Dileep George: Brain-Inspired AI | Lex Fridman Podcast #115

Lex Fridman 0:00 The following is a conversation with Daleep George, a researcher at the intersection of neuroscience and artificial intelligence, co-founder of Vicarious with Scott Phoenix, and formerly co-founder of Numenta with Jeff Hawkins, who's been on this podcast, and Donna Dubinsky. From his early work on hierarchical temporal memory to recursive cortical networks to today, Daleep's always sought to engineer intelligence that is closely inspired by the human brain. As a side note, I think we understand very little about the fundamental principles underlying the function of the human brain, but the little we do know gives hints that may be more useful for engineering intelligence than any idea in mathematics, computer science, physics, and scientific fields outside of biology. And so the brain is a kind of existence proof that says it's possible, keep at it. I should also say that brain-inspired AI is often overhyped and used as fodder just as quantum computing for, uh, marketing speak. But I'm not afraid of exploring these sometimes overhyped areas since where there's smoke, there's sometimes fire. Quick summary of the ads. Three sponsors, Babbel, Raycon earbuds, and MasterClass. Please consider supporting this podcast by clicking the special links in the description to get the discount. It really is the best way to support this podcast. If you enjoy this thing, subscribe on YouTube, review it with five stars on Apple Podcasts, support on Patreon, or connect with me on Twitter @LexFridman. As usual, I'll do a few minutes of ads now and never any ads in the middle that can break the flow of the conversation. This show is sponsored by Babbel, an app and website that gets you speaking in a new language within weeks. Go to babbel.com and use code LEX to get three months free. They offer 14 languages, including Spanish, French, Italian, German, and yes, Russian. Daily lessons are 10 to 15 minutes, super easy, effective, designed by over 100 language experts. Let me read a few lines from the Russian poem, (Russian) by Alexander Blok that you'll start to understand if you sign up to Babbel. (Russian) . Now, I say that you'll only start to understand this poem because Russian starts with language and ends with vodka. Now, the latter part is definitely not endorsed or provided by Babbel and will probably lose me the sponsorship, but once you graduate from Babbel, you can enroll in my advanced course of late night Russian conversation over vodka. I have not yet developed an app for that. It's in progress. So get started by visiting babbel.com and use code LEX to get three months free. This show is sponsored by Raycon earbuds. Get them at buyraycon.com/lex. They've become my main method of listening to podcasts, audiobooks, and music when I run, do push-ups and pull-ups, or just living life. In fact, I often listen to brown noise with them when I'm thinking deeply about something. It helps me focus. They're super comfortable, pair easily, great sound, great bass, six hours of playtime. I've been putting in a lot of miles to get ready for a potential ultra-marathon and listening to audiobooks on World War II. The sound is rich and really comes in clear. So again, get them at buyraycon.com/lex. This show is sponsored by MasterClass. Sign up at masterclass.com/lex to get a discount and to support this podcast. When I first heard about MasterClass, I thought it was too good to be true. I still think it's too good to be true. For 180 bucks a year, you get an all-access pass to watch courses from, to list some of my favorites, Chris Hadfield on space exploration, Neil deGrasse Tyson on scientific thinking and communication, Will Wright, creator of SimCity and Sims, on game design. Every time I do this read (laughs) , I really want to play a city builder game. Carlos Santana on guitar, Garry Kasparov on chess, Daniel Negreanu on poker, and many more. Chris Hadfield explaining how rockets work and the experience of being launched into space alone is worth the money. By the way, you can watch it on basically any device. Once again, sign up at masterclass.com to get a discount and to support this podcast. And now, here's my conversation with Daleep George.

Lex Fridman 4:50 Do you think we need to understand a brain in order to build it?

Dileep George 4:54 Yes, if you want to build a brain, we definitely need to understand how it works. So Blue Brain or Henry Markram's project, uh, is trying to build a brain without understanding it. Like, you know, just trying to, uh, put details of the brain from neuroscience experiments into a giant simulation, uh, by putting more and more neurons, more and more details. Uh, but that is not going to work, uh, because when it doesn't perform as, uh, what you expect it to do, then what do you do? Do you just keep adding more details? How do you debug it? So it's a... So unless you understand, unless you have a theory about how the system is supposed to work, how the pieces are supposed to fit together, what they're going to contribute, you can't, you can't build it.

Lex Fridman 5:42 At the functional level. Understand. So can you actually linger on and describe the Blue Brain project? It's kinda fascinating, uh, principle and idea to try to simulate the brain. As we're talking about the human brain, right?

Dileep George 5:57 Right. Human brains and rat brains or cat brains have lots in common. The cortex, uh, the neocortex structure is very similar. So initially they were trying to just simulate a cat brain. Uh, and, uh-

Lex Fridman 6:14 To understand the nature of evil.

Dileep George 6:16 (laughs) And understand the nature of evil. Uh- Yeah. ... or, uh, as it happens in most of these simulations, uh, you, you easily get one thing out, which is oscillations, you know? Yeah. If you, you sim- if you simulate a large number of neurons, they oscillate, uh, and, uh, you can adjust the parameters and say that, "Oh, oscillations match the rhythm that we see in the brain," et cetera. But, uh-

Lex Fridman 6:40 Oh, I see. So like, uh, so the idea is, uh, is the simulation at the level of, uh, individual neurons?

Dileep George 6:47 Yeah. So the Blue Brain Project, the original idea as proposed was, um, you, you put very detailed bio physical neurons, uh, bio physical models of neurons, and you interconnect them according to the statistics of connections that we have found from real neuroscience experiments. And then, uh, turn it on and, uh, see what happens. Uh, and, and these neural models are, you know, incredibly complicated in themselves, right, because these neurons are, uh, modeled using, uh, this, uh, idea called Hodgkin-Huxley models, which are about how signals propagate in a cable. And there are a- active dendrites, all those phenomena which those phenomena themselves we don't understand that well. Uh, and then, uh, we put in connectivity, which is part guesswork, part, you know, observed. And of course, if you do not have any theory about how it is supposed to work, uh, we, you, you know, we just have to take whatever comes out of it as, "Okay, this is something interesting." You know?

Lex Fridman 7:55 But in your sense, like these models of the way signal travels along... or like with the axons and all the basic models, that's... they're too crude?

Dileep George 8:04 Oh. Well, actually they are pretty detailed and pretty sophisticated, and they do replicate the neural dynamics. If you take a single neuron and, uh, you, you try to, uh, turn on the different channels, the calcium channels and, uh, uh, the different receptors, uh, and see what effect of, uh, turning on or off those channels are in the neurons' spike output, people have built pretty sophisticated models of that. And, and they are, I, I would say, um, you know, in the regime of correct.

Lex Fridman 8:41 Well, see, the correctness... That's interesting because you've mentioned at several levels, uh, the correctness is measured by looking at some kind of aggregate statistics.

Dileep George 8:49 It, it would be more, uh, the, the spiking dynamics of a single neuron.

Lex Fridman 8:53 Spiking dynamics of a single neuron. Okay.

Dileep George 8:55 Yeah. Uh, and, and yeah, these models, because they are, they are going to the level of mechanism, right?

Lex Fridman 9:00 Mm-hmm.

Dileep George 9:00 So they are basically looking at, uh, okay, what, what is the effect of turning on an ion channel? Uh, and, um, and you can, you can model that using electric circuits, you know? And then so th- their model... S- so it is not just a, uh, function fitting. It is... People are looking at the mechanism underlying it and, uh, putting that in terms of electric circuit, uh, theory, signal propagation theory and, and modeling that. And so those models are sophisticated, but getting a single neuron's model 99% right does not still tell you how to... You know, it would be the analog of getting a transistor model right and now trying to build a microprocessor. Uh, and if you, if you just, uh, observe... you know, if you did not understand how a microprocessor works, uh, but you say, "Oh, I have... I now can model one transistor well, and now I will just try to interconnect, uh, the transistors according to whatever I could, you know, guess from the, the experiments and try to simulate it," um, then it is very unlikely that you will produce a functioning microprocessor. Um, you want to... you know, when you want to, uh, produce a functioning microprocessor, you want to understand Boolean logic, how does... how do the, the gates work, all those things, and then, you know, understand how do those gates get implemented using transistors.

