understand the mind
00:14:55 without getting into all the messy details of neuroscience?
00:14:59 Like you kind of mentioned to you it’s appealing
00:15:03 to try to understand the mechanisms at the lowest level,
00:15:06 but do you think that’s needed,
00:15:07 that’s required to understand how the mind works?
00:15:11 That’s an important part of the whole picture,
00:15:14 but I would be the last person on earth
00:15:18 to suggest that that reality
00:15:23 renders psychology in its own right unproductive.
00:15:29 I trained as a psychologist.
00:15:31 I am fond of saying that I have learned much more
00:15:35 from psychology than I have from neuroscience.
00:15:38 To me, psychology is a hugely important discipline.
00:15:43 And one thing that warms in my heart is that
00:15:50 ways of investigating behavior
00:15:54 that have been native to cognitive psychology
00:15:58 since it’s dawn in the 60s
00:16:01 are starting to become,
00:16:03 they’re starting to become interesting to AI researchers
00:16:07 for a variety of reasons.
00:16:09 And that’s been exciting for me to see.
00:16:11 Can you maybe talk a little bit about what you see
00:16:14 as beautiful aspects of psychology,
00:16:19 maybe limiting aspects of psychology?
00:16:21 I mean, maybe just start it off as a science, as a field.
00:16:25 To me, it was when I understood what psychology is,
00:16:29 analytical psychology,
00:16:30 like the way it’s actually carried out,
00:16:32 it was really disappointing to see two aspects.
00:16:36 One is how small the N is,
00:16:39 how small the number of subject is in the studies.
00:16:43 And two, it was disappointing to see
00:16:45 how controlled the entire,
00:16:47 how much it was in the lab.
00:16:50 It wasn’t studying humans in the wild.
00:16:52 There was no mechanism for studying humans in the wild.
00:16:55 So that’s where I became a little bit disillusioned
00:16:57 to psychology.
00:16:59 And then the modern world of the internet
00:17:01 is so exciting to me.
00:17:02 The Twitter data or YouTube data,
00:17:05 data of human behavior on the internet becomes exciting
00:17:08 because the N grows and then in the wild grows.
00:17:11 But that’s just my narrow sense.
00:17:13 Like, do you have a optimistic or pessimistic
00:17:16 cynical view of psychology?
00:17:18 How do you see the field broadly?
00:17:21 When I was in graduate school,
00:17:22 it was early enough that there was still a thrill
00:17:27 in seeing that there were ways of doing,
00:17:32 there were ways of doing experimental science
00:17:36 that provided insight to the structure of the mind.
00:17:40 One thing that impressed me most when I was at that stage
00:17:43 in my education was neuropsychology,
00:17:46 looking at, analyzing the behavior of populations
00:17:51 who had brain damage of different kinds
00:17:55 and trying to understand what the specific deficits were
00:18:02 that arose from a lesion in a particular part of the brain.
00:18:06 And the kind of experimentation that was done
00:18:08 and that’s still being done to get answers in that context
00:18:13 was so creative and it was so deliberate.
00:18:18 It was good science.
00:18:21 An experiment answered one question but raised another
00:18:24 and somebody would do an experiment
00:18:25 that answered that question.
00:18:26 And you really felt like you were narrowing in on
00:18:29 some kind of approximate understanding
00:18:31 of what this part of the brain was for.
00:18:34 Do you have an example from memory
00:18:36 of what kind of aspects of the mind
00:18:39 could be studied in this kind of way?
00:18:41 Oh, sure.
00:18:42 I mean, the very detailed neuropsychological studies
00:18:45 of language function,
00:18:49 looking at production and reception
00:18:52 and the relationship between visual function,
00:18:57 reading and auditory and semantic.
00:19:00 There were these, and still are, these beautiful models
00:19:03 that came out of that kind of research
00:19:05 that really made you feel like you understood something
00:19:08 that you hadn’t understood before
00:19:10 about how language processing is organized in the brain.
00:19:15 But having said all that,
00:19:20 I think you are, I mean, I agree with you
00:19:25 that the cost of doing highly controlled experiments
00:19:30 is that you, by construction, miss out on the richness
00:19:36 and complexity of the real world.
00:19:39 One thing that, so I was drawn into science
00:19:42 by what in those days was called connectionism,
00:19:44 which is, of course, what we now call deep learning.
00:19:49 And at that point in history,
00:19:50 neural networks were primarily being used
00:19:54 in order to model human cognition.
00:19:56 They weren’t yet really useful for industrial applications.
00:20:00 So you always found neural networks
00:20:02 in biological form beautiful.
00:20:04 Oh, neural networks were very concretely the thing
00:20:07 that drew me into science.
00:20:09 I was handed, are you familiar with the PDP books
00:20:13 from the 80s when I was in,
00:20:15 I went to medical school before I went into science.
00:20:18 And, yeah.
00:20:19 Really, interesting.
00:20:20 Wow.
00:20:21 I also did a graduate degree in art history,
00:20:23 so I’m kind of exploring.
00:20:26 Well, art history, I understand.
00:20:28 That’s just a curious, creative mind.
00:20:31 But medical school, with the dream of what,
00:20:33 if we take that slight tangent?
00:20:36 What, did you want to be a surgeon?
00:20:39 I actually was quite interested in surgery.
00:20:41 I was interested in surgery and psychiatry.
00:20:44 And I thought, I must be the only person on the planet
00:20:49 who was torn between those two fields.
00:20:52 And I said exactly that to my advisor in medical school,
00:20:56 who turned out, I found out later,
00:20:59 to be a famous psychoanalyst.
00:21:01 And he said to me, no, no, it’s actually not so uncommon
00:21:05 to be interested in surgery and psychiatry.
00:21:07 And he conjectured that the reason
00:21:10 that people develop these two interests
00:21:12 is that both fields are about going beneath the surface
00:21:16 and kind of getting into the kind of secret.
00:21:19 I mean, maybe you understand this as someone
00:21:21 who was interested in psychoanalysis.
00:21:23 There’s sort of a, there’s a cliche phrase
00:21:26 that people use now, like in NPR,
00:21:28 the secret life of blankety blank, right?
00:21:31 And that was part of the thrill of surgery,
00:21:33 was seeing the secret activity
00:21:38 that’s inside everybody’s abdomen and thorax.
00:21:40 That’s a very poetic way to connect it to disciplines
00:21:43 that are very, practically speaking,
00:21:45 different from each other.
00:21:46 That’s for sure, that’s for sure, yes.
00:21:48 So how did we get onto medical school?
00:21:52 So I was in medical school
00:21:53 and I was doing a psychiatry rotation
00:21:57 and my kind of advisor in that rotation
00:22:02 asked me what I was interested in.
00:22:04 And I said, well, maybe psychiatry.
00:22:07 He said, why?
00:22:09 And I said, well, I’ve always been interested
00:22:11 in how the brain works.
00:22:13 I’m pretty sure that nobody’s doing scientific research
00:22:16 that addresses my interests,
00:22:19 which are, I didn’t have a word for it then,
00:22:21 but I would have said about cognition.
00:22:25 And he said, well, you know, I’m not sure that’s true.
00:22:27 You might be interested in these books.
00:22:29 And he pulled down the PDB books from his shelf
00:22:32 and they were still shrink wrapped.
00:22:33 He hadn’t read them, but he handed them to me.
00:22:36 He said, you feel free to borrow these.
00:22:38 And that was, you know, I went back to my dorm room
00:22:41 and I just, you know, read them cover to cover.
00:22:43 And what’s PDB?
00:22:44 Parallel distributed processing,
00:22:46 which was one of the original names for deep learning.
00:22:50 And so I apologize for the romanticized question,
00:22:55 but what idea in the space of neuroscience
00:22:58 and the space of the human brain is to you
00:23:00 the most beautiful, mysterious, surprising?
00:23:03 What had always fascinated me,
00:23:08 even when I was a pretty young kid, I think,
00:23:12 was the paradox that lies in the fact
00:23:21 that the brain is so mysterious
00:23:25 and seems so distant.
00:23:30 But at the same time,
00:23:32 it’s responsible for the full transparency
00:23:37 of everyday life.
00:23:39 The brain is literally what makes everything obvious
00:23:41 and familiar.
00:23:43 And there’s always one in the room with you.
00:23:47 Yeah.
00:23:48 I used to teach, when I taught at Princeton,
00:23:50 I used to teach a cognitive neuroscience course.
00:23:53 And the very last thing I would say to the students was,
00:23:56 you know, people often,
00:24:00 when people think of scientific inspiration,
00:24:04 the metaphor is often, well, look to the stars.
00:24:08 The stars will inspire you to wonder at the universe
00:24:12 and think about your place in it and how things work.
00:24:15 And I’m all for looking at the stars,
00:24:18 but I’ve always been much more inspired.
00:24:21 And my sense of wonder comes from the,
00:24:25 not from the distant, mysterious stars,
00:24:28 but from the extremely intimately close brain.
00:24:34 Yeah.
00:24:35 There’s something just endlessly fascinating
00:24:38 to me about that.
00:24:40 The, like, just like you said,
00:24:41 the one that’s close and yet distant
00:24:45 in terms of our understanding of it.
00:24:48 Do you, are you also captivated by the fact
00:24:53 that this very conversation is happening
00:24:56 because two brains are communicating so that?
00:24:57 Yes, exactly.
00:24:59 The, I guess what I mean is the subjective nature
00:25:03 of the experience, if it can take a small attention
00:25:06 into the mystical of it, the consciousness,
00:25:10 or when you were saying you’re captivated
00:25:13 by the idea of the brain,
00:25:14 are you talking about specifically
00:25:16 the mechanism of cognition?
00:25:18 Or are you also just, like, at least for me,
00:25:23 it’s almost like paralyzing the beauty and the mystery
00:25:26 of the fact that it creates the entirety of the experience,
00:25:29 not just the reasoning capability, but the experience.
00:25:32 Well, I definitely resonate with that latter thought.
00:25:38 And I often find discussions of artificial intelligence
00:25:45 to be disappointingly narrow.
00:25:50 Speaking as someone who has always had an interest in art.
00:25:55 Right.
00:25:56 I was just gonna go there
00:25:57 because it sounds like somebody who has an interest in art.
00:26:00 Yeah, I mean, there are many layers
00:26:04 to full bore human experience.
00:26:08 And in some ways it’s not enough to say,
00:26:12 oh, well, don’t worry, we’re talking about cognition,
00:26:15 but we’ll add emotion, you know?
00:26:17 There’s an incredible scope
00:26:21 to what humans go through in every moment.
00:26:25 And yes, so that’s part of what fascinates me,
00:26:33 is that our brains are producing that.
00:26:40 But at the same time, it’s so mysterious to us.
00:26:43 How?
00:26:46 Our brains are literally in our heads
00:26:49 producing this experience.
00:26:50 Producing the experience.
00:26:52 And yet it’s so mysterious to us.
00:26:55 And so, and the scientific challenge
00:26:57 of getting at the actual explanation for that
00:27:00 is so overwhelming.
00:27:03 That’s just, I don’t know.
00:27:05 Certain people have fixations on particular questions
00:27:08 and that’s always, that’s just always been mine.
00:27:11 Yeah, I would say the poetry of that is fascinating.
00:27:14 And I’m really interested in natural language as well.
00:27:16 And when you look at artificial intelligence community,
00:27:19 it always saddens me how much
00:27:23 when you try to create a benchmark
00:27:25 for the community to gather around,
00:27:28 how much of the magic of language is lost
00:27:30 when you create that benchmark.
00:27:33 That there’s something, we talk about experience,
00:27:35 the music of the language, the wit,
00:27:38 the something that makes a rich experience,
00:27:41 something that would be required to pass
00:27:43 the spirit of the Turing test is lost in these benchmarks.
00:27:47 And I wonder how to get it back in
00:27:50 because it’s very difficult.
00:27:51 The moment you try to do like real good rigorous science,
00:27:55 you lose some of that magic.
00:27:56 When you try to study cognition
00:28:00 in a rigorous scientific way,
00:28:01 it feels like you’re losing some of the magic.
00:28:03 The seeing cognition in a mechanistic way
00:28:07 that AI folk at this stage in our history.
00:28:10 Well, I agree with you, but at the same time,
00:28:13 one thing that I found really exciting
00:28:18 about that first wave of deep learning models in cognition
00:28:22 was the fact that the people who were building these models
00:28:29 were focused on the richness and complexity
00:28:32 of human cognition.