Lex Fridman 10:24 Yeah. There's actually... I remember... This reminds me, there's a paper, maybe you're familiar with it, that I remember going through in a reading group that approaches a microprocessor from a perspective of a neuroscientist.

Dileep George 10:36 Yeah.

Lex Fridman 10:36 I think it, it basically... it, uh, uses all the tools that we have of neuroscience to try to understand like as if we just aliens showed up to study computers, uh-

Dileep George 10:47 Yeah.

Lex Fridman 10:48 ... and, and to see if, if those tools can be used to get any kind of sense of how the microprocessor works. And I think the final... the takeaway from th- at least this initial, uh, exploration is that (laughs) we're screwed. There's no way that the tools of neuroscience would be able to get us to anything, like not even Boolean logic. I mean, it's just a- a- any aspect of the architecture, of the, uh, function, of the processes involved, uh, the, the clocks, the, the timing, all that, you can't figure that out from the tools of neuroscience.

Dileep George 11:24 Yeah. So I'm very familiar with this, this particular paper.

Lex Fridman 11:27 Yeah.

Dileep George 11:27 I think it was, uh, called, um, Can, uh, a Neuroscientist Understand a Microprocessor or-

Lex Fridman 11:34 Yeah.

Dileep George 11:34 ... something like that. Following the methodology in that paper, even ele- electrical engineer would not understand microprocessors. So I couldn't have so... (laughs)

Lex Fridman 11:43 (laughs)

Dileep George 11:43 (laughs) So I co- so I don't think it is that bad in the sense of saying, um, neuroscientists do find valuable things, uh, by observing the brain. Uh, they, they do find good insights. Um, but those insight cannot be put together just as a simulation. You have to, you have to investigate what are the computational underpinnings- pinnings of those findings? How do, do all of them fit together from an information processing perspective? You have to, you have to... Somebody has to, uh, painstakingly put those things together and build hypotheses. Um, so I don't want to diss all of neuroscience as saying, "Oh, they're not finding anything." No, that, you know, that, uh, that paper almost went to that level of, uh, uh, neuroscientists will never understand. Uh, no, that- that's not true. I think they do find lots of useful things, but it has to be put together in a, in a computational framework.

Lex Fridman 12:40 Yeah. I mean, but, you know, just...... the AI systems will be listening to this podcast a hundred years from now and, and they will probably, uh, th- there's some non-zero probability they'll find your words laughable. It's like, "Oh, remember-"

Dileep George 12:54 (laughs) .

Lex Fridman 12:54 "... human, humans thought they understood something about the brain, they were totally clueless." There's a sense about neuroscience that we may be in the very, very early days of understanding, uh, the brain. But I mean, that's one perspective. In your perspective, how far are we into understanding, uh, any aspect of the brain? So the, the, the dynamics of the individual neuron communication to the how when they, in, in a collective sense, how they're able to store information, transfer information, how intelligence then emerges, all that kind of stuff. Wh- where are we on that timeline?

Dileep George 13:35 Yeah. So, you know, timelines are very, very hard to predict.

Lex Fridman 13:38 (laughs) .

Dileep George 13:38 And we can, of course, be wrong. Uh, and you can be wrong in, on either side. Uh, you know, we know that, uh, you know, when we look back, uh, the first flight was-

Lex Fridman 13:49 Yeah.

Dileep George 13:49 ... in 1903. Uh, in 1900, there was a New York Times article on flying machines that do not fly. And, uh, and m- m- you know, humans might not fly for another 100 years. That was what that article, uh, stated. And, uh, so, but no, they, they flew three years after that. So this, you know, it's very hard to, um, so-

Lex Fridman 14:11 Well, and on that point, one of the Wright brothers, uh, I think two years before, uh, said that, uh, like he said like some number like 50 years, he, he, he has become convinced that it's, it's, um, it's impossible.

Dileep George 14:29 Even during their experimentation, yeah, yeah, yeah.

Lex Fridman 14:31 Yeah. Do you... I mean, that's a tribute to when, uh, that's like the entrepreneurial battle of like depression of going through just like-

Dileep George 14:38 Right.

Lex Fridman 14:38 ... thinking i- there's, this is impossible.

Dileep George 14:40 Right.

Lex Fridman 14:40 But there, yeah, there's something, even the person that's in it is not able to see-

Dileep George 14:45 Right.

Lex Fridman 14:45 ... uh, estimate correctly.

Dileep George 14:47 E- exactly. But I can, I can tell from the point of, you know, objectively what are the things that we know about the brain and how that can be used to build AI models which can then go back and inform how the brain works. Um, so my way of understanding the brain would be to basically say look at the insights neuroscientists have found, understand that from, uh, a computational angle, information processing angle, build models using that. And then building th- that model which, which functions, you know, which is a functional model, which is, which is doing the task that we want the model to do. It is not just trying to model a phenomena in the brain. It is, it is trying to do what the brain is trying to do on the, uh, on the whole, uh, functional level. And building that model will help you fill in the missing pieces that, you know, biology just gives you the hints and building the model, you know, fills in the rest of the, the pieces of the puzzle. And then you can go and connect that back to biology and say, "Okay, now it makes sense that this part of n- uh, the brain is, uh, doing this or this layer in the cortical circuit is doing this." Uh, and, and, and then continue this iteratively because n- now that will inform new experiments in neuroscience. And, of course, you know, building the model and verifying that in the real world will u- will also tell you more about does the model actually work? Uh, and you can refine the model, find better ways of putting these neuroscience insights together.

Lex Fridman 16:19 Mm-hmm.

Dileep George 16:19 So, so I would say it is, it is, you know, it... So you, neuroscientists alone, just from experimentation, will not be able to build a model of the, of the brain, uh, or a functional model of the brain. So we, you know, there, there's, uh, lots of efforts which are very impressive efforts in collecting more and more connectivity data from the brain. Uh, you know, how, how are the microcircuits of the brain connected with each other?

Lex Fridman 16:45 Those are beautiful, by the way.

Dileep George 16:47 Those are beautiful. Uh.

Lex Fridman 16:49 (laughs) .

Dileep George 16:49 And at the same time, those, those do not itself, um, by themselves convey the story of how does it work.

Lex Fridman 16:57 Yeah.

Dileep George 16:57 Uh, and, and somebody has to understand, okay, why are they connected like that and what, what are those things doing? Uh, and, and we do that by building models in AI using hints from neuroscience and, and repeat the

Dileep George 17:11 cycle.

Lex Fridman 17:12 So what aspect of the brain are useful in this whole endeavor? Which, by the way, I should say you're, you're both a neuroscientist and an AI person. Uh, I guess the dream is to both understand the brain and to build AGI systems. So you're, it's like an engineer's perspective of trying to understand the brain. So what aspects of the brain, uh, functionally speaking, like you said, do you find interesting?

Dileep George 17:38 Yeah. Quite a lot of things. All right. So one is, um, you know, if you look at the visual cortex, uh, uh, and, and, you, you know, the, the, uh, visual cortex is, is a large part of the brain. Uh, I forgot the ex- exact fraction, but it is, it's a, it- a huge part of our brain area is, uh, occupied by just, just vision. Um, so vision, visual cortex is not just a feedforward cascade of neurons. Um, uh, there are a lot more feedback connections in the brain compared to the feedforward connections. And, and it is surprising to the level of detail neuroscientists have actually studied this. If you, if you go into neuroscience literature and poke around and ask, you know, have they studied what will be the effect of poking a neuron in, uh, level IT, uh, in level V1 and, uh, um, have they studied that? Uh, and you'll say, "Yes, they have studied that." (laughs) .