00:28:34 So an early debate in cognitive science,
00:28:40 which I sort of witnessed as a grad student
00:28:41 was about something that sounds very dry,
00:28:44 which is the formation of the past tense.
00:28:47 But there were these two camps.
00:28:49 One said, well, the mind encodes certain rules
00:28:54 and it also has a list of exceptions
00:28:57 because of course, the rule is add ED,
00:29:00 but that’s not always what you do.
00:29:01 So you have to have a list of exceptions.
00:29:05 And then there were the connectionists
00:29:06 who evolved into the deep learning people who said,
00:29:10 well, if you look carefully at the data,
00:29:13 if you actually look at corpora, like language corpora,
00:29:18 it turns out to be very rich
00:29:20 because yes, there are most verbs
00:29:25 that you just tack on ED, and then there are exceptions,
00:29:28 but there are rules that the exceptions aren’t just random.
00:29:36 There are certain clues to which verbs
00:29:39 should be exceptional.
00:29:41 And then there are exceptions to the exceptions.
00:29:44 And there was a word that was kind of deployed
00:29:47 in order to capture this, which was quasi regular.
00:29:51 In other words, there are rules, but it’s messy.
00:29:54 And there’s either structure even among the exceptions.
00:29:58 And it would be, yeah, you could try to write down,
00:30:01 we could try to write down the structure
00:30:03 in some sort of closed form,
00:30:04 but really the right way to understand
00:30:07 how the brain is handling all this,
00:30:09 and by the way, producing all of this,
00:30:11 is to build a deep neural network
00:30:14 and train it on this data
00:30:15 and see how it ends up representing all of this richness.
00:30:18 So the way that deep learning
00:30:21 was deployed in cognitive psychology
00:30:23 was that was the spirit of it.
00:30:25 It was about that richness.
00:30:29 And that’s something that I always found very compelling,
00:30:31 still do.
00:30:33 Is there something especially interesting
00:30:36 and profound to you
00:30:37 in terms of our current deep learning neural network,
00:30:40 artificial neural network approaches,
00:30:42 and whatever we do understand
00:30:46 about the biological neural networks in our brain?
00:30:49 Is there, there’s quite a few differences.
00:30:52 Are some of them to you,
00:30:54 either interesting or perhaps profound
00:30:58 in terms of the gap we might want to try to close
00:31:03 in trying to create a human level intelligence?
00:31:07 What I would say here is something
00:31:08 that a lot of people are saying,
00:31:10 which is that one seeming limitation
00:31:16 of the systems that we’re building now
00:31:18 is that they lack the kind of flexibility,
00:31:22 the readiness to sort of turn on a dime
00:31:25 when the context calls for it
00:31:28 that is so characteristic of human behavior.
00:31:32 So is that connected to you to the,
00:31:34 like which aspect of the neural networks in our brain
00:31:37 is that connected to?
00:31:39 Is that closer to the cognitive science level of,
00:31:45 now again, see like my natural inclination
00:31:47 is to separate into three disciplines of neuroscience,
00:31:51 cognitive science and psychology.
00:31:54 And you’ve already kind of shut that down
00:31:56 by saying you’re kind of see them as separate,
00:31:58 but just to look at those layers,
00:32:01 I guess where is there something about the lowest layer
00:32:05 of the way the neural neurons interact
00:32:09 that is profound to you in terms of this difference
00:32:13 to the artificial neural networks,
00:32:15 or is all the key differences
00:32:17 at a higher level of abstraction?
00:32:20 One thing I often think about is that,
00:32:24 if you take an introductory computer science course
00:32:27 and they are introducing you to the notion
00:32:29 of Turing machines,
00:32:31 one way of articulating
00:32:36 what the significance of a Turing machine is,
00:32:39 is that it’s a machine emulator.
00:32:42 It can emulate any other machine.
00:32:47 And that to me,
00:32:52 that way of looking at a Turing machine
00:32:56 really sticks with me.
00:32:57 I think of humans as maybe sharing
00:33:01 in some of that character.
00:33:05 We’re capacity limited,
00:33:06 we’re not Turing machines obviously,
00:33:07 but we have the ability to adapt behaviors
00:33:11 that are very much unlike anything we’ve done before,
00:33:15 but there’s some basic mechanism
00:33:17 that’s implemented in our brain
00:33:18 that allows us to run software.
00:33:22 But just on that point, you mentioned Turing machine,
00:33:24 but nevertheless, it’s fundamentally
00:33:26 our brains are just computational devices in your view.
00:33:29 Is that what you’re getting at?
00:33:31 It was a little bit unclear to this line you drew.
00:33:35 Is there any magic in there
00:33:37 or is it just basic computation?
00:33:40 I’m happy to think of it as just basic computation,
00:33:43 but mind you, I won’t be satisfied
00:33:46 until somebody explains to me
00:33:48 what the basic computations are
00:33:49 that are leading to the full richness of human cognition.
00:33:54 It’s not gonna be enough for me
00:33:56 to understand what the computations are
00:33:58 that allow people to do arithmetic or play chess.
00:34:02 I want the whole thing.
00:34:06 And a small tangent,
00:34:07 because you kind of mentioned coronavirus,
00:34:10 there’s group behavior.
00:34:12 Oh, sure.
00:34:13 Is there something interesting
00:34:14 to your search of understanding the human mind
00:34:18 where behavior of large groups
00:34:21 or just behavior of groups is interesting,
00:34:24 seeing that as a collective mind,
00:34:25 as a collective intelligence,
00:34:27 perhaps seeing the groups of people
00:34:28 as a single intelligent organisms,
00:34:31 especially looking at the reinforcement learning work
00:34:34 you’ve done recently.
00:34:35 Well, yeah, I can’t.
00:34:36 I mean, I have the honor of working
00:34:41 with a lot of incredibly smart people
00:34:43 and I wouldn’t wanna take any credit
00:34:45 for leading the way on the multiagent work
00:34:48 that’s come out of my group or DeepMind lately,
00:34:51 but I do find it fascinating.
00:34:53 And I mean, I think it can’t be debated.
00:35:00 You know, human behavior arises within communities.
00:35:06 That just seems to me self evident.
00:35:08 But to me, it is self evident,
00:35:11 but that seems to be a profound aspects
00:35:14 of something that created.
00:35:16 That was like, if you look at like 2001 Space Odyssey
00:35:19 when the monkeys touched the…
00:35:21 Yeah.
00:35:22 That’s the magical moment I think Yuval Harari argues
00:35:25 that the ability of our large numbers of humans
00:35:29 to hold an idea, to converge towards idea together,
00:35:31 like you said, shaking hands versus bumping elbows,
00:35:34 somehow converge without being in a room altogether,
00:35:40 just kind of this like distributed convergence
00:35:43 towards an idea over a particular period of time
00:35:46 seems to be fundamental to just every aspect
00:35:51 of our cognition, of our intelligence,
00:35:53 because humans, I will talk about reward,
00:35:56 but it seems like we don’t really have
00:35:58 a clear objective function under which we operate,
00:36:01 but we all kind of converge towards one somehow.
00:36:04 And that to me has always been a mystery
00:36:07 that I think is somehow productive
00:36:09 for also understanding AI systems.
00:36:13 But I guess that’s the next step.
00:36:16 The first step is try to understand the mind.
00:36:18 Well, I don’t know.
00:36:19 I mean, I think there’s something to the argument
00:36:22 that that kind of like strictly bottom up approach
00:36:27 is wrongheaded.
00:36:29 In other words, there are basic phenomena,
00:36:34 basic aspects of human intelligence
00:36:36 that can only be understood in the context of groups.
00:36:43 I’m perfectly open to that.
00:36:44 I’ve never been particularly convinced by the notion
00:36:48 that we should consider intelligence
00:36:52 to inhere at the level of communities.
00:36:55 I don’t know why, I’m sort of stuck on the notion
00:36:58 that the basic unit that we want to understand
00:37:01 is individual humans.
00:37:02 And if we have to understand that
00:37:05 in the context of other humans, fine.
00:37:08 But for me, intelligence is just,
00:37:11 I stubbornly define it as something
00:37:14 that is an aspect of an individual human.
00:37:18 That’s just my, I don’t know if that’s a matter of taste.
00:37:20 I’m with you, but that could be the reductionist dream
00:37:22 of a scientist because you can understand a single human.
00:37:26 It also is very possible that intelligence can only arise
00:37:30 when there’s multiple intelligences.
00:37:33 When there’s multiple sort of, it’s a sad thing,
00:37:37 if that’s true, because it’s very difficult to study.
00:37:39 But if it’s just one human,
00:37:42 that one human would not be homosapien,
00:37:44 would not become that intelligent.
00:37:46 That’s a possibility.
00:37:48 I’m with you.
00:37:50 One thing I will say along these lines
00:37:52 is that I think a serious effort
00:38:01 to understand human intelligence
00:38:05 and maybe to build humanlike intelligence
00:38:09 needs to pay just as much attention
00:38:11 to the structure of the environment
00:38:14 as to the structure of the cognizing system,
00:38:20 whether it’s a brain or an AI system.
00:38:23 That’s one thing I took away actually
00:38:24 from my early studies with the pioneers
00:38:27 of neural network research,
00:38:29 people like Jay McClelland and John Cohen.
00:38:34 The structure of cognition is really,
00:38:38 it’s only partly a function of the architecture of the brain
00:38:44 and the learning algorithms that it implements.
00:38:46 What really shapes it is the interaction of those things
00:38:51 with the structure of the world
00:38:54 in which those things are embedded.
00:38:56 And that’s especially important for,
00:38:58 that’s made most clear in reinforcement learning
00:39:00 where the simulated environment is,
00:39:03 you can only learn as much as you can simulate.
00:39:05 And that’s what DeepMind made very clear
00:39:09 with the other aspect of the environment,
00:39:11 which is the self play mechanism of the other agent,
00:39:15 of the competitive behavior,
00:39:16 which the other agent becomes the environment essentially.
00:39:20 And that’s, I mean, one of the most exciting ideas in AI
00:39:24 is the self play mechanism that’s able to learn successfully.
00:39:27 So there you go.
00:39:28 There’s a thing where competition is essential
00:39:31 for learning, at least in that context.
00:39:35 So if we can step back into another sort of beautiful world,
00:39:37 which is the actual mechanics,
00:39:42 the dirty mess of it of the human brain,
00:39:44 is there something for people who might not know?
00:39:49 Is there something you can comment on
00:39:51 or describe the key parts of the brain
00:39:53 that are important for intelligence or just in general,
00:39:56 what are the different parts of the brain
00:39:58 that you’re curious about that you’ve studied
00:40:01 and that are just good to know about
00:40:03 when you’re thinking about cognition?
00:40:06 Well, my area of expertise, if I have one,
00:40:11 is prefrontal cortex.
00:40:14 So, you know. What’s that?
00:40:16 Where do we?
00:40:18 It depends on who you ask.
00:40:21 The technical definition is anatomical.
00:40:25 There are parts of your brain
00:40:30 that are responsible for motor behavior
00:40:32 and they’re very easy to identify.
00:40:35 And the region of your cerebral cortex,
00:40:40 the sort of outer crust of your brain
00:40:43 that lies in front of those
00:40:46 is defined as the prefrontal cortex.
00:40:49 And when you say anatomical, sorry to interrupt,
00:40:51 so that’s referring to sort of the geographic region
00:40:57 as opposed to some kind of functional definition.
00:41:00 Exactly, so this is kind of the coward’s way out.
00:41:04 I’m telling you what the prefrontal cortex is
00:41:06 just in terms of what part of the real estate it occupies.
00:41:09 It’s the thing in the front of the brain.
00:41:10 Yeah, exactly.
00:41:11 And in fact, the early history
00:41:14 of neuroscientific investigation
00:41:20 of what this front part of the brain does
00:41:23 is sort of funny to read
00:41:25 because it was really World War I
00:41:32 that started people down this road
00:41:34 of trying to figure out what different parts of the brain,
00:41:37 the human brain do in the sense
00:41:39 that there were a lot of people with brain damage
00:41:42 who came back from the war with brain damage.
00:41:44 And that provided, as tragic as that was,
00:41:47 it provided an opportunity for scientists
00:41:49 to try to identify the functions of different brain regions.
00:41:53 And that was actually incredibly productive,
00:41:56 but one of the frustrations that neuropsychologists faced
00:41:59 was they couldn’t really identify exactly
00:42:02 what the deficit was that arose from damage
00:42:05 to these most kind of frontal parts of the brain.
00:42:08 It was just a very difficult thing to pin down.
00:42:13 There were a couple of neuropsychologists
00:42:16 who identified through a large amount
00:42:20 of clinical experience and close observation,
00:42:23 they started to put their finger on a syndrome
00:42:26 that was associated with frontal damage.
00:42:27 Actually, one of them was a Russian neuropsychologist
00:42:30 named Luria, who students of cognitive psychology still read.
00:42:36 And what he started to figure out was that
00:42:41 the frontal cortex was somehow involved in flexibility,
00:42:48 in guiding behaviors that required someone
00:42:52 to override a habit, or to do something unusual,
00:42:57 or to change what they were doing in a very flexible way
00:43:01 from one moment to another.
00:43:02 So focused on like new experiences.
00:43:05 And so the way your brain processes
00:43:08 and acts in new experiences.
00:43:10 Yeah, what later helped bring this function
00:43:14 into better focus was a distinction
00:43:17 between controlled and automatic behavior,
00:43:19 or in other literatures, this is referred to
00:43:23 as habitual behavior versus goal directed behavior.
00:43:28 So it’s very, very clear that the human brain
00:43:33 has pathways that are dedicated to habits,
00:43:36 to things that you do all the time,
00:43:39 and they need to be automatized
00:43:42 so that they don’t require you to concentrate too much.
00:43:45 So that leaves your cognitive capacity
00:43:47 free to do other things.
00:43:49 Just think about the difference
00:43:51 between driving when you’re learning to drive
00:43:55 versus driving after you’re a fairly expert.
00:43:59 There are brain pathways that slowly absorb
00:44:03 those frequently performed behaviors
00:44:07 so that they can be habits, so that they can be automatic.
00:44:12 That’s kind of like the purest form of learning.
00:44:14 I guess it’s happening there, which is why,
00:44:18 I mean, this is kind of jumping ahead,
00:44:20 which is why that perhaps is the most useful for us
00:44:22 to focusing on and trying to see
00:44:24 how artificial intelligence systems can learn.
00:44:27 Is that the way you think?
00:44:28 It’s interesting.
00:44:29 I do think about this distinction
00:44:30 between controlled and automatic,
00:44:31 or goal directed and habitual behavior a lot
00:44:34 in thinking about where we are in AI research.
00:44:42 But just to finish the kind of dissertation here,
00:44:46 the role of the prefrontal cortex
00:44:51 is generally understood these days
00:44:54 sort of in contradistinction to that habitual domain.
00:45:00 In other words, the prefrontal cortex
00:45:02 is what helps you override those habits.
00:45:05 It’s what allows you to say,
00:45:07 well, what I usually do in this situation is X,
00:45:10 but given the context, I probably should do Y.
00:45:14 I mean, the elbow bump is a great example, right?
00:45:18 Reaching out and shaking hands
00:45:19 is probably a habitual behavior,
00:45:22 and it’s the prefrontal cortex that allows us
00:45:26 to bear in mind that there’s something unusual
00:45:28 going on right now, and in this situation,
00:45:31 I need to not do the usual thing.
00:45:34 The kind of behaviors that Luria reported,
00:45:38 and he built tests for detecting these kinds of things,
00:45:42 were exactly like this.
00:45:43 So in other words, when I stick out my hand,
00:45:47 I want you instead to present your elbow.
00:45:49 A patient with frontal damage
00:45:51 would have a great deal of trouble with that.
00:45:53 Somebody proffering their hand would elicit a handshake.
00:45:58 The prefrontal cortex is what allows us to say,
00:46:00 hold on, hold on, that’s the usual thing,
00:46:03 but I have the ability to bear in mind
00:46:07 even very unusual contexts and to reason about
00:46:10 what behavior is appropriate there.
00:46:13 Just to get a sense, are us humans special
00:46:17 in the presence of the prefrontal cortex?
00:46:20 Do mice have a prefrontal cortex?
00:46:22 Do other mammals that we can study?
00:46:25 If no, then how do they integrate new experiences?
00:46:30 Yeah, that’s a really tricky question
00:46:33 and a very timely question
00:46:35 because we have revolutionary new technologies
00:46:44 for monitoring, measuring,
00:46:48 and also causally influencing neural behavior
00:46:52 in mice and fruit flies.
00:46:57 And these techniques are not fully available
00:47:00 even for studying brain function in monkeys,
00:47:06 let alone humans.
00:47:08 And so it’s a very sort of, for me at least,
00:47:12 a very urgent question whether the kinds of things
00:47:16 that we wanna understand about human intelligence
00:47:18 can be pursued in these other organisms.
00:47:22 And to put it briefly, there’s disagreement.
00:47:26 People who study fruit flies will often tell you,
00:47:32 hey, fruit flies are smarter than you think.
00:47:35 And they’ll point to experiments where fruit flies
00:47:37 were able to learn new behaviors,
00:47:40 were able to generalize from one stimulus to another
00:47:44 in a way that suggests that they have abstractions
00:47:47 that guide their generalization.
00:47:51 I’ve had many conversations in which
00:47:53 I will start by observing,
00:47:58 recounting some observation about mouse behavior
00:48:05 where it seemed like mice were taking an awfully long time
00:48:09 to learn a task that for a human would be profoundly trivial.
00:48:13 And I will conclude from that,
00:48:16 that mice really don’t have the cognitive flexibility
00:48:18 that we want to explain.
00:48:20 And then a mouse researcher will say to me,
00:48:21 well, hold on, that experiment may not have worked
00:48:26 because you asked a mouse to deal with stimuli
00:48:31 and behaviors that were very unnatural for the mouse.
00:48:34 If instead you kept the logic of the experiment the same,
00:48:38 but presented the information in a way
00:48:44 that aligns with what mice are used to dealing with
00:48:46 in their natural habitats,
00:48:48 you might find that a mouse actually has more intelligence
00:48:51 than you think.
00:48:52 And then they’ll go on to show you videos
00:48:54 of mice doing things in their natural habitat,
00:48:57 which seem strikingly intelligent,
00:49:00 dealing with physical problems.
00:49:02 I have to drag this piece of food back to my lair,
00:49:07 but there’s something in my way
00:49:08 and how do I get rid of that thing?
00:49:10 So I think these are open questions
00:49:13 to put it, to sum that up.
00:49:15 And then taking a small step back related to that
00:49:18 is you kind of mentioned we’re taking a little shortcut
00:49:21 by saying it’s a geographic part of the prefrontal cortex
00:49:26 is a region of the brain.
00:49:28 But if we, what’s your sense in a bigger philosophical view,
00:49:33 prefrontal cortex and the brain in general,
00:49:36 do you have a sense that it’s a set of subsystems
00:49:38 in the way we’ve kind of implied
00:49:41 that are pretty distinct or to what degree is it that
00:49:46 or to what degree is it a giant interconnected mess
00:49:49 where everything kind of does everything
00:49:51 and it’s impossible to disentangle them?
00:49:54 I think there’s overwhelming evidence
00:49:57 that there’s functional differentiation,
00:50:00 that it’s clearly not the case
00:50:03 that all parts of the brain are doing the same thing.
00:50:07 This follows immediately from the kinds of studies
00:50:11 of brain damage that we were chatting about before.
00:50:14 It’s obvious from what you see
00:50:18 if you stick an electrode in the brain
00:50:19 and measure what’s going on at the level of neural activity.
00:50:25 Having said that, there are two other things to add,
00:50:30 which kind of, I don’t know,
00:50:32 maybe tug in the other direction.
00:50:34 One is that it’s when you look carefully
00:50:39 at functional differentiation in the brain,
00:50:42 what you usually end up concluding,
00:50:44 at least this is my observation of the literature,
00:50:48 is that the differences between regions are graded
00:50:52 rather than being discreet.
00:50:55 So it doesn’t seem like it’s easy
00:50:57 to divide the brain up into true modules
00:51:03 that have clear boundaries and that have
00:51:07 you know, clear channels of communication between them.
00:51:16 And this applies to the prefrontal cortex?
00:51:18 Yeah, oh yeah.
00:51:18 The prefrontal cortex is made up
00:51:20 of a bunch of different subregions,
00:51:23 the functions of which are not clearly defined
00:51:27 and the borders of which seem to be quite vague.
00:51:32 And then there’s another thing that’s popping up
00:51:34 in very recent research, which, you know, which,
00:51:40 involves application of these new techniques,
00:51:44 which there are a number of studies that suggest that
00:51:48 parts of the brain that we would have previously thought
00:51:51 were quite focused in their function
00:51:57 are actually carrying signals
00:51:59 that we wouldn’t have thought would be there.
00:52:01 For example, looking in the primary visual cortex,
00:52:04 which is classically thought of as basically
00:52:07 the first cortical way station
00:52:09 for processing visual information.
00:52:10 Basically what it should care about is, you know,
00:52:12 where are the edges in this scene that I’m viewing?
00:52:17 It turns out that if you have enough data,
00:52:19 you can recover information from primary visual cortex
00:52:22 about all sorts of things.
00:52:23 Like, you know, what behavior the animal is engaged
00:52:26 in right now and how much reward is on offer
00:52:29 in the task that it’s pursuing.
00:52:31 So it’s clear that even regions whose function
00:52:36 is pretty well defined at a core screen
00:52:40 are nonetheless carrying some information
00:52:42 about information from very different domains.
00:52:47 So, you know, the history of neuroscience
00:52:49 is sort of this oscillation between the two views
00:52:52 that you articulated, you know, the kind of modular view
00:52:55 and then the big, you know, mush view.
00:52:57 And, you know, I think, I guess we’re gonna end up
00:53:01 somewhere in the middle.
00:53:02 Which is unfortunate for our understanding
00:53:05 because there’s something about our, you know,
00:53:08 conceptual system that finds it’s easy to think about
00:53:11 a modularized system and easy to think about
00:53:13 a completely undifferentiated system.
00:53:15 But something that kind of lies in between is confusing.
00:53:19 But we’re gonna have to get used to it, I think.
00:53:21 Unless we can understand deeply the lower level mechanism
00:53:24 of neuronal communication.
00:53:25 Yeah, yeah.
00:53:26 But on that topic, you kind of mentioned information.
00:53:29 Just to get a sense, I imagine something
00:53:31 that there’s still mystery and disagreement on
00:53:34 is how does the brain carry information and signal?
00:53:38 Like what in your sense is the basic mechanism
00:53:43 of communication in the brain?
00:53:46 Well, I guess I’m old fashioned in that I consider
00:53:52 the networks that we use in deep learning research
00:53:54 to be a reasonable approximation to, you know,
00:53:59 the mechanisms that carry information in the brain.
00:54:02 So the usual way of articulating that is to say,
00:54:06 what really matters is a rate code.
00:54:08 What matters is how quickly is an individual neuron spiking?
00:54:14 You know, what’s the frequency at which it’s spiking?
00:54:16 Is it right?
00:54:17 So the timing of the spike.
00:54:18 Yeah, is it firing fast or slow?
00:54:20 Let’s, you know, let’s put a number on that.
00:54:22 And that number is enough to capture
00:54:24 what neurons are doing.
00:54:26 There’s, you know, there’s still uncertainty
00:54:30 about whether that’s an adequate description
00:54:34 of how information is transmitted within the brain.
00:54:39 There, you know, there are studies that suggest
00:54:42 that the precise timing of spikes matters.
00:54:46 There are studies that suggest that there are computations
00:54:50 that go on within the dendritic tree, within a neuron,
00:54:54 that are quite rich and structured
00:54:57 and that really don’t equate to anything that we’re doing
00:54:59 in our artificial neural networks.
00:55:02 Having said that, I feel like we can get,
00:55:05 I feel like we’re getting somewhere
00:55:08 by sticking to this high level of abstraction.
00:55:11 Just the rate, and by the way,
00:55:13 we’re talking about the electrical signal.
00:55:16 I remember reading some vague paper somewhere recently
00:55:20 where the mechanical signal, like the vibrations
00:55:23 or something of the neurons, also communicates information.
00:55:28 I haven’t seen that, but.
00:55:30 There’s somebody who was arguing
00:55:32 that the electrical signal, this is in a nature paper,
00:55:36 something like that, where the electrical signal
00:55:38 is actually a side effect of the mechanical signal.
00:55:43 But I don’t think that changes the story.
00:55:46 But it’s almost an interesting idea
00:55:49 that there could be a deeper, it’s always like in physics
00:55:52 with quantum mechanics, there’s always a deeper story
00:55:55 that could be underlying the whole thing.