Lex Fridman 18:35 (laughs) . So every pos- every possible combination (laughs) -

Dileep George 18:37 Yeah. I mean, it's, it's, uh, it's not a random exploration at all. It's, uh, very hypothesis-driven, right?

Lex Fridman 18:43 Yeah.

Dileep George 18:43 They, they, they are very, uh, experimental neuroscientists are very, very systematic in how they probe the brain.

Lex Fridman 18:48 Yeah.

Dileep George 18:49 Uh, because experiments are very costly to conduct. They take a lot of preparation, they, they need a lot of control. So they, they are very hypothesis-driven-

Lex Fridman 18:57 Yeah.

Dileep George 18:57 ... in how they probe the brain. And, um, often what I find is that when we have a question in, um, in AI, uh, about have, has anybody probed, uh, probed how lateral connections in the brain works? And when you go and read the literature, yes, people have probed it and people have probed it very systematically. And, and they have hypotheses about how those lateral connections are-... supposedly contributing to visual processing. Uh, but, of course, they haven't built very, very functional detail models of it.

Lex Fridman 19:28 By the way, how do they... In those studies, sorry to interrupt, uh, do they, do they stimulate, like, a neuron in one particular area of the visual cortex and then see how the travel, uh, the signal travels kind of thing?

Dileep George 19:39 Fascinate, very, very fascinating experiments, right? You know, so I can, I can give you one example I was impressed with. Um, this is, um... So before going to that, let me, like, let me give you a, uh, you know, a, uh, overview of how the, the layers in the cortex are organized, right?

Lex Fridman 19:54 Sure.

Dileep George 19:54 Uh, visual cortex is organized into roughly four hierarchical levels. Okay. So, uh, V1, V2, V4, IT and in V1-

Lex Fridman 20:03 What happened to V3?

Dileep George 20:04 Uh, well, yeah, there is another pathway. Uh-

Lex Fridman 20:06 Okay.

Dileep George 20:07 Okay. So there is a... This is-

Lex Fridman 20:08 (laughs)

Dileep George 20:08 This, I'm, I'm talking about just object recognition pathway.

Lex Fridman 20:10 All right, cool.

Dileep George 20:11 Okay. Uh, and then, um, in V1 itself, um, i- uh, so it's, there is a very detailed micro circuit in V1 itself. There is, there is organization within a level itself. Um, the cortical sheet is organized into, uh, you know, multiple layers and there are columnar structure. And, and this, this l- layer wise and columnar structure is repeated in V1, V2, V4, uh, uh, IT, all of them, right? Uh, and, and the connections between these layers within a level w- you know, in V1 itself, there are six layers, roughly. And the connections between them, there is a particular structure to them. Uh, and, um, now... So one example of, uh, an experiment, uh, uh, people did is, when I, when you present a stimulus, uh, which is, um, let's say requires, um, separating the foreground from the background of an object. So it is a, it's a textured triangle on a textured background. Uh, and, um, you can check, does the surface settle first or does the contour settle first?

Lex Fridman 21:19 Settle?

Dileep George 21:20 Settle in the sense that the... So w- when you fi- finally form the percept of the, of the, uh, triangle-

Lex Fridman 21:27 Mm-hmm.

Dileep George 21:28 ... you understand where the contours of the triangle are, and you also know where the inside of the triangle is, right? That's when you form the final percept. Um, now you can ask, what is the dynamics of forming that final percept?

Lex Fridman 21:42 Mm-hmm.

Dileep George 21:42 Um, do the, uh, d- do the neurons, um, first find the edges and converge on where the edges are-

Lex Fridman 21:51 Mm-hmm.

Dileep George 21:51 ... and then they find the inner surfaces? Or does it go the other way around?

Lex Fridman 21:55 The other way around.

Dileep George 21:56 Um, so-

Lex Fridman 21:57 So what's the answer?

Dileep George 21:58 Uh, in this case, it, it turns out that it fi- first settles on the edges, it, it converges on the edge hypothesis first, and then the, the surfaces are filled in from the edges to the inside.

Lex Fridman 22:11 That's fascinating.

Dileep George 22:11 Uh, and, and the detail to which you can study this, it's, it's amazing that you can actually not only find, um, the temporal dynamics of when this happens. Uh, uh, and then you can also find which layer in the, you know, in V1, which layer is encoding, uh, the edges, which layer is encoding the surfaces, and, um, which layer is encoding the feedback, which layer is encoding the feed-forward, and what, what's the combination of them that produces the final percept.

Lex Fridman 22:40 Mm-hmm.

Dileep George 22:41 Um, and these kinds of experiments stand out when you try to explain illusions. Um, one, one example of a favorite illusion of mine is the Kanizsa triangle. I don't know whether you're familiar with this one. So this is, um, uh, this is an example where it's a triangle, uh, but, you know, the, the corners of the, only the corners of the triangle are shown in the stimulate, stimulus.

Lex Fridman 23:03 Mm-hmm.

Dileep George 23:03 Uh, so they, they look like kind of Pac-Man. Uh-

Lex Fridman 23:06 Oh, the black Pac-Man. Yeah, yeah, yeah.

Dileep George 23:08 Exactly. Uh, and then-

Lex Fridman 23:09 Then you start to see.

Dileep George 23:10 ... your visual system hallucinates the edges.

Lex Fridman 23:13 Yeah.

Dileep George 23:13 Um, and you can, you, you know, you, when you look at it, you will see a faint edge, right?

Lex Fridman 23:17 Yeah.

Dileep George 23:17 And you can go inside the brain and look, you know, do actually neurons signal the presence of this edge? And, and if they signal, how do they do it? Because they are not receiving anything from the input.

Lex Fridman 23:31 Mm-hmm.

Dileep George 23:31 And the, the input is black for those neurons, right?

Lex Fridman 23:34 Yeah.

Dileep George 23:35 Uh, so how do they signal it? When does the signaling happen? You know, does it, you know... So, so if a real contour is present in the input, then the signa- you know, the neurons immediately signal, oh, okay, there is a, there is an edge here. When, when it is an illusory edge, um, it is clearly not in the input.

Lex Fridman 23:53 Right.

Dileep George 23:53 It is coming from the context. So those neurons fire later. And, and you can say that, okay, these are... It's the feedback connections that is causing them to fire. Uh, and, and they happen later and you can f- uh, find the dynamics of them. Oh, so, so these studies are-

Lex Fridman 24:10 Hm.

Dileep George 24:10 ... pretty impressive and, and very detailed.

Lex Fridman 24:13 So by the way, just, uh, just a sta- step back.

Dileep George 24:16 Yeah.

Lex Fridman 24:17 You said, uh, that there may be more feedback connections than feed-forward connections.

Dileep George 24:21 Yeah.

Lex Fridman 24:22 Uh, first of all, this is just for like machine learning folks.

Dileep George 24:27 Yeah.

Lex Fridman 24:27 I mean, that for, that's crazy that there's all these feedback connections. I mean, we often think about, uh, thank, thanks to deep learning, you s- you start to think about, um, the, the human brain as a kind of feed-forward mechanism.

Dileep George 24:43 Right.

Lex Fridman 24:44 Uh, so what the heck are these feedback connections?

Dileep George 24:48 Yeah. Yeah.

Lex Fridman 24:49 What's their dy- what's the dynamics? Or what are we supposed to think about them?

Dileep George 24:54 Yeah. So this is, this fits into a very beautiful picture about how the brain works, right? Um, so the, the beautiful picture of how the brain works is that our brain is building a model of the world. Uh, I know. So our visual system is building a model of how objects behave in the world. And, and we are constantly projecting that model back onto the world. So what we are seeing is not just a feed-forward thing that just gets interpreted in, in a feed-forward part.