00:55:57 But you think it’s basically the rate of spiking
00:56:00 that gets us, that’s like the lowest hanging fruit
00:56:02 that can get us really far.
00:56:04 This is a classical view.
00:56:06 I mean, this is not, the only way in which this stance
00:56:10 would be controversial is in the sense
00:56:13 that there are members of the neuroscience community
00:56:17 who are interested in alternatives.
00:56:18 But this is really a very mainstream view.
00:56:21 The way that neurons communicate
00:56:22 is that neurotransmitters arrive,
00:56:30 they wash up on a neuron, the neuron has receptors
00:56:34 for those transmitters, the meeting of the transmitter
00:56:39 with these receptors changes the voltage of the neuron.
00:56:42 And if enough voltage change occurs, then a spike occurs,
00:56:46 one of these like discrete events.
00:56:48 And it’s that spike that is conducted down the axon
00:56:52 and leads to neurotransmitter release.
00:56:54 This is just like neuroscience 101.
00:56:56 This is like the way the brain is supposed to work.
00:56:59 Now, what we do when we build artificial neural networks
00:57:03 of the kind that are now popular in the AI community
00:57:08 is that we don’t worry about those individual spikes.
00:57:11 We just worry about the frequency
00:57:14 at which those spikes are being generated.
00:57:16 And people talk about that as the activity of a neuron.
00:57:22 And so the activity of units in a deep learning system
00:57:27 is broadly analogous to the spike rate of a neuron.
00:57:32 There are people who believe that there are other forms
00:57:38 of communication in the brain.
00:57:39 In fact, I’ve been involved in some research recently
00:57:41 that suggests that the voltage fluctuations
00:57:46 that occur in populations of neurons
00:57:49 that are sort of below the level of spike production
00:57:54 may be important for communication.
00:57:57 But I’m still pretty old school in the sense
00:58:00 that I think that the things that we’re building
00:58:02 in AI research constitute reasonable models
00:58:06 of how a brain would work.
00:58:10 Let me ask just for fun a crazy question, because I can.
00:58:14 Do you think it’s possible we’re completely wrong
00:58:17 about the way this basic mechanism
00:58:20 of neuronal communication, that the information
00:58:23 is stored in some very different kind of way in the brain?
00:58:26 Oh, heck yes.
00:58:27 I mean, look, I wouldn’t be a scientist
00:58:29 if I didn’t think there was any chance we were wrong.
00:58:32 But I mean, if you look at the history
00:58:36 of deep learning research as it’s been applied
00:58:39 to neuroscience, of course the vast majority
00:58:42 of deep learning research these days isn’t about neuroscience.
00:58:45 But if you go back to the 1980s,
00:58:49 there’s sort of an unbroken chain of research
00:58:52 in which a particular strategy is taken,
00:58:54 which is, hey, let’s train a deep learning system.
00:59:00 Let’s train a multi layer neural network
00:59:04 on this task that we trained our rat on,
00:59:09 or our monkey on, or this human being on.
00:59:12 And then let’s look at what the units
00:59:15 deep in the system are doing.
00:59:17 And let’s ask whether what they’re doing
00:59:20 resembles what we know about what neurons
00:59:23 deep in the brain are doing.
00:59:24 And over and over and over and over,
00:59:28 that strategy works in the sense that
00:59:32 the learning algorithms that we have access to,
00:59:34 which typically center on back propagation,
00:59:37 they give rise to patterns of activity,
00:59:42 patterns of response,
00:59:45 patterns of neuronal behavior in these artificial models
00:59:48 that look hauntingly similar to what you see in the brain.
00:59:53 And is that a coincidence?
00:59:57 At a certain point, it starts looking like such coincidence
01:00:00 is unlikely to not be deeply meaningful, yeah.
01:00:03 Yeah, the circumstantial evidence is overwhelming.
01:00:07 But it could be.
01:00:07 But you’re always open to total flipping at the table.
01:00:10 Hey, of course.
01:00:11 So you have coauthored several recent papers
01:00:15 that sort of weave beautifully between the world
01:00:17 of neuroscience and artificial intelligence.
01:00:20 And maybe if we could, can we just try to dance around
01:00:26 and talk about some of them?
01:00:27 Maybe try to pick out interesting ideas
01:00:29 that jump to your mind from memory.
01:00:32 So maybe looking at, we were talking about
01:00:34 the prefrontal cortex, the 2018, I believe, paper
01:00:38 called the Prefrontal Cortex
01:00:40 as a Meta Reinforcement Learning System.
01:00:42 What, is there a key idea
01:00:44 that you can speak to from that paper?
01:00:47 Yeah, I mean, the key idea is about meta learning.
01:00:53 What is meta learning?
01:00:54 Meta learning is, by definition,
01:01:00 a situation in which you have a learning algorithm
01:01:06 and the learning algorithm operates in such a way
01:01:09 that it gives rise to another learning algorithm.
01:01:14 In the earliest applications of this idea,
01:01:17 you had one learning algorithm sort of adjusting
01:01:20 the parameters on another learning algorithm.
01:01:23 But the case that we’re interested in this paper
01:01:25 is one where you start with just one learning algorithm
01:01:29 and then another learning algorithm kind of emerges
01:01:33 out of thin air.
01:01:35 I can say more about what I mean by that.
01:01:36 I don’t mean to be scurrentist,
01:01:39 but that’s the idea of meta learning.
01:01:44 It relates to the old idea in psychology
01:01:46 of learning to learn.
01:01:49 Situations where you have experiences
01:01:54 that make you better at learning something new.
01:01:57 A familiar example would be learning a foreign language.
01:02:01 The first time you learn a foreign language,
01:02:02 it may be quite laborious and disorienting
01:02:06 and novel, but let’s say you’ve learned
01:02:10 two foreign languages.
01:02:12 The third foreign language, obviously,
01:02:14 is gonna be much easier to pick up.
01:02:15 And why?
01:02:16 Because you’ve learned how to learn.
01:02:18 You know how this goes.
01:02:20 You know, okay, I’m gonna have to learn how to conjugate.
01:02:22 I’m gonna have to…
01:02:23 That’s a simple form of meta learning
01:02:26 in the sense that there’s some slow learning mechanism
01:02:30 that’s helping you kind of update
01:02:33 your fast learning mechanism.
01:02:34 Does that make sense?
01:02:35 So how from our understanding from the psychology world,
01:02:40 from neuroscience, our understanding
01:02:43 how meta learning might work in the human brain,
01:02:47 what lessons can we draw from that
01:02:49 that we can bring into the artificial intelligence world?
01:02:53 Well, yeah, so the origin of that paper
01:02:55 was in AI work that we were doing in my group.
01:03:00 We were looking at what happens
01:03:03 when you train a recurrent neural network
01:03:06 using standard reinforcement learning algorithms.
01:03:10 But you train that network, not just in one task,
01:03:12 but you train it in a bunch of interrelated tasks.
01:03:15 And then you ask what happens when you give it
01:03:18 yet another task in that sort of line of interrelated tasks.
01:03:23 And what we started to realize is that
01:03:29 a form of meta learning spontaneously happens
01:03:31 in recurrent neural networks.
01:03:33 And the simplest way to explain it is to say
01:03:39 a recurrent neural network has a kind of memory
01:03:43 in its activation patterns.
01:03:45 It’s recurrent by definition in the sense
01:03:47 that you have units that connect to other units,
01:03:50 that connect to other units.
01:03:51 So you have sort of loops of connectivity,
01:03:53 which allows activity to stick around
01:03:55 and be updated over time.
01:03:57 In psychology we call, in neuroscience
01:03:59 we call this working memory.
01:04:00 It’s like actively holding something in mind.
01:04:04 And so that memory gives
01:04:09 the recurrent neural network a dynamics, right?
01:04:13 The way that the activity pattern evolves over time
01:04:17 is inherent to the connectivity
01:04:19 of the recurrent neural network, okay?
01:04:21 So that’s idea number one.
01:04:23 Now, the dynamics of that network are shaped
01:04:26 by the connectivity, by the synaptic weights.
01:04:29 And those synaptic weights are being shaped
01:04:31 by this reinforcement learning algorithm
01:04:33 that you’re training the network with.
01:04:37 So the punchline is if you train a recurrent neural network
01:04:41 with a reinforcement learning algorithm
01:04:43 that’s adjusting its weights,
01:04:44 and you do that for long enough,
01:04:47 the activation dynamics will become very interesting, right?
01:04:50 So imagine I give you a task
01:04:53 where you have to press one button or another,
01:04:56 left button or right button.
01:04:57 And there’s some probability
01:05:00 that I’m gonna give you an M&M
01:05:02 if you press the left button,
01:05:04 and there’s some probability I’ll give you an M&M
01:05:06 if you press the other button.
01:05:07 And you have to figure out what those probabilities are
01:05:09 just by trying things out.
01:05:12 But as I said before,
01:05:13 instead of just giving you one of these tasks,
01:05:15 I give you a whole sequence.
01:05:17 You know, I give you two buttons
01:05:18 and you figure out which one’s best.
01:05:19 And I go, good job, here’s a new box.
01:05:22 Two new buttons, you have to figure out which one’s best.
01:05:24 Good job, here’s a new box.
01:05:25 And every box has its own probabilities
01:05:27 and you have to figure it out.
01:05:28 So if you train a recurrent neural network
01:05:30 on that kind of sequence of tasks,
01:05:33 what happens, it seemed almost magical to us
01:05:37 when we first started kind of realizing what was going on.
01:05:41 The slow learning algorithm that’s adjusting
01:05:43 the synaptic weights,
01:05:46 those slow synaptic changes give rise to a network dynamics
01:05:51 that themselves, that, you know,
01:05:53 the dynamics themselves turn into a learning algorithm.
01:05:56 So in other words, you can tell this is happening
01:05:59 by just freezing the synaptic weights saying,
01:06:01 okay, no more learning, you’re done.
01:06:03 Here’s a new box, figure out which button is best.
01:06:07 And the recurrent neural network will do this just fine.
01:06:09 There’s no, like it figures out which button is best.
01:06:13 It kind of transitions from exploring the two buttons
01:06:16 to just pressing the one that it likes best
01:06:18 in a very rational way.
01:06:20 How is that happening?
01:06:21 It’s happening because the activity dynamics
01:06:24 of the network have been shaped by the slow learning process
01:06:28 that’s occurred over many, many boxes.
01:06:30 And so what’s happened is that this slow learning algorithm
01:06:34 that’s slowly adjusting the weights
01:06:37 is changing the dynamics of the network,
01:06:39 the activity dynamics into its own learning algorithm.
01:06:43 And as we were kind of realizing that this is a thing,
01:06:51 it just so happened that the group that was working on this
01:06:53 included a bunch of neuroscientists
01:06:56 and it started kind of ringing a bell for us,
01:06:59 which is to say that we thought this sounds a lot
01:07:02 like the distinction between synaptic learning
01:07:06 and activity, synaptic memory
01:07:08 and activity based memory in the brain.
01:07:11 And it also reminded us of recurrent connectivity
01:07:15 that’s very characteristic of prefrontal function.
01:07:19 So this is kind of why it’s good to have people working
01:07:22 on AI that know a little bit about neuroscience
01:07:26 and vice versa, because we started thinking
01:07:29 about whether we could apply this principle to neuroscience.
01:07:32 And that’s where the paper came from.
01:07:33 So the kind of principle of the recurrence
01:07:37 they can see in the prefrontal cortex,
01:07:39 then you start to realize that it’s possible
01:07:43 for something like an idea of a learning
01:07:46 to learn emerging from this learning process
01:07:50 as long as you keep varying the environment sufficiently.
01:07:54 Exactly, so the kind of metaphorical transition
01:07:59 we made to neuroscience was to think,
01:08:00 okay, well, we know that the prefrontal cortex
01:08:03 is highly recurrent.
01:08:04 We know that it’s an important locus for working memory
01:08:08 for activation based memory.
01:08:11 So maybe the prefrontal cortex
01:08:13 supports reinforcement learning.
01:08:15 In other words, what is reinforcement learning?
01:08:19 You take an action, you see how much reward you got,
01:08:21 you update your policy of behavior.
01:08:24 Maybe the prefrontal cortex is doing that sort of thing
01:08:26 strictly in its activation patterns.
01:08:28 It’s keeping around a memory in its activity patterns
01:08:31 of what you did, how much reward you got,
01:08:35 and it’s using that activity based memory
01:08:38 as a basis for updating behavior.
01:08:41 But then the question is, well,
01:08:42 how did the prefrontal cortex get so smart?
01:08:44 In other words, where did these activity dynamics come from?
01:08:48 How did that program that’s implemented
01:08:50 in the recurrent dynamics of the prefrontal cortex arise?
01:08:54 And one answer that became evident in this work was,
01:08:58 well, maybe the mechanisms that operate
01:09:00 on the synaptic level, which we believe are mediated
01:09:05 by dopamine, are responsible for shaping those dynamics.
01:09:10 So this may be a silly question,
01:09:12 but because this kind of several temporal sort of classes
01:09:19 of learning are happening and the learning to learnism
01:09:23 emerges, can you keep building stacks of learning
01:09:28 to learn to learn, learning to learn to learn
01:09:30 to learn to learn because it keeps,
01:09:32 I mean, basically abstractions of more powerful abilities
01:09:37 to generalize of learning complex rules.
01:09:41 Yeah, that’s overstretching this kind of mechanism.
01:09:46 Well, one of the people in AI who started thinking
01:09:51 about meta learning from very early on,
01:09:54 Jürgen Schmidhuber sort of cheekily suggested,
01:09:59 I think it may have been in his PhD thesis,
01:10:03 that we should think about meta, meta, meta,
01:10:06 meta, meta, meta learning.
01:10:08 That’s really what’s gonna get us to true intelligence.
01:10:13 Certainly there’s a poetic aspect to it
01:10:15 and it seems interesting and correct
01:10:19 that that kind of levels of abstraction would be powerful,
01:10:21 but is that something you see in the brain?
01:10:23 This kind of, is it useful to think of learning
01:10:27 in these meta, meta, meta way or is it just meta learning?
01:10:32 Well, one thing that really fascinated me
01:10:35 about this mechanism that we were starting to look at,
01:10:39 and other groups started talking
01:10:41 about very similar things at the same time.
01:10:44 And then a kind of explosion of interest
01:10:47 in meta learning happened in the AI community
01:10:48 shortly after that.
01:10:50 I don’t know if we had anything to do with that,
01:10:52 but I was gratified to see that a lot of people
01:10:55 started talking about meta learning.
01:10:57 One of the things that I liked about the kind of flavor
01:11:01 of meta learning that we were studying was that
01:11:04 it didn’t require anything special.
01:11:05 It was just, if you took a system that had
01:11:08 some form of memory that the function of which
01:11:12 could be shaped by pick URL algorithm,
01:11:16 then this would just happen, right?
01:11:19 I mean, there are a lot of forms of,
01:11:21 there are a lot of meta learning algorithms
01:11:23 that have been proposed since then
01:11:24 that are fascinating and effective
01:11:26 in their domains of application.
01:11:29 But they’re engineered, they’re things that somebody
01:11:32 had to say, well, gee, if we wanted meta learning
01:11:34 to happen, how would we do that?
01:11:35 Here’s an algorithm that would,
01:11:37 but there’s something about the kind of meta learning
01:11:39 that we were studying that seemed to me special
01:11:42 in the sense that it wasn’t an algorithm.
01:11:44 It was just something that automatically happened
01:11:48 if you had a system that had memory
01:11:51 and it was trained with a reinforcement learning algorithm.
01:11:54 And in that sense, it can be as meta as it wants to be.
01:11:59 There’s no limit on how abstract the meta learning can get
01:12:04 because it’s not reliant on a human engineering
01:12:07 a particular meta learning algorithm to get there.
01:12:11 And that’s, I also, I don’t know,
01:12:15 I guess I hope that that’s relevant in the brain.
01:12:17 I think there’s a kind of beauty
01:12:19 in the ability of this emergent.
01:12:23 The emergent aspect of it, as opposed to engineered.
01:12:26 Exactly, it’s something that just, it just happens
01:12:29 in a sense, in a sense, you can’t avoid this happening.
01:12:33 If you have a system that has memory
01:12:35 and the function of that memory is shaped
01:12:39 by reinforcement learning, and this system is trained
01:12:42 in a series of interrelated tasks, this is gonna happen.
01:12:46 You can’t stop it.
01:12:48 As long as you have certain properties,
01:12:50 maybe like a recurrent structure to.
01:12:52 You have to have memory.
01:12:53 It actually doesn’t have to be a recurrent neural network.
01:12:55 One of, a paper that I was honored to be involved
01:12:58 with even earlier, used a kind of slot based memory.
01:13:02 Do you remember the title?
01:13:03 Just for people to understand.
01:13:05 It was Memory Augmented Neural Networks.
01:13:08 I think it was, I think the title was
01:13:10 Meta Learning in Memory Augmented Neural Networks.
01:13:14 And it was the same exact story.
01:13:17 If you have a system with memory,
01:13:21 here it was a different kind of memory,
01:13:22 but the function of that memory is shaped
01:13:26 by reinforcement learning.
01:13:29 Here it was the reads and writes that occurred
01:13:34 on this slot based memory.
01:13:36 This will just happen.
01:13:39 But this brings us back to something I was saying earlier
01:13:42 about the importance of the environment.
01:13:46 This will happen if the system is being trained
01:13:49 in a setting where there’s like a sequence of tasks
01:13:53 that all share some abstract structure.
01:13:56 Sometimes we talk about task distributions.
01:13:59 And that’s something that’s very obviously true
01:14:04 of the world that humans inhabit.
01:14:09 Like if you just kind of think about what you do every day,
01:14:13 you never do exactly the same thing
01:14:16 that you did the day before.
01:14:17 But everything that you do sort of has a family resemblance.
01:14:21 It shares a structure with something that you did before.
01:14:23 And so the real world is sort of
01:14:29 saturated with this kind of, this property.
01:14:32 It’s endless variety with endless redundancy.
01:14:37 And that’s the setting in which
01:14:38 this kind of meta learning happens.
01:14:40 And it does seem like we’re just so good at finding,
01:14:44 just like in this emergent phenomena you described,
01:14:47 we’re really good at finding that redundancy,
01:14:50 finding those similarities, the family resemblance.
01:14:53 Some people call it sort of, what is it?
01:14:56 Melanie Mitchell was talking about analogies.
01:14:59 So we’re able to connect concepts together
01:15:01 in this kind of way,
01:15:03 in this same kind of automated emergent way,
01:15:06 which there’s so many echoes here
01:15:08 of psychology and neuroscience.
01:15:10 And obviously now with reinforcement learning
01:15:15 with recurrent neural networks at the core.
01:15:18 If we could talk a little bit about dopamine,
01:15:20 you have really, you’re a part of coauthoring
01:15:23 really exciting recent paper, very recent,
01:15:26 in terms of release on dopamine
01:15:28 and temporal difference learning.
01:15:31 Can you describe the key ideas of that paper?
01:15:34 Sure, yeah.
01:15:35 I mean, one thing I want to pause to do
01:15:37 is acknowledge my coauthors
01:15:39 on actually both of the papers we’re talking about.
01:15:41 So this dopamine paper.
01:15:42 I’ll just, I’ll certainly post all their names.
01:15:45 Okay, wonderful.
01:15:46 Yeah, because I’m sort of abashed
01:15:49 to be the spokesperson for these papers
01:15:51 when I had such amazing collaborators on both.
01:15:55 So it’s a comfort to me to know
01:15:56 that you’ll acknowledge them.
01:15:58 Yeah, there’s an incredible team there, but yeah.
01:16:00 Oh yeah, it’s such a, it’s so much fun.
01:16:03 And in the case of the dopamine paper,
01:16:06 we also collaborated with Naochit at Harvard,
01:16:09 who, you know, obviously a paper simply
01:16:11 wouldn’t have happened without him.
01:16:12 But so you were asking for like a thumbnail sketch of.
01:16:17 Yeah, thumbnail sketch or key ideas or, you know,
01:16:20 things, the insights that are, you know,
01:16:22 continuing on our kind of discussion here
01:16:24 between neuroscience and AI.
01:16:26 Yeah, I mean, this was another,
01:16:28 a lot of the work that we’ve done so far
01:16:30 is taking ideas that have bubbled up in AI
01:16:35 and, you know, asking the question of whether the brain
01:16:39 might be doing something related,
01:16:41 which I think on the surface sounds like something
01:16:45 that’s really mainly of use to neuroscience.
01:16:49 We see it also as a way of validating
01:16:53 what we’re doing on the AI side.
01:16:55 If we can gain some evidence that the brain
01:16:57 is using some technique that we’ve been trying out
01:17:01 in our AI work, that gives us confidence
01:17:05 that, you know, it may be a good idea,
01:17:07 that it’ll, you know, scale to rich, complex tasks,
01:17:11 that it’ll interface well with other mechanisms.
01:17:14 So you see it as a two way road.
01:17:16 Yeah, for sure. Just because a particular paper
01:17:18 is a little bit focused on from one to the,
01:17:21 from AI, from neural networks to neuroscience.
01:17:25 Ultimately the discussion, the thinking,
01:17:28 the productive longterm aspect of it
01:17:30 is the two way road nature of the whole interaction.
01:17:33 Yeah, I mean, we’ve talked about the notion
01:17:36 of a virtuous circle between AI and neuroscience.
01:17:39 And, you know, the way I see it,
01:17:42 that’s always been there since the two fields,
01:17:47 you know, jointly existed.
01:17:50 There have been some phases in that history
01:17:52 when AI was sort of ahead.
01:17:53 There are some phases when neuroscience was sort of ahead.
01:17:56 I feel like given the burst of innovation
01:18:00 that’s happened recently on the AI side,
01:18:03 AI is kind of ahead in the sense that
01:18:06 there are all of these ideas that we, you know,
01:18:10 for which it’s exciting to consider
01:18:12 that there might be neural analogs.
01:18:16 And neuroscience, you know,
01:18:19 in a sense has been focusing on approaches
01:18:22 to studying behavior that come from, you know,
01:18:24 that are kind of derived from this earlier era
01:18:27 of cognitive psychology.
01:18:29 And, you know, so in some ways fail to connect
01:18:33 with some of the issues that we’re grappling with in AI.
01:18:36 Like how do we deal with, you know,
01:18:37 large, you know, complex environments.
01:18:41 But, you know, I think it’s inevitable
01:18:45 that this circle will keep turning
01:18:47 and there will be a moment
01:18:49 in the not too different distant future
01:18:51 when neuroscience is pelting AI researchers
01:18:54 with insights that may change the direction of our work.
01:18:58 Just a quick human question.
01:19:00 Is it, you have parts of your brain,
01:19:05 this is very meta, but they’re able to both think
01:19:08 about neuroscience and AI.
01:19:10 You know, I don’t often meet people like that.
01:19:14 So do you think, let me ask a meta plasticity question.
01:19:19 Do you think a human being can be both good at AI
01:19:22 and neuroscience?
01:19:23 It’s like what, on the team at DeepMind,
01:19:26 what kind of human can occupy these two realms?
01:19:30 And is that something you see everybody should be doing,
01:19:33 can be doing, or is that a very special few
01:19:36 can kind of jump?
01:19:37 Just like we talk about art history,
01:19:39 I would think it’s a special person
01:19:41 that can major in art history
01:19:43 and also consider being a surgeon.
01:19:46 Otherwise known as a dilettante.
01:19:48 A dilettante, yeah.
01:19:50 Easily distracted.
01:19:52 No, I think it does take a special kind of person
01:19:58 to be truly world class at both AI and neuroscience.
01:20:02 And I am not on that list.
01:20:05 I happen to be someone whose interest in neuroscience
01:20:10 and psychology involved using the kinds
01:20:15 of modeling techniques that are now very central in AI.
01:20:20 And that sort of, I guess, bought me a ticket
01:20:24 to be involved in all of the amazing things
01:20:26 that are going on in AI research right now.
01:20:29 I do know a few people who I would consider
01:20:32 pretty expert on both fronts,
01:20:34 and I won’t embarrass them by naming them,
01:20:36 but there are exceptional people out there
01:20:40 who are like this.
01:20:41 The one thing that I find is a barrier
01:20:45 to being truly world class on both fronts
01:20:49 is just the complexity of the technology
01:20:54 that’s involved in both disciplines now.
01:20:58 So the engineering expertise that it takes
01:21:02 to do truly frontline, hands on AI research
01:21:07 is really, really considerable.
01:21:10 The learning curve of the tools,
01:21:11 just like the specifics of just whether it’s programming
01:21:15 or the kind of tools necessary to collect the data,
01:21:17 to manage the data, to distribute, to compute,
01:21:19 all that kind of stuff.