Lex Fridman 25:23 Mm-hmm.

Dileep George 25:23 It... We are constantly projecting our expectations onto the world. And, and what the final percept is a combination of what we project onto the world-... uh, combined with what the actual sensory input is.

Lex Fridman 25:36 Mm-hmm.

Dileep George 25:36 Uh-

Lex Fridman 25:36 Al- almost like trying to calculate the difference and then trying to interpret the difference.

Dileep George 25:41 Yeah. It's- it's- um, I wouldn't put it as calculating the difference. It's more like, what is the best explanation for the- the input stimulus based on the model of the world I have?

Lex Fridman 25:52 Got it. Got it. And that's where all the illusions come in. And that's... But that's- that's an incredibly efficient so- uh, uh, efficient process. So the feedback mechanism, it just helps you constantly, uh... Yeah. So hallucinate how the world should be based on your world model-

Dileep George 26:08 Exactly.

Lex Fridman 26:08 ... and then just looking at, uh, if there's novelty, uh, like, trying to explain it.

Dileep George 26:16 Yeah.

Lex Fridman 26:16 Like that... Hence, that's why movement... We detect movement really well. There's all these kinds of things. And that- this is, like, at all different levels of the c- cortex, you're saying, that-

Dileep George 26:26 Everything.

Lex Fridman 26:26 ... this happens at the lowest level, at the highest level.

Dileep George 26:29 Exactly. Yes. Yeah. Fee- in fact, feedback connections are more prevalent in- everywhere in the cortex. And- and, um, so one way to think about it, and there's a lot of evidence for this, is inference, um... So, you know, so basically if you have a model of the world and when- when some evidence comes in, what we are doing is inference, right?

Lex Fridman 26:49 Mm-hmm.

Dileep George 26:49 You are trying to now explain this evidence using your model of the world.

Lex Fridman 26:55 Yep.

Dileep George 26:55 And this inference includes projecting your model onto the evidence and, uh, taking the evidence, uh, back into the model and- and doing an iterative procedure. Um, and, uh, this iterative procedure is what happens using the feedforward feedback propagation. Uh, and feedback affects what you see in the world and, you know, it also affects feedforward propagation. And examples are everywhere. We- we see these kinds of things everywhere. The idea that there can be multiple competing hypotheses, uh, in our model trying to explain the same evidence, and then you have to kind of make them compete. And one hypothesis will explain away the other hypothesis through this competition process.

Lex Fridman 27:42 D- w- wait, what? So y- you have competing models of the world that try to explain... What do you mean by explain away?

Dileep George 27:50 So this is a classic example in, uh, uh, graphical models, probabilistic models. Um, so if you-

Lex Fridman 27:56 Uh, what are those?

Dileep George 27:57 Uh, okay. Um... (laughs)

Lex Fridman 28:00 (laughs) Oh, I think it's useful to mention because we'll talk about them more.

Dileep George 28:04 Yeah. Yeah. So neural networks are one class of machine learning models. Um, uh, you know, you have distributed set of, uh, nodes, which are called the neurons. A- you know, each one is doing a dot product. And you can- you can approximate any function using this, uh, multilevel, uh, network of neurons. So that's, uh, uh, a class of models which are used for- useful for function approximation. There is another class of models in machine learning, uh, called probabilistic graphical models. And you can think of them as each node in that model is variable, which is- which is talking about something. You know, it- it can be a variable representing, is- is an edge present in the input or not? Uh, and at the top of the, uh, uh, network, a node can be, uh, representing, is there an object present in the, uh, world or not? And- and then... So it can... It is- it is another way of encoding knowledge.

Lex Fridman 29:03 Mm-hmm.

Dileep George 29:05 And, uh, um, and then you... Once you encode the knowledge, you can, uh, do inference in the right way, you know. How... What is the best way to, uh, y- you know, explain some set of evidence using this model that you encoded? You know.

Lex Fridman 29:19 Mm-hmm.

Dileep George 29:19 So when you encode the model, you are encoding the relationship between these different variables. How is the edge connected to my, uh, the model of the object? How is the surface connected to the model of the object? Um, and then, um, of course, this is a very distributed, complicated model. And inference is how do you explain a piece of evidence when- when a set of stimulus comes in? If somebody tells me there is a 50% probability that there is an edge here in this part of the model, how does that affect my belief on whether I should think that there should be a... Is the square present in the image?

Lex Fridman 29:53 Mm-hmm.

Dileep George 29:54 So- so this is the process of inference. So one example of inference is having this explaining away effect between multiple causes. So, uh, graphical models can be used to represent causality in the world. Um, so let's say, um, y- you know, uh, your, uh, alarm, uh, the, uh, at home can be, uh, triggered by a, uh, burglar getting into your house, uh, or it can be triggered by an earthquake.

Lex Fridman 30:24 Yeah.

Dileep George 30:24 Both- both can be causes of the alarm going off. So now, you- you are dri- y- you know, y- you're in your office, you heard burglar alarm going off. You are heading, uh, home, uh, thinking that there's a burglar got in.

Lex Fridman 30:37 Mm-hmm.

Dileep George 30:37 But while driving home, if you hear on the radio that there was an earthquake in the vicinity, now your hype- you know, uh, strength of evidence for, uh, a burglar getting into your house is diminished.

Lex Fridman 30:49 Mm-hmm.

Dileep George 30:50 Because now that- that piece of evidence is explained by the earthquake in- being present. So if you- if you think about these two causes explaining a lower level, uh, variable, which is alarm, now w- w- what we are seeing is that increasing the evidence for some cause ex- you know, there is evidence coming in from below for alarm being present, and initially it was flowing to a burglar being present. But now since somebody, uh... Some- this evi- there is side evidence for this other cause, it explains away this evidence, and it- evidence will now flow to the other cause. Th- this is, you know, two competing causal, uh, things trying to explain the same evidence.

Lex Fridman 31:28 And- and the brain has a similar kind of mechanism-

Dileep George 31:30 Yes.

Lex Fridman 31:31 ... for, uh, for doing so.

Dileep George 31:32 Yes.

Lex Fridman 31:32 That's kind of interesting. I mean, and that-How's

Lex Fridman 31:35 that all encoded (sighs) in the brain? Like, where is the storage of information? Are we talking, just maybe to get it, uh, a little bit more specific, is it in the hardware of the actual connections? Is it in the chemical communication? Is it electrical communication? Do we, do we know?

Dileep George 31:54 So, so this is, you know, a, a paper that we are bringing out soon-

Lex Fridman 31:57 Which one is this?

Dileep George 31:58 Um, this is the cortical microcircuits paper that-

Lex Fridman 32:01 Got it.

Dileep George 32:01 ... I sent you a draft of. Of course, this is, uh, a lot of it is still hypothesis. One hypothesis that a, you can think of a cortical column as encoding a, a concept. A concept, you know, think of it as a, uh, a, um, conc- an example of a concept is, um, is an edge present or not? Or is, is an object present or not? Okay. So it can, you can think of it as a binary variable. A binary random variable. The presence of an edge or not, or the presence of an object or not. So, each cortical column can be thought of as representing that one concept, one variable. And then the connections between these cortical columns are basically encoding the relationship between these random variables.

Lex Fridman 32:45 Mm-hmm.

Dileep George 32:45 And then there are connections within the cortical column. There are... Each cortical column im- is implemented using multiple layers of neurons-

Lex Fridman 32:53 Mm-hmm.

Dileep George 32:53 ... with very, very, very rich, um, structure there, you know. There are thousands of neurons in a cortical column. Uh-

Lex Fridman 33:00 But, but that structure is similar across the different cortical columns.

Dileep George 33:03 Correct.

Lex Fridman 33:04 Yeah.

Dileep George 33:04 Correct. And also, these cortical columns collect, connect to a substructure called thalamus in the, uh, in... So all, all cortical columns pass through this substructure. So, our hypothesis is that, yeah, the connections between the cortical columns implement this, uh, you know, that's where the knowledge is stored about, you know, how these different connups- concepts connect to each other. And then the, the neurons inside this cortical column and in thalamus in combination implement this, um, actual computations needed for inference, which includes explaining a way and competing between the different, uh, hypotheses. Um, and it is all very... So what is amazing is that, um, neuroscientists have actually done ex- experiments to the tune of showing these things. Uh, they might not be putting it in the overall inference framework, but they will show things like if I poke this higher level neuron, uh, it will inhibit, through this complicated loop through the thalamus, it will inhibit this other column.

Lex Fridman 34:09 Mm-hmm.

Dileep George 34:09 Uh, so they will, they will do such (laughs) experiments.

Lex Fridman 34:11 But do they use terminology of concepts, for example? So, so you're-

Dileep George 34:16 No.

Lex Fridman 34:16 I mean, (sighs) uh, is it, uh, is it something where it's easy to anthropomorphize and think about concepts like, uh, you start moving into logic-based kind of reasoning systems. So, um, are we to think of concepts in that kind of way? Or is it s-, uh, is it a lot messier, a lot more gray area? You know, even, even more gray, even more messy than, uh, the artificial neural network kinds-

Dileep George 34:47 Um-

Lex Fridman 34:47 ... kinds of abstractions?

Dileep George 34:48 Easiest way to think of it as a variable, right? It's a binary variable.

Lex Fridman 34:51 Mm-hmm.

Dileep George 34:52 Which is showing the presence or absence of something.

Lex Fridman 34:55 S- but I guess what I'm asking is, is that something, uh, that... Are we supposed to think of something that's human interpretable of that something?

Dileep George 35:04 It doesn't need to be. It doesn't need to be human interpretable. There is no need for it to be human interpretable. Uh, but it's, it's almost like, um, you, you will be able to find some interpretation of it, uh, because it is connected to the other things-

Lex Fridman 35:21 Yeah. So it's-

Dileep George 35:21 ... that you know about.

Lex Fridman 35:22 And, and the, the point is it's useful somehow.

Dileep George 35:25 Yeah.

Lex Fridman 35:25 It's useful as a entity in the graph that... in connecting to the other entities that are... let's call them concepts.

Dileep George 35:34 Right.

Lex Fridman 35:34 Okay. So, uh, by the way, what's... Are these the cortical m- microcircuits?

Dileep George 35:40 Correct. These are the cortical microcircuits. You know, that's what neuroscientists use to talk about the circuits in, in, uh, within a level of the cortex. So you can think of, you know, let's think of in neural network, you know, artificial neural network terms, you know. People talk about the architecture of the... you know. So how many, how many layers they build. Uh, you know, what is the fan in, fan out, et cetera. That is the macro architecture.

Lex Fridman 36:03 Mm-hmm.

Dileep George 36:03 Uh, so... And then within a layer of the neural network, you can... I- I- you know, the cortical neural network is much more structured with- you know, within a level. There is a lot more intricate, uh, structure there. Uh, but ev- um, even within an artificial neural network, you can think of n- feature detection plus pooling as one-

Lex Fridman 36:23 Mm-hmm.

Dileep George 36:24 ... one level. And so that is kind of a microcircuit. Uh, it, it's much more, uh, complex in the real brain. Uh, and, uh, and... So within a level, whatever is that circuitry within a column of the cortex and between the layers of the cortex, that's the microcircuitry.

Lex Fridman 36:41 I love that terminology. Uh, machine learning people don't use the circuit terminology.

Dileep George 36:45 Right.

Lex Fridman 36:45 But they should. It's, uh, it's a nice... So, okay. Uh, okay. So that's, uh, that, that's the, the cortical microcircuit. So what's interesting about,

Lex Fridman 36:56 uh, w- what can we say? What is the paper that, uh, you're working on, uh, propose about the ideas around these cortical microcircuits?

Dileep George 37:04 So this is a fully functional model for the microcircuits of the visual cortex.

Lex Fridman 37:10 So the, the paper focuses... and your idea and our discussions now is focusing on vision.

Dileep George 37:15 Yeah.

Lex Fridman 37:16 The, uh, visual cortex. Okay.

Dileep George 37:18 Yeah. I-

Lex Fridman 37:19 This is a model, this is a full model. This is-

Dileep George 37:21 Th- this-

Lex Fridman 37:21 This is how vision works.

Dileep George 37:22 Uh, well, this is, this is a, yeah, model of-

Lex Fridman 37:25 A hypothesis.

Dileep George 37:25 Yeah.

Lex Fridman 37:25 A hypothesis.

Dileep George 37:26 So, okay, so let me, let me step back, uh, a bit. Um, so we looked at neuroscience for insights on how to build a vision model.Right. And, and, and we synthesized all those insights into a computational model. This is called the recursive cortical network model that we, we used for breaking captchas and, uh, and we are using the same model for robotic picking and, uh, uh, tracking of objects.

Lex Fridman 37:50 And that, again, is a vision system.

Dileep George 37:52 That's a vision system.

Lex Fridman 37:52 Com- computer vision system.

Dileep George 37:54 That's a computer vision system.

Lex Fridman 37:55 Takes in images and outputs what?

Dileep George 37:59 On one side, it outputs the class of the image, uh, and also segments the image. Uh, and you can also ask it further queries, "Where is the edge of the object? Where is the interior of the object?" So-

Lex Fridman 38:10 Got it.

Dileep George 38:11 ... so it's a, it's a model that you build to answer multiple questions. So you're not trying to build a model for just classification or just recommendation et cetera. It's a, it's a, it's a joint model that can do multiple things. Um, and, um, so, so that's the model that we built using insights from neuroscience. And some of those insights are, what is the role of feedback connections? What is the role of lateral connections? Uh, so all those things went into the model. The m- the model actually uses feedback connections.

Lex Fridman 38:40 All these ideas from neu- from neuroscience.

Dileep George 38:43 Yeah.

Lex Fridman 38:43 Uh, so what, what, what the heck is a r- recursive cortical network? Like what, what are the architecture approaches, interesting aspects here, which is essentially a brain-inspired approach to computer vision?

Dileep George 38:56 Yeah. So there are multiple layers to this question. Again, go from the very, very top and then zoom in. Okay?

Lex Fridman 39:03 Mm-hmm.

Dileep George 39:03 So one important thing, constraint that went into the model is that you should not think vision, think of vision as something in isolation. We should not think perception as something, as a pre-processor for cognition.

Lex Fridman 39:18 Mm-hmm.

Dileep George 39:18 Perception and cognition are interconnected. And so you should not think of one problem in separation from the other problem. Um, and so that means if you finally want to have a system that understand concepts, uh, about the world and can learn a, you know, very conceptual model of the world, and can reason and connect to language, all of those things, you need to, you need to have, think all the way through and make sure that your perception system is compatible with your cognition system and language system and all of them. And one aspect of that is top-down controllability. Um-

Lex Fridman 39:53 What does that mean-

Dileep George 39:54 So that means-

Lex Fridman 39:55 ... in this context?

Dileep George 39:55 You know, so, so think of, uh, you know, you can close your eyes and think about the details of one object. Right? I can, I can zoom in further and further.

Lex Fridman 40:05 Mm-hmm.

Dileep George 40:05 I can, you know... So, so think of the bottle in front of me, right?

Lex Fridman 40:08 Mm-hmm.