01:21:20 And on the neuroscience, I guess, side,
01:21:22 there’ll be all different sets of tools.
01:21:24 Exactly, especially with the recent explosion
01:21:26 in neuroscience methods.
01:21:28 So having said all that,
01:21:32 I think the best scenario for both neuroscience
01:21:39 and AI is to have people interacting
01:21:44 who live at every point on this spectrum
01:21:48 from exclusively focused on neuroscience
01:21:51 to exclusively focused on the engineering side of AI.
01:21:55 But to have those people inhabiting a community
01:22:01 where they’re talking to people who live elsewhere
01:22:03 on the spectrum.
01:22:04 And I may be someone who’s very close to the center
01:22:08 in the sense that I have one foot in the neuroscience world
01:22:12 and one foot in the AI world,
01:22:14 and that central position, I will admit,
01:22:17 prevents me, at least someone
01:22:19 with my limited cognitive capacity,
01:22:21 from having true technical expertise in either domain.
01:22:26 But at the same time, I at least hope
01:22:30 that it’s worthwhile having people around
01:22:32 who can kind of see the connections.
01:22:34 Yeah, the community, the emergent intelligence
01:22:39 of the community when it’s nicely distributed is useful.
01:22:43 Exactly, yeah.
01:22:44 So hopefully that, I mean, I’ve seen that work,
01:22:46 I’ve seen that work out well at DeepMind.
01:22:48 There are people who, I mean, even if you just focus
01:22:52 on the AI work that happens at DeepMind,
01:22:55 it’s been a good thing to have some people around
01:22:59 doing that kind of work whose PhDs are in neuroscience
01:23:03 or psychology.
01:23:04 Every academic discipline has its kind of blind spots
01:23:09 and kind of unfortunate obsessions and its metaphors
01:23:16 and its reference points,
01:23:18 and having some intellectual diversity is really healthy.
01:23:24 People get each other unstuck, I think.
01:23:28 I see it all the time at DeepMind.
01:23:30 And I like to think that the people
01:23:33 who bring some neuroscience background to the table
01:23:35 are helping with that.
01:23:37 So one of my probably the deepest passion for me,
01:23:41 what I would say, maybe we kind of spoke off mic
01:23:44 a little bit about it, but that I think is a blind spot
01:23:49 for at least robotics and AI folks
01:23:51 is human robot interaction, human agent interaction.
01:23:55 Maybe do you have thoughts about how we reduce the size
01:24:01 of that blind spot?
01:24:02 Do you also share the feeling that not enough folks
01:24:07 are studying this aspect of interaction?
01:24:10 Well, I’m actually pretty intensively interested
01:24:14 in this issue now, and there are people in my group
01:24:17 who’ve actually pivoted pretty hard over the last few years
01:24:20 from doing more traditional cognitive psychology
01:24:24 and cognitive neuroscience to doing experimental work
01:24:28 on human agent interaction.
01:24:30 And there are a couple of reasons that I’m
01:24:33 pretty passionately interested in this.
01:24:35 One is it’s kind of the outcome of having thought
01:24:42 for a few years now about what we’re up to.
01:24:46 Like what are we doing?
01:24:49 Like what is this AI research for?
01:24:53 So what does it mean to make the world a better place?
01:24:57 I think I’m pretty sure that means making life better
01:24:59 for humans.
01:25:02 And so how do you make life better for humans?
01:25:05 That’s a proposition that when you look at it carefully
01:25:10 and honestly is rather horrendously complicated,
01:25:15 especially when the AI systems
01:25:18 that you’re building are learning systems.
01:25:25 They’re not, you’re not programming something
01:25:29 that you then introduce to the world
01:25:31 and it just works as programmed,
01:25:33 like Google Maps or something.
01:25:36 We’re building systems that learn from experience.
01:25:39 So that typically leads to AI safety questions.
01:25:43 How do we keep these things from getting out of control?
01:25:45 How do we keep them from doing things that harm humans?
01:25:49 And I mean, I hasten to say,
01:25:51 I consider those hugely important issues.
01:25:54 And there are large sectors of the research community
01:25:58 at DeepMind and of course elsewhere
01:26:00 who are dedicated to thinking hard all day,
01:26:03 every day about that.
01:26:04 But there’s, I guess I would say a positive side to this too
01:26:09 which is to say, well, what would it mean
01:26:13 to make human life better?
01:26:15 And how can we imagine learning systems doing that?
01:26:21 And in talking to my colleagues about that,
01:26:23 we reached the initial conclusion
01:26:25 that it’s not sufficient to philosophize about that.
01:26:30 You actually have to take into account
01:26:32 how humans actually work and what humans want
01:26:37 and the difficulties of knowing what humans want
01:26:41 and the difficulties that arise
01:26:43 when humans want different things.
01:26:47 And so human agent interaction has become,
01:26:50 a quite intensive focus of my group lately.
01:26:56 If for no other reason that,
01:26:59 in order to really address that issue in an adequate way,
01:27:04 you have to, I mean, psychology becomes part of the picture.
01:27:07 Yeah, and so there’s a few elements there.
01:27:10 So if you focus on solving like the,
01:27:12 if you focus on the robotics problem,
01:27:14 let’s say AGI without humans in the picture
01:27:18 is you’re missing fundamentally the final step.
01:27:22 When you do want to help human civilization,
01:27:24 you eventually have to interact with humans.
01:27:27 And when you create a learning system, just as you said,
01:27:31 that will eventually have to interact with humans,
01:27:34 the interaction itself has to be become,
01:27:37 has to become part of the learning process.
01:27:40 So you can’t just watch, well, my sense is,
01:27:43 it sounds like your sense is you can’t just watch humans
01:27:46 to learn about humans.
01:27:48 You have to also be part of the human world.
01:27:50 You have to interact with humans.
01:27:51 Yeah, exactly.
01:27:52 And I mean, then questions arise that start imperceptibly,
01:27:57 but inevitably to slip beyond the realm of engineering.
01:28:02 So questions like, if you have an agent
01:28:05 that can do something that you can’t do,
01:28:10 under what conditions do you want that agent to do it?
01:28:13 So if I have a robot that can play Beethoven sonatas
01:28:24 better than any human, in the sense that the sensitivity,
01:28:30 the expression is just beyond what any human,
01:28:33 do I want to listen to that?
01:28:36 Do I want to go to a concert and hear a robot play?
01:28:38 These aren’t engineering questions.
01:28:41 These are questions about human preference
01:28:44 and human culture.
01:28:45 Psychology bordering on philosophy.
01:28:47 Yeah, and then you start asking,
01:28:50 well, even if we knew the answer to that,
01:28:54 is it our place as AI engineers
01:28:57 to build that into these agents?
01:28:59 Probably the agents should interact with humans
01:29:03 beyond the population of AI engineers
01:29:05 and figure out what those humans want.
01:29:08 And then when you start,
01:29:10 I referred this the moment ago,
01:29:11 but even that becomes complicated.
01:29:14 Be quote, what if two humans want different things?
01:29:19 And you have only one agent that’s able to interact with them
01:29:22 and try to satisfy their preferences.
01:29:24 Then you’re into the realm of economics
01:29:30 and social choice theory and even politics.
01:29:33 So there’s a sense in which,
01:29:35 if you kind of follow what we’re doing
01:29:37 to its logical conclusion,
01:29:39 then it goes beyond questions of engineering and technology
01:29:45 and starts to shade imperceptibly into questions
01:29:48 about what kind of society do you want?
01:29:51 And actually, once that dawned on me,
01:29:55 I actually felt,
01:29:58 I don’t know what the right word is,
01:29:59 quite refreshed in my involvement in AI research.
01:30:03 It was almost like building this kind of stuff
01:30:06 is gonna lead us back to asking really fundamental questions
01:30:10 about what is this,
01:30:13 what’s the good life and who gets to decide
01:30:16 and bringing in viewpoints from multiple sub communities
01:30:23 to help us shape the way that we live.
01:30:27 There’s something, it started making me feel like
01:30:30 doing AI research in a fully responsible way, would,
01:30:38 could potentially lead to a kind of like cultural renewal.
01:30:42 Yeah, it’s the way to understand human beings
01:30:48 at the individual, at the societal level.
01:30:50 It may become a way to answer all the silly human questions
01:30:54 of the meaning of life and all those kinds of things.
01:30:57 Even if it doesn’t give us a way
01:30:58 of answering those questions,
01:30:59 it may force us back to thinking about them.
01:31:03 And it might bring, it might restore a certain,
01:31:06 I don’t know, a certain depth to,
01:31:10 or even dare I say spirituality to the way that,
01:31:16 to the world, I don’t know.
01:31:18 Maybe that’s too grandiose.
01:31:19 Well, I’m with you.
01:31:21 I think it’s AI will be the philosophy of the 21st century,
01:31:27 the way which will open the door.
01:31:29 I think a lot of AI researchers are afraid to open that door
01:31:32 of exploring the beautiful richness
01:31:35 of the human agent interaction, human AI interaction.
01:31:39 I’m really happy that somebody like you
01:31:42 have opened that door.
01:31:43 And one thing I often think about is the usual schema
01:31:49 for thinking about human agent interaction
01:31:54 as this kind of dystopian, oh, our robot overlords.
01:32:00 And again, I hasten to say AI safety is hugely important.
01:32:03 And I’m not saying we shouldn’t be thinking
01:32:06 about those risks, totally on board for that.
01:32:09 But there’s, having said that,
01:32:17 what often follows for me is the thought
01:32:18 that there’s another kind of narrative
01:32:22 that might be relevant, which is,
01:32:24 when we think of humans gaining more and more information
01:32:31 about human life, the narrative there is usually
01:32:36 that they gain more and more wisdom
01:32:38 and they get closer to enlightenment
01:32:40 and they become more benevolent.
01:32:43 And the Buddha is like, that’s a totally different narrative.
01:32:47 And why isn’t it the case that we imagine
01:32:50 that the AI systems that we’re creating
01:32:52 are just gonna, like, they’re gonna figure out
01:32:53 more and more about the way the world works
01:32:55 and the way that humans interact
01:32:56 and they’ll become beneficent.
01:32:59 I’m not saying that will happen.
01:33:00 I don’t honestly expect that to happen
01:33:05 without some careful, setting things up very carefully.
01:33:08 But it’s another way things could go, right?
01:33:11 And yeah, and I would even push back on that.
01:33:13 I personally believe that the most trajectories,
01:33:18 natural human trajectories will lead us towards progress.
01:33:25 So for me, there is a kind of sense
01:33:28 that most trajectories in AI development
01:33:30 will lead us into trouble.
01:33:32 To me, and we over focus on the worst case.
01:33:37 It’s like in computer science,
01:33:38 theoretical computer science has been this focus
01:33:40 on worst case analysis.
01:33:42 There’s something appealing to our human mind
01:33:45 at some lowest level to be good.
01:33:47 I mean, we don’t wanna be eaten by the tiger, I guess.
01:33:50 So we wanna do the worst case analysis.
01:33:52 But the reality is that shouldn’t stop us
01:33:55 from actually building out all the other trajectories
01:33:58 which are potentially leading to all the positive worlds,
01:34:01 all the enlightenment.
01:34:04 There’s a book, Enlightenment Now,
01:34:05 with Steven Pinker and so on.
01:34:06 This is looking generally at human progress.
01:34:09 And there’s so many ways that human progress
01:34:12 can happen with AI.
01:34:13 And I think you have to do that research.
01:34:16 You have to do that work.
01:34:17 You have to do the, not just the AI safety work
01:34:20 of the one worst case analysis.
01:34:22 How do we prevent that?
01:34:23 But the actual tools and the glue
01:34:27 and the mechanisms of human AI interaction
01:34:31 that would lead to all the positive actions that can go.
01:34:34 It’s a super exciting area, right?
01:34:36 Yeah, we should be spending,
01:34:38 we should be spending a lot of our time saying
01:34:40 what can go wrong.
01:34:42 I think it’s harder to see that there’s work to be done
01:34:47 to bring into focus the question of what it would look like
01:34:51 for things to go right.
01:34:54 That’s not obvious.
01:34:57 And we wouldn’t be doing this if we didn’t have the sense
01:34:59 there was huge potential, right?
01:35:01 We’re not doing this for no reason.
01:35:05 We have a sense that AGI would be a major boom to humanity.
01:35:10 But I think it’s worth starting now,
01:35:13 even when our technology is quite primitive,
01:35:15 asking exactly what would that mean?
01:35:19 We can start now with applications
01:35:21 that are already gonna make the world a better place,
01:35:22 like solving protein folding.
01:35:25 I think DeepMind has gotten heavy
01:35:27 into science applications lately,
01:35:30 which I think is a wonderful, wonderful move
01:35:34 for us to be making.
01:35:36 But when we think about AGI,
01:35:37 when we think about building fully intelligent
01:35:39 agents that are gonna be able to, in a sense,
01:35:42 do whatever they want,
01:35:45 we should start thinking about
01:35:46 what do we want them to want, right?
01:35:48 What kind of world do we wanna live in?
01:35:52 That’s not an easy question.
01:35:54 And I think we just need to start working on it.
01:35:56 And even on the path to,
01:35:58 it doesn’t have to be AGI,
01:35:59 but just intelligent agents that interact with us
01:36:02 and help us enrich our own existence on social networks,
01:36:06 for example, on recommender systems of various intelligence.
01:36:08 And there’s so much interesting interaction
01:36:10 that’s yet to be understood and studied.
01:36:12 And how do you create,
01:36:15 I mean, Twitter is struggling with this very idea,
01:36:19 how do you create AI systems
01:36:21 that increase the quality and the health of a conversation?
01:36:24 For sure.
01:36:25 That’s a beautiful human psychology question.
01:36:28 And how do you do that
01:36:29 without deception being involved,
01:36:34 without manipulation being involved,
01:36:38 maximizing human autonomy?
01:36:42 And how do you make these choices in a democratic way?
01:36:45 How do we face the,
01:36:50 again, I’m speaking for myself here.
01:36:52 How do we face the fact that
01:36:55 it’s a small group of people
01:36:57 who have the skillset to build these kinds of systems,
01:37:01 but what it means to make the world a better place
01:37:05 is something that we all have to be talking about.
01:37:09 Yeah, the world that we’re trying to make a better place
01:37:14 includes a huge variety of different kinds of people.
01:37:18 Yeah, how do we cope with that?
01:37:19 This is a problem that has been discussed
01:37:22 in gory, extensive detail in social choice theory.
01:37:28 One thing I’m really interested in
01:37:29 and one thing I’m really enjoying
01:37:32 about the recent direction work has taken
01:37:35 in some parts of my team is that,
01:37:36 yeah, we’re reading the AI literature,
01:37:38 we’re reading the neuroscience literature,
01:37:39 but we’ve also started reading economics
01:37:42 and, as I mentioned, social choice theory,
01:37:44 even some political theory,
01:37:45 because it turns out that it all becomes relevant.
01:37:50 It all becomes relevant.
01:37:53 But at the same time,
01:37:55 we’ve been trying not to write philosophy papers,
01:38:00 we’ve been trying not to write physician papers.
01:38:01 We’re trying to figure out ways
01:38:03 of doing actual empirical research
01:38:05 that kind of take the first small steps
01:38:07 to thinking about what it really means
01:38:10 for humans with all of their complexity
01:38:13 and contradiction and paradox
01:38:18 to be brought into contact with these AI systems
01:38:22 in a way that really makes the world a better place.
01:38:25 Often, reinforcement learning frameworks
01:38:27 actually kind of allow you to do that,
01:38:30 machine learning, and so that’s the exciting thing about AI
01:38:33 is it allows you to reduce the unsolvable problem,
01:38:37 philosophical problem, into something more concrete
01:38:40 that you can get ahold of.
01:38:41 Yeah, and it allows you to kind of define the problem
01:38:43 in some way that allows for growth in the system
01:38:49 that’s sort of, you know,
01:38:51 you’re not responsible for the details, right?
01:38:54 You say, this is generally what I want you to do,
01:38:56 and then learning takes care of the rest.
01:38:59 Of course, the safety issues arise in that context,
01:39:04 but I think also some of these positive issues
01:39:05 arise in that context.
01:39:06 What would it mean for an AI system
01:39:09 to really come to understand what humans want?
01:39:14 And with all of the subtleties of that, right?
01:39:18 You know, humans want help with certain things,
01:39:24 but they don’t want everything done for them, right?
01:39:27 There is, part of the satisfaction
01:39:29 that humans get from life is in accomplishing things.
01:39:32 So if there were devices around that did everything for,
01:39:34 you know, I often think of the movie WALLI, right?
01:39:37 That’s like dystopian in a totally different way.
01:39:39 It’s like, the machines are doing everything for us.
01:39:41 That’s not what we wanted.
01:39:43 You know, anyway, I find this, you know,
01:39:46 this opens up a whole landscape of research
01:39:50 that feels affirmative and exciting.
01:39:52 To me, it’s one of the most exciting, and it’s wide open.
01:39:56 We have to, because it’s a cool paper,
01:39:58 talk about dopamine.
01:39:59 Oh yeah, okay, so I can.
01:40:01 We were gonna, I was gonna give you a quick summary.
01:40:04 Yeah, a quick summary of, what’s the title of the paper?
01:40:09 I think we called it a distributional code for value
01:40:14 in dopamine based reinforcement learning, yes.
01:40:19 So that’s another project that grew out of pure AI research.
01:40:25 A number of people at DeepMind and a few other places
01:40:29 had started working on a new version
01:40:32 of reinforcement learning,
01:40:35 which was defined by taking something
01:40:38 in traditional reinforcement learning and just tweaking it.
01:40:41 So the thing that they took
01:40:42 from traditional reinforcement learning was a value signal.
01:40:46 So at the center of reinforcement learning,
01:40:49 at least most algorithms, is some representation
01:40:52 of how well things are going,
01:40:54 your expected cumulative future reward.
01:40:57 And that’s usually represented as a single number.
01:41:01 So if you imagine a gambler in a casino
01:41:04 and the gambler’s thinking, well, I have this probability
01:41:07 of winning such and such an amount of money,
01:41:09 and I have this probability of losing such and such
01:41:11 an amount of money, that situation would be represented
01:41:14 as a single number, which is like the expected,
01:41:17 the weighted average of all those outcomes.
01:41:20 And this new form of reinforcement learning said,
01:41:23 well, what if we generalize that
01:41:26 to a distributional representation?
01:41:28 So now we think of the gambler as literally thinking,
01:41:30 well, there’s this probability
01:41:32 that I’ll win this amount of money,
01:41:33 and there’s this probability
01:41:34 that I’ll lose that amount of money,
01:41:35 and we don’t reduce that to a single number.
01:41:37 And it had been observed through experiments,
01:41:40 through just trying this out,
01:41:42 that that kind of distributional representation
01:41:45 really accelerated reinforcement learning
01:41:49 and led to better policies.
01:41:52 What’s your intuition about,
01:41:53 so we’re talking about rewards.
01:41:55 Yeah.
01:41:56 So what’s your intuition why that is, why does it do that?
01:41:58 Well, it’s kind of a surprising historical note,
01:42:02 at least surprised me when I learned it,
01:42:04 that this had been proven to be true.
01:42:07 This had been tried out in a kind of heuristic way.
01:42:09 People thought, well, gee, what would happen if we tried?
01:42:12 And then it had this, empirically,
01:42:14 it had this striking effect.
01:42:17 And it was only then that people started thinking,
01:42:19 well, gee, wait, why?
01:42:21 Wait, why?
01:42:22 Why is this working?
01:42:23 And that’s led to a series of studies
01:42:26 just trying to figure out why it works, which is ongoing.
01:42:29 But one thing that’s already clear from that research
01:42:31 is that one reason that it helps
01:42:34 is that it drives richer representation learning.
01:42:39 So if you imagine two situations
01:42:43 that have the same expected value,
01:42:45 the same kind of weighted average value,
01:42:48 standard deep reinforcement learning algorithms
01:42:51 are going to take those two situations
01:42:53 and kind of, in terms of the way
01:42:55 they’re represented internally,
01:42:56 they’re gonna squeeze them together
01:42:58 because the thing that you’re trying to represent,
01:43:02 which is their expected value, is the same.
01:43:04 So all the way through the system,
01:43:06 things are gonna be mushed together.
01:43:08 But what if those two situations
01:43:11 actually have different value distributions?
01:43:13 They have the same average value,
01:43:16 but they have different distributions of value.
01:43:19 In that situation, distributional learning
01:43:22 will maintain the distinction between these two things.
01:43:25 So to make a long story short,
01:43:26 distributional learning can keep things separate
01:43:30 in the internal representation
01:43:32 that might otherwise be conflated or squished together.
01:43:35 And maintaining those distinctions
01:43:36 can be useful when the system is now faced
01:43:40 with some other task where the distinction is important.
01:43:43 If we look at the optimistic
01:43:44 and pessimistic dopamine neurons.
01:43:46 So first of all, what is dopamine?
01:43:50 Oh, God.
01:43:51 Why is this at all useful
01:43:58 to think about in the artificial intelligence sense?
01:44:00 But what do we know about dopamine in the human brain?
01:44:04 What is it?
01:44:05 Why is it useful?
01:44:06 Why is it interesting?
01:44:07 What does it have to do with the prefrontal cortex
01:44:09 and learning in general?
01:44:10 Yeah, so, well, this is also a case
01:44:15 where there’s a huge amount of detail and debate.
01:44:19 But one currently prevailing idea
01:44:24 is that the function of this neurotransmitter dopamine
01:44:29 resembles a particular component
01:44:33 of standard reinforcement learning algorithms,
01:44:36 which is called the reward prediction error.
01:44:39 So I was talking a moment ago
01:44:41 about these value representations.
01:44:44 How do you learn them?
01:44:45 How do you update them based on experience?
01:44:46 Well, if you made some prediction about a future reward
01:44:51 and then you get more reward than you were expecting,
01:44:54 then probably retrospectively,
01:44:56 you want to go back and increase the value representation
01:45:00 that you attached to that earlier situation.
01:45:03 If you got less reward than you were expecting,
01:45:06 you should probably decrement that estimate.
01:45:08 And that’s the process of temporal difference.
01:45:10 Exactly, this is the central mechanism
01:45:12 of temporal difference learning,
01:45:12 which is one of several sort of the backbone
01:45:17 of our momentarium in NRL.
01:45:20 And this connection between the reward prediction error
01:45:25 and dopamine was made in the 1990s.
01:45:31 And there’s been a huge amount of research
01:45:33 that seems to back it up.
01:45:35 Dopamine may be doing other things,
01:45:37 but this is clearly, at least roughly,
01:45:39 one of the things that it’s doing.
01:45:42 But the usual idea was that dopamine
01:45:45 was representing these reward prediction errors,
01:45:48 again, in this like kind of single number way
01:45:51 that representing your surprise with a single number.
01:45:56 And in distributional reinforcement learning,
01:45:58 this kind of new elaboration of the standard approach,
01:46:03 it’s not only the value function
01:46:06 that’s represented as a single number,
01:46:08 it’s also the reward prediction error.
01:46:10 And so what happened was that Will Dabney,
01:46:16 one of my collaborators who was one of the first people
01:46:18 to work on distributional temporal difference learning,
01:46:22 talked to a guy in my group, Zeb Kurt Nelson,
01:46:25 who’s a computational neuroscientist,
01:46:27 and said, gee, you know, is it possible
01:46:29 that dopamine might be doing something
01:46:31 like this distributional coding thing?
01:46:33 And they started looking at what was in the literature,
01:46:35 and then they brought me in,
01:46:36 and we started talking to Nao Uchida,
01:46:39 and we came up with some specific predictions
01:46:41 about if the brain is using
01:46:43 this kind of distributional coding,
01:46:45 then in the tasks that Nao has studied,
01:46:47 you should see this, this, this, and this,
01:46:49 and that’s where the paper came from.
01:46:50 We kind of enumerated a set of predictions,
01:46:53 all of which ended up being fairly clearly confirmed,
01:46:57 and all of which leads to at least some initial indication
01:47:00 that the brain might be doing something
01:47:02 like this distributional coding,
01:47:03 that dopamine might be representing surprise signals
01:47:06 in a way that is not just collapsing everything
01:47:09 to a single number, but instead is kind of respecting
01:47:12 the variety of future outcomes, if that makes sense.
01:47:16 So yeah, so that’s showing, suggesting possibly
01:47:19 that dopamine has a really interesting
01:47:21 representation scheme in the human brain
01:47:25 for its reward signal.
01:47:27 Exactly. That’s fascinating.