Dileep George 40:08 And, and now you can think about, okay, what the cap of that bottle looks. Uh, and now we can think about what's the texture on that bottle of the, uh, uh, the cap. You know, you can think about, you know, what will happen if, uh, something hits that. Uh, so you can, you can, you can manipulate your visual knowledge in, uh, cognition-driven ways.

Lex Fridman 40:32 Yes.

Dileep George 40:32 Uh, and so this top-down controllability, uh, and being-

Lex Fridman 40:36 Ah.

Dileep George 40:36 ... able to simulate scenarios in the world.

Lex Fridman 40:40 So you're not just a passive, uh, player in this perception game, you, you can-

Dileep George 40:46 Got it.

Lex Fridman 40:46 ... you can control it. You can-

Dileep George 40:47 Got it.

Lex Fridman 40:47 You g- you have imagination.

Dileep George 40:49 Correct, correct. So, so, so basically, you know, basically having a generative network-

Lex Fridman 40:54 Yeah.

Dileep George 40:54 ... uh, which is a model, and, and it is not just some arbitrary generative network. It has to be, it has to be built in a way that it is controllable top-down.

Lex Fridman 41:03 Mm-hmm.

Dileep George 41:03 It is, it is not just trying to generate a whole picture at once. Uh, you know, it's not trying to generate photorealistic things of the world. You, you know, you don't have good photorealistic models of the world. Human brains do not have... If I, if I, for example, ask you the question, uh, "What is the color of the letter E in the Google logo?" You have no idea. (laughs)

Lex Fridman 41:23 No idea.

Dileep George 41:24 Although you have seen it millions of times. (laughs)

Lex Fridman 41:25 I think, I think-

Dileep George 41:25 Or not millions of time, hundreds of times. (laughs) So...

Lex Fridman 41:28 Yeah.

Dileep George 41:28 Uh, so it's not... Our model is not photorealistic. But, but e- it is, but it has other properties that we can manipulate it, uh, in the, uh... And you can think about filling in a different color in that logo. You can think about expanding the, the letter E.

Lex Fridman 41:42 Yeah.

Dileep George 41:42 You know, you, you can see what'll in... So you can imagine the consequence of, uh, you know, actions that you have never performed. So, so these are the kind of characteristics the generative model need to have. So this is one constraint that went into our model. Like, you know, so this is... When you read the, just the perception side of the paper, it is not obvious that this was a constraint into the, int- that went into the model, this top-down controllability of the generative model.

Lex Fridman 42:06 Uh, so what, what is, uh, top-down controllability in a model look like? It's a really interesting concept, fascinating concept.

Dileep George 42:14 Yeah.

Lex Fridman 42:15 What does that... Is that the recursed, recursiveness gives you that? Or how, how do you, how do you do it?

Dileep George 42:20 It's, um, quite a few things. It's like, what, what does the model factor, uh, factorize? You know, what are the, what is the model representing as different pieces in the puzzle? Like, you know, so-

Lex Fridman 42:30 Mm-hmm.

Dileep George 42:30 ... so in the RCN, uh, network, it, it thinks of the world, you know, so for example, the background of an image is modeled separately from the foreground of the image.

Lex Fridman 42:41 Got it.

Dileep George 42:41 So, so the, the objects are separate from the background. They're different entities.

Lex Fridman 42:45 So there's a kind of segmentation that's built in fundamentally to this structure.

Dileep George 42:49 Correct. And, and, and then even that object is composed of parts. And also, and another one is the, the shape of the object, uh, is differently modeled from the texture of the object.

Lex Fridman 43:02 Got it. So there's like these, um, uh, I've been... U- you know who Francois Chollet is?

Dileep George 43:08 Yes. Yeah.

Lex Fridman 43:08 Uh, he, he's... (laughs) So there's, uh, he, he developed this, like, IQ test type of thing for-

Dileep George 43:14 Right.

Lex Fridman 43:14 ... Arc Challenge for... And, uh, it's kind of cool that there's, um, these concepts, priors, that he defines that you bring to the table in order to be able to reason about basic shapes and things in an IQ test.

Dileep George 43:28 Right.

Lex Fridman 43:28 So here you're making it q- quite explicit that here, here are the things that you should be... The, these are like distinct things that you should be able to, uh, model in this.

Dileep George 43:40 Yeah. Keep in mind that you, you can derive this from much more general principles. It doesn't... You don't need to explicitly put it as, oh, objects versus foreground versus background, uh, the surface versus structure. No, these are, these are derivable from-... uh, more fundamental principles of how, you know, w- what's the property of continuity of natural signals.

Lex Fridman 44:02 What's the property of continuity of natural signals?

Dileep George 44:05 Yeah.

Lex Fridman 44:06 By the way, that sounds very poetic. But yeah-

Dileep George 44:08 (laughs) .

Lex Fridman 44:08 ... uh, so you're saying that's a, a, th- there's s- some low level properties from which emerges the idea that shapes should be different than-

Dileep George 44:16 Exa-

Lex Fridman 44:16 ... like, uh, there should be parts of an object, there should be, I mean-

Dileep George 44:20 Exactly.

Lex Fridman 44:20 ... kind of like Francoise talk- I mean, there's an objectness. There's all-

Dileep George 44:23 Got it.

Lex Fridman 44:23 ... these things that it's kinda crazy that we humans, uh, I guess evolved to have because it's useful for us-

Dileep George 44:30 Correct.

Lex Fridman 44:31 ... to perceive the world.

Dileep George 44:32 Correct. Correct.

Lex Fridman 44:32 Yeah.

Dileep George 44:32 And it, it derives mostly from the properties of natural signals. And-

Lex Fridman 44:36 Yeah.

Dileep George 44:37 ... and so, um-

Lex Fridman 44:39 Natural signals. So natural signals are the kind of things we'll perceive in an, in the natural world.

Dileep George 44:44 Correct.

Lex Fridman 44:45 I don't know. I don't, I don't know why that sounds so beautiful. Natural signals, yeah.

Dileep George 44:48 As opposed to a QR code, right? Which is an artificial signal that we created. Humans are not very good at classifying QR codes. We are very good at saying w- well, something is a cat or a dog.

Lex Fridman 44:58 Yeah.

Dileep George 44:58 But not very good at, you know, w- classifi- where computers are very good at classifying QR codes. Um, so our, our visual system is tuned for natural signals. Uh, and there are fundamental assumptions in the architecture that are derived from natural signals, uh, m- properties.

Lex Fridman 45:15 I wonder when you take, uh, hallucinogenic drugs, does that go into natural-

Dileep George 45:20 (laughs) .

Lex Fridman 45:20 ... or is that closer to the QR code?

Dileep George 45:22 Uh, it-

Lex Fridman 45:23 That's a wh-

Dileep George 45:23 ... it's still natural.

Lex Fridman 45:23 It's still natural? Yeah.

Dileep George 45:25 Because it's, it, it is still operating y- using your brains.

Lex Fridman 45:28 By the way on that, on that topic, I, I mean, I haven't been following, I think, they're becoming legalized in certain... I can't wait until they become legalized to a degree that you c- like, vision science researchers could study it.

Dileep George 45:40 Yeah.

Lex Fridman 45:40 Just like, through, through medical, chemical ways modify... there could be ethical concerns but modif- that's another way to study the brain is to be, be able to chemically modify it. There's probably, um, probably very long a way to, to figure out how to do it ethically.

Dileep George 45:59 Yeah. But I th- I think there are studies on that already. Uh-

Lex Fridman 46:02 Already?

Dileep George 46:02 ... yeah, I think so. Uh-

Lex Fridman 46:04 But you keep-

Dileep George 46:04 ... because it's, it's not unethical to give, uh, it to rats. So-

Lex Fridman 46:08 Oh, that's true. That's true. (laughs) .

Dileep George 46:11 (laughs) .

Lex Fridman 46:11 There's a lot, there's a lot (laughs) , there's drugged up rats out there. Okay.

Dileep George 46:14 Yeah.

Lex Fridman 46:14 Cool. Sorry, sorry to have... so okay, so there's, uh, uh, so there's these, uh, low level, uh, things from natural signals that, uh, that, that, uh, they can, can-

Dileep George 46:26 From which these properties will emerge.

Lex Fridman 46:28 Yes.

Dileep George 46:29 Uh, but it is still a very hard problem on how to encode that, you know. So you don't, you know, there is no, uh... so, uh, you mentioned, um, the, the, uh, the priors, uh, Francoise wanted to encode in ca- uh, in the, uh, abstract reasoning challenge.

Lex Fridman 46:45 Mm.

Dileep George 46:45 But it is not straight forward how to encode those priors. Um, so, so some of those, uh, challenges, like, you know, the object rec- uh, uh, completion challenges are things that we purely use our visual system to do. It is, uh, it looks like abstract reasoning but it is purely an output of a w- the, the vision system. For example, completing the corners of that kinds of triangle, completing the lines of that kinds of triangle. It's a purely visual system property. It, you know, it's in there, there's no abstract reasoning involved. It, it uses all these priors but it is stored in our visual system in a particular way that is amenable to inference. And, and, and that is one of the things that we tackled in the, you know, it's basically saying, okay, these are the prior knowledge, uh, which, which will be derived from the work, but then how is that prior knowledge represented in the model such that inference, when, when some piece of evidence comes in, can be done very efficiently and in a very distributed way.

Lex Fridman 47:41 Mm-hmm.

Dileep George 47:41 Um, because it is very, there are so many ways of representing knowledge which is not amenable to very quick, i- inference, you know, quick look-ups. Uh-

Lex Fridman 47:50 Mm-hmm.

Dileep George 47:50 ... and so that's one, um, core part of what we tackled in, uh, the RCN model. Um, uh, how do you encode visual knowledge to, uh, to very quick inference. And yeah.

Lex Fridman 48:03 Can you maybe comment on, uh, so w- to folks listening to this, and in general, may be familiar with different kinds of architectures of neural networks, what, what are we talking about with RCN? Uh, w- what are the, what does the architecture look like? What are different components? Is it close to neural networks? Is it far away from neural networks? What does it look like?

Dileep George 48:22 Yeah. So, so you can, uh, think of the delta between the model and the convolutional neural network if, if people are familiar with convolutional neural networks. So convolutional neural networks have this feedforward processing cascade which is called, uh, feature detectors and pooling, and that is repeated in the, in the hierarch- in a, in a, uh, multi-level, uh, system. Um, and if you, if you want to, an intuitive idea of what, what is happening, feature detectors are, uh, you know, detecting interesting occurrences in the input.

Lex Fridman 48:54 Mm-hmm.

Dileep George 48:54 It can be a line, a corner, a, an eye, or a piece of texture, et cetera. And, uh, the pooling n- neurons are doing some local transformation of that and making it invariant to local transformations. So this is what the structure of convolutional neural network is. Um, recursive cortical network has a similar structure when you look at just the feedforward pathway but in addition to that, it is also structured in a way that it is generative so that it c- it can run it backward and combine the forward with the backward.

Lex Fridman 49:27 Mm-hmm.

Dileep George 49:28 Another aspect that it has is it has lateral connections. These lateral connections-

Lex Fridman 49:35 Interesting.

Dileep George 49:36 ... um, which is bet- between... so if you have an edge here and an edge here, it has connections between these edges, it is not just feedforward connections, it is, um, something between these edges which is, uh, the, the nodes are presenting these edges which is to enforce compatibility between them. So otherwise what will happen is that-

Lex Fridman 49:54 Like constraints?

Dileep George 49:55 It's a constraint. It's basically if you, if you do just feature detection followed by pooling...... then your, your transformations in different parts of the visual field are not coordinated.

Lex Fridman 50:06 Mm-hmm.

Dileep George 50:07 Uh, and so you can- you will create a jagged... When you, when you generate from the model, you will create jagged, um, uh, things and uncoordinated transformations. So these lateral connections are enforcing the- the transformations.

Lex Fridman 50:22 Is the whole thing still differentiable?

Dileep George 50:24 Uh, no.

Lex Fridman 50:25 Okay.

Dileep George 50:25 No. (laughs) It's not, it's not trained using, uh, back prop.

Lex Fridman 50:30 Okay. That's really important. So, uh-

Dileep George 50:32 Yeah.

Lex Fridman 50:32 So there's this feed forward, there is feedback mechanisms, there are some interesting connectivity things. It's still layered, like, uh-

Dileep George 50:39 Yes.

Lex Fridman 50:40 ... str-

Dileep George 50:40 There are multiple levels, yes.

Lex Fridman 50:41 Struc- multi- multiple, uh, layers. Okay. Very, very interesting. Uh, and yeah, okay, so the interconnection between, um, adjacent, the connections across service constraints-

Dileep George 50:52 Yeah.

Lex Fridman 50:53 ... that- that keep the thing stable.

Dileep George 50:55 Correct.

Lex Fridman 50:56 Okay. So w- what else?

Dileep George 50:58 Uh, and then there's this idea of doing inference. Uh, a neural network does not do inference on the fly. So an example of why this inference is important is,

Dileep George 51:09 you know, so one of the first applications, uh, that we showed in the paper was to track, uh, text-based CAPTCHAs.

Lex Fridman 51:17 What are CAPTCHAs, by the way? I mean... (laughs)

Dileep George 51:20 Uh... Yeah. (laughs)

Lex Fridman 51:20 By the way, one of the most awesome, like, the people don't use this term anymore as human computation, I think. Uh, I love this term. The guy who created CAPTCHAs, I think, came up with this term.

Dileep George 51:30 Yeah.

Lex Fridman 51:31 I love it. Anyway, uh-

Dileep George 51:33 Yeah. Yeah.

Lex Fridman 51:33 Uh, wh- what, what are CAPTCHAs?

Dileep George 51:35 So CAPTCHAs are those strings that you fill in, uh, when you're, you know, when- if you're open- opening a new account in Google, they show you a picture. Uh, you know, usually, it used to be set of garbled letters, uh, that you have to kind of, uh, figure out what- what- what is that string of characters and type it. And the reason CAPTCHAs exist is because, you know, um, Google or Twitter do not want automatic creation of accounts. You can use a computer to create millions of accounts, uh, and, uh, use that for, you know, nefarious purposes. Uh, so you want to make sure that, to the extent possible, the interaction that, uh, their system is having is with a human. So it's a- it's called a human interaction proof. A CAPTCHA is a human interaction proof.

Lex Fridman 52:22 Yeah.

Dileep George 52:23 Um, so- so this is- CAPTCHAs are, by design, things that are easy for humans to solve, but hard for computers.

Lex Fridman 52:30 Hard for robots, yeah.

Dileep George 52:31 Yeah. Um, so- and text-based CAPTCHAs were- was the one which is prevalent until around 2014. Because at that time, text-based CAPTCHAs were hard for computers to crack. Even now, they are actually, in the sense of an arbitrary text-based CAPTCHA will be unsolvable even now. But with the techniques that we have developed, it can be, you know, you can quickly develop a mechanism that solves all the CAPTCHA, uh-

Lex Fridman 52:58 They- they've probably gotten a lot harder too. The people-

Dileep George 53:01 Correct.

Lex Fridman 53:01 (laughs) They've been getting cleverer and cleverer at generating these text CAPTCHAs.

Dileep George 53:05 Correct. Correct.

Lex Fridman 53:06 Yeah.

Dileep George 53:06 Right.

Lex Fridman 53:06 So, okay, so that was one of the things you've tested it on, is these kinds of CAPTCHAs-

Dileep George 53:10 Yeah.