01:47:29 That’s another beautiful example of AI
01:47:32 revealing something nice about neuroscience,
01:47:34 potentially suggesting possibilities.
01:47:36 Well, you never know.
01:47:37 So the minute you publish a paper like that,
01:47:39 the next thing you think is, I hope that replicates.
01:47:42 Like, I hope we see that same thing in other data sets,
01:47:44 but of course, several labs now
01:47:47 are doing the followup experiments, so we’ll know soon.
01:47:50 But it has been a lot of fun for us
01:47:52 to take these ideas from AI
01:47:54 and kind of bring them into neuroscience
01:47:56 and see how far we can get.
01:47:58 So we kind of talked about it a little bit,
01:48:01 but where do you see the field of neuroscience
01:48:04 and artificial intelligence heading broadly?
01:48:07 Like, what are the possible exciting areas
01:48:12 that you can see breakthroughs in the next,
01:48:15 let’s get crazy, not just three or five years,
01:48:17 but the next 10, 20, 30 years
01:48:22 that would make you excited
01:48:26 and perhaps you’d be part of?
01:48:29 On the neuroscience side,
01:48:32 there’s a great deal of interest now
01:48:34 in what’s going on in AI.
01:48:36 And at the same time,
01:48:41 I feel like, so neuroscience,
01:48:45 especially the part of neuroscience
01:48:50 that’s focused on circuits and systems,
01:48:54 kind of like really mechanism focused,
01:48:57 there’s been this explosion in new technology.
01:49:01 And up until recently,
01:49:05 the experiments that have exploited this technology
01:49:08 have not involved a lot of interesting behavior.
01:49:13 And this is for a variety of reasons,
01:49:16 one of which is in order to employ
01:49:18 some of these technologies,
01:49:19 you actually have to, if you’re studying a mouse,
01:49:22 you have to head fix the mouse.
01:49:23 In other words, you have to like immobilize the mouse.
01:49:26 And so it’s been tricky to come up
01:49:28 with ways of eliciting interesting behavior
01:49:30 from a mouse that’s restrained in this way,
01:49:33 but people have begun to create
01:49:35 very interesting solutions to this,
01:49:39 like virtual reality environments
01:49:41 where the animal can kind of move a track ball.
01:49:43 And as people have kind of begun to explore
01:49:48 what you can do with these technologies,
01:49:50 I feel like more and more people are asking,
01:49:52 well, let’s try to bring behavior into the picture.
01:49:55 Let’s try to like reintroduce behavior,
01:49:58 which was supposed to be what this whole thing was about.
01:50:01 And I’m hoping that those two trends,
01:50:05 the kind of growing interest in behavior
01:50:09 and the widespread interest in what’s going on in AI,
01:50:14 will come together to kind of open a new chapter
01:50:17 in neuroscience research where there’s a kind of
01:50:22 a rebirth of interest in the structure of behavior
01:50:25 and its underlying substrates,
01:50:27 but that that research is being informed
01:50:31 by computational mechanisms
01:50:33 that we’re coming to understand in AI.
01:50:36 If we can do that, then we might be taking a step closer
01:50:39 to this utopian future that we were talking about earlier
01:50:43 where there’s really no distinction
01:50:44 between psychology and neuroscience.
01:50:46 Neuroscience is about studying the mechanisms
01:50:50 that underlie whatever it is the brain is for,
01:50:54 and what is the brain for?
01:50:56 What is the brain for? It’s for behavior.
01:50:58 I feel like we could maybe take a step toward that now
01:51:03 if people are motivated in the right way.
01:51:06 You also asked about AI.
01:51:08 So that was a neuroscience question.
01:51:10 You said neuroscience, that’s right.
01:51:12 And especially places like DeepMind
01:51:13 are interested in both branches.
01:51:15 So what about the engineering of intelligence systems?
01:51:20 I think one of the key challenges
01:51:24 that a lot of people are seeing now in AI
01:51:28 is to build systems that have the kind of flexibility
01:51:34 and the kind of flexibility that humans have in two senses.
01:51:38 One is that humans can be good at many things.
01:51:41 They’re not just expert at one thing.
01:51:44 And they’re also flexible in the sense
01:51:45 that they can switch between things very easily
01:51:49 and they can pick up new things very quickly
01:51:52 because they very ably see what a new task has in common
01:51:57 with other things that they’ve done.
01:52:01 And that’s something that our AI systems
01:52:05 just blatantly do not have.
01:52:09 There are some people who like to argue
01:52:11 that deep learning and deep RL
01:52:13 are simply wrong for getting that kind of flexibility.
01:52:17 I don’t share that belief,
01:52:20 but the simpler fact of the matter
01:52:22 is we’re not building things yet
01:52:23 that do have that kind of flexibility.
01:52:25 And I think the attention of a large part
01:52:28 of the AI community is starting to pivot to that question.
01:52:31 How do we get that?
01:52:33 That’s gonna lead to a focus on abstraction.
01:52:38 It’s gonna lead to a focus on
01:52:40 what in psychology we call cognitive control,
01:52:43 which is the ability to switch between tasks,
01:52:45 the ability to quickly put together a program of behavior
01:52:49 that you’ve never executed before,
01:52:51 but you know makes sense for a particular set of demands.
01:52:55 It’s very closely related to what the prefrontal cortex does
01:52:59 on the neuroscience side.
01:53:01 So I think it’s gonna be an interesting new chapter.
01:53:05 So that’s the reasoning side and cognition side,
01:53:07 but let me ask the over romanticized question.
01:53:10 Do you think we’ll ever engineer an AGI system
01:53:13 that we humans would be able to love
01:53:17 and that would love us back?
01:53:19 So have that level and depth of connection?
01:53:26 I love that question.
01:53:27 And it relates closely to things
01:53:31 that I’ve been thinking about a lot lately,
01:53:33 in the context of this human AI research.
01:53:36 There’s social psychology research
01:53:41 in particular by Susan Fisk at Princeton
01:53:44 the department where I used to work,
01:53:48 where she dissects human attitudes toward other humans
01:53:54 into a sort of two dimensional scheme.
01:53:59 And one dimension is about ability.
01:54:03 How able, how capable is this other person?
01:54:10 But the other dimension is warmth.
01:54:11 So you can imagine another person who’s very skilled
01:54:15 and capable, but is very cold.
01:54:19 And you wouldn’t really like highly,
01:54:22 you might have some reservations about that other person.
01:54:26 But there’s also a kind of reservation
01:54:28 that we might have about another person
01:54:31 who elicits in us or displays a lot of human warmth,
01:54:34 but is not good at getting things done.
01:54:37 We reserve our greatest esteem really
01:54:40 for people who are both highly capable
01:54:43 and also quite warm.
01:54:47 That’s like the best of the best.
01:54:49 This isn’t a normative statement I’m making.
01:54:53 This is just an empirical statement.
01:54:55 This is what humans seem…
01:54:57 These are the two dimensions that people seem to kind of like
01:54:59 along which people size other people up.
01:55:02 And in AI research,
01:55:03 there’s a lot of people who think that humans are
01:55:06 very capable, and in AI research,
01:55:08 we really focus on this capability thing.
01:55:11 We want our agents to be able to do stuff.
01:55:13 This thing can play go at a superhuman level.
01:55:15 That’s awesome.
01:55:16 But that’s only one dimension.
01:55:18 What about the other dimension?
01:55:20 What would it mean for an AI system to be warm?
01:55:25 And I don’t know, maybe there are easy solutions here.
01:55:27 Like we can put a face on our AI systems.
01:55:30 It’s cute, it has big ears.
01:55:32 I mean, that’s probably part of it.
01:55:33 But I think it also has to do with a pattern of behavior.
01:55:36 A pattern of what would it mean for an AI system
01:55:40 to display caring, compassionate behavior
01:55:43 in a way that actually made us feel like it was for real?
01:55:47 That we didn’t feel like it was simulated.
01:55:49 We didn’t feel like we were being duped.
01:55:53 To me, people talk about the Turing test
01:55:55 or some descendant of it.
01:55:57 I feel like that’s the ultimate Turing test.
01:56:01 Is there an AI system that can not only convince us
01:56:05 that it knows how to reason
01:56:07 and it knows how to interpret language,
01:56:09 but that we’re comfortable saying,
01:56:12 yeah, that AI system’s a good guy.
01:56:15 On the warmth scale, whatever warmth is,
01:56:18 we kind of intuitively understand it,
01:56:20 but we also wanna be able to, yeah,
01:56:25 we don’t understand it explicitly enough yet
01:56:29 to be able to engineer it.
01:56:30 Exactly.
01:56:31 And that’s an open scientific question.
01:56:33 You kind of alluded it several times
01:56:35 in the human AI interaction.
01:56:37 That’s a question that should be studied
01:56:38 and probably one of the most important questions
01:56:42 as we move to AGI.
01:56:43 We humans are so good at it.
01:56:46 Yeah.
01:56:46 It’s not just that we’re born warm.
01:56:50 I suppose some people are warmer than others
01:56:53 given whatever genes they manage to inherit.
01:56:55 But there are also learned skills involved.
01:57:01 There are ways of communicating to other people
01:57:04 that you care, that they matter to you,
01:57:07 that you’re enjoying interacting with them, right?
01:57:11 And we learn these skills from one another.
01:57:14 And it’s not out of the question
01:57:16 that we could build engineered systems.
01:57:20 I think it’s hopeless, as you say,
01:57:21 that we could somehow hand design
01:57:23 these sorts of behaviors.
01:57:26 But it’s not out of the question
01:57:27 that we could build systems that kind of,
01:57:30 we instill in them something that sets them out
01:57:34 in the right direction,
01:57:35 so that they end up learning what it is
01:57:39 to interact with humans
01:57:40 in a way that’s gratifying to humans.
01:57:44 I mean, honestly, if that’s not where we’re headed,
01:57:49 I want out.
01:57:50 I think it’s exciting as a scientific problem,
01:57:54 just as you described.
01:57:56 I honestly don’t see a better way to end it
01:57:59 than talking about warmth and love.
01:58:01 And Matt, I don’t think I’ve ever had such a wonderful
01:58:05 conversation where my questions were so bad
01:58:07 and your answers were so beautiful.
01:58:09 So I deeply appreciate it.
01:58:10 I really enjoyed it.
01:58:11 Thanks for talking to me.
01:58:12 Well, it’s been very fun.
01:58:13 As you can probably tell,
01:58:17 there’s something I like about kind of thinking
01:58:19 outside the box and like,
01:58:21 so it’s good having an opportunity to do that.
01:58:22 Awesome.
01:58:23 Thanks so much for doing it.
01:58:25 Thanks for listening to this conversation
01:58:27 with Matt Bopenik.
01:58:28 And thank you to our sponsors,
01:58:30 The Jordan Harbinger Show
01:58:32 and Magic Spoon Low Carb Keto Cereal.
01:58:36 Please consider supporting this podcast
01:58:38 by going to jordanharbinger.com slash lex
01:58:41 and also going to magicspoon.com slash lex
01:58:44 and using code lex at checkout.
01:58:48 Click the links, buy all the stuff.
01:58:50 It’s the best way to support this podcast
01:58:52 and the journey I’m on in my research and the startup.
01:58:57 If you enjoy this thing, subscribe on YouTube,
01:58:59 review it with the five stars in Apple Podcasts,
01:59:02 support it on Patreon, follow on Spotify
01:59:05 or connect with me on Twitter at lexfreedman.
01:59:08 Again, spelled miraculously without the E,
01:59:12 just F R I D M A N.
01:59:15 And now let me leave you with some words
01:59:17 from neurologist V.S. Amarachandran.
01:59:20 How can a three pound mass of jelly
01:59:23 that you can hold in your palm imagine angels,
01:59:26 contemplate the meaning of an infinity
01:59:28 and even question its own place in the cosmos?
01:59:31 Especially awe inspiring is the fact that any single brain,
01:59:35 including yours, is made up of atoms
01:59:38 that were forged in the hearts
01:59:40 of countless far flung stars billions of years ago.
01:59:45 These particles drifted for eons and light years
01:59:48 until gravity and change brought them together here now.
01:59:53 These atoms now form a conglomerate, your brain,
01:59:57 that can not only ponder the very stars they gave at birth,
02:00:00 but can also think about its own ability to think
02:00:04 and wonder about its own ability to wander.
02:00:07 With the arrival of humans, it has been said,
02:00:10 the universe has suddenly become conscious of itself.
02:00:14 This truly is the greatest mystery of all.
02:00:18 Thank you for listening and hope to see you next time.