Lex Fridman 53:10 ... in 2014, '15.

Dileep George 53:12 Correct.

Lex Fridman 53:12 That kind of stuff.

Dileep George 53:13 Right. Right.

Lex Fridman 53:13 So what, uh ... W- what I mean, why- by the way, why CAPTCHAs? Why?

Dileep George 53:18 Yeah. Yeah. Even now, I would say CAPTCHA is a very- very good challenge problem, uh, if you want to understand how human perception works and if you want to build, uh, systems that work like the human brain. Uh, and I wouldn't say CAPTCHA is a solved problem. We have cracked the fundamental defense of CAPTCHAs, but it is not solved in the way that humans solve it. Um, so I can give you an example. I can, um, take a five-year-old child who has just learned characters, uh, and, uh, show them any new CAPTCHA that we create.

Lex Fridman 53:52 Mm-hmm.

Dileep George 53:53 They will be able to solve it. Uh, I can show you pretty much any new CAPTCHA, uh, from any new website. You'll be able to solve it without getting any training examples from that particular style of CAPTCHA.

Lex Fridman 54:05 You're assuming I'm human, yeah.

Dileep George 54:07 Yes, yeah. (laughs) Uh, that's right. So if you are human, uh, other- if you c- otherwise, I will be able to figure that out-

Lex Fridman 54:13 Yeah. Well...

Dileep George 54:14 (laughs) Using this one. But, uh-

Lex Fridman 54:16 This- this whole podcast is just a Turing test.

Dileep George 54:18 That's right. (laughs)

Lex Fridman 54:18 A long- a long Turing test. Anyway, I'm sorry. So yeah.

Dileep George 54:22 Uh...

Lex Fridman 54:22 So humans- humans can figure it out with very few examples.

Dileep George 54:25 Eh, or no training examples.

Lex Fridman 54:27 No training.

Dileep George 54:27 Like, no training examples from that particular style of CAPTCHA.

Lex Fridman 54:30 Yeah.

Dileep George 54:30 Um, and- and so you can, you know, so, uh, even now, this is unreachable for the current deep learning system. So basically, there is no- I don't think a system exists where you can basically say, "Train on whatever you want," and then now say, "Hey, I will show you a new CAPTCHA which I did not show you in- in the- in the training setup. Will the system be able to solve it?" Um, it still doesn't exist. So that is the magic of human perception, yeah. And Doug Hofstadter, uh, put this, uh, very beautifully in, uh, one of his, uh, talks. The- the central problem in AI is what is the letter A?

Lex Fridman 55:10 Mm-hmm.

Dileep George 55:10 If you can- if you can build a system that reliably can detect all the variations of the letter A, you don't even need to go to the- (laughs) the-

Lex Fridman 55:19 The B and the C, yeah.

Dileep George 55:20 Yeah. You don't even need to go to the B and the C or the strings of characters. And, uh, so that- that is the speed at which, you know, with which we, uh, tackle that problem.

Lex Fridman 55:29 W- what does he mean by that? I mean, is- is it, uh, like without training examples, try to figure out the fundamental, uh, elements that make up the letter A-

Dileep George 55:41 Right.

Lex Fridman 55:42 ... in all of its forms?

Dileep George 55:43 In all of its forms. It can be- A can be made with two humans standing leaning against each other, holding hands, you know.

Lex Fridman 55:48 Yeah.

Dileep George 55:49 Uh, and, uh, it can be made of leaves. It can be-

Lex Fridman 55:52 Yeah. You might have to understand, uh, everything about this world in order to understand the letter A. Yeah.

Dileep George 55:57 Exactly. Exactly. So-

Lex Fridman 55:58 So it's common sense reasoning, essentially.

Dileep George 56:00 Right.

Lex Fridman 56:00 Yeah.

Dileep George 56:00 Right. Right. So- so to finally, to really solve, finally to say that we have solved CAPTCHA, uh, you have to solve the whole problem. (laughs)

Lex Fridman 56:10 (laughs) Yeah. Okay. So what- how does, uh, this kind of, uh, RCN architecture help us to get, do a b- better job with that kind of thing?

Dileep George 56:19 Yeah. So, uh, as I mentioned, one of the important things was being able to do inference, being able to dynamically do inference.

Lex Fridman 56:26 Uh, can you, can you, uh, c- can you, uh, clarify what you mean? 'Cause, 'cause you said, like, neural networks don't do inference.

Dileep George 56:33 Yeah.

Lex Fridman 56:33 So, what do you mean by inference in this context, then?

Dileep George 56:36 So, okay. So, in CAPTCHAs, what they do to confuse people is to make these characters crowd together.

Lex Fridman 56:43 Yes.

Dileep George 56:43 Okay? And when you make the characters crowd together, what happens is that you will now start seeing combinations of characters as some other new character-

Lex Fridman 56:52 Mm-hmm.

Dileep George 56:52 ... or, or an existing character. So, you would, you would put an R and N together, it will start looking like an M.

Lex Fridman 56:58 Mm-hmm.

Dileep George 56:58 Uh, and, and, so locally, they are, you know, th- there is very strong evidence for it being, uh, some, uh, incorrect character. But globally, the only explanation that fits together is something that is different from what you get, find locally.

Lex Fridman 57:16 Yes.

Dileep George 57:16 So, th- so, so this is, uh, inference. You are basically taking, uh, local evidence and putting it in the global context, and often coming to a conclusion locally which is conflicting with the local information.

Lex Fridman 57:30 So, actually, (laughs) so you mean inference, like, uh, in the way it's used t- when you talk about reasoning, for example?

Dileep George 57:37 Correct.

Lex Fridman 57:37 Uh, as opposed to, like, inference which is a s- uh, well, neural, with artificial neural networks, which is a single pass through the network.

Dileep George 57:44 Correct.

Lex Fridman 57:44 Okay.

Dileep George 57:45 Correct, correct.

Lex Fridman 57:45 So, like, you're, you're basically doing some basic forms of reasoning.

Dileep George 57:48 Correct.

Lex Fridman 57:48 Like integration of, like, uh, how local things fit into the, the global picture.

Dileep George 57:54 Right.

Lex Fridman 57:54 Got it.

Dileep George 57:54 And, and, and things like explaining a way coming to this one.

Lex Fridman 57:56 Got it.

Dileep George 57:56 Because you are, you are, uh, explaining that piece of evidence, uh, as something else, uh, because globally, that's the only thing that makes sense. Um-

Lex Fridman 58:05 Yeah.

Dileep George 58:06 So now, uh, you can amortize this inference by, you know, in a neural network, if you want to do this, what you, that you can, you can brute force it. You can just show it all combinations of things. (laughs)

Lex Fridman 58:19 Yeah.

Dileep George 58:19 That you want to, you want to, uh, your reasoning to work over.

Lex Fridman 58:22 Mm-hmm.

Dileep George 58:23 And you can, you know, like, just drain the hell out of that neural network, and it will look like it is doing, uh, you know, inference on the fly, but it is, it is really just doing amortized inference. It is because you, you have shown it a lot of these combinations during training time.

Lex Fridman 58:39 Mm-hmm.

Dileep George 58:40 Um, so what you want to do is be able to do dynamic inference, rather than just being able to show all those combinations in the training time. And that's something we emphasized in the model.

Lex Fridman 58:51 What does it mean, dynamic inference? Is that, that has to do with the feedback thing?

Dileep George 58:56 Yes.

Lex Fridman 58:56 Like, what, what is dynamic... I, I, I'm trying to visualize what dynamic inference would be in this case. Like, what is it doing with the input? It's showing the input the first time.

Dileep George 59:07 Yeah.

Lex Fridman 59:07 And is, is, like, what's changing over temporally, over t- what's the dynamics of this inference process?

Dileep George 59:14 So, so you can think of it as you have, um, at the top of the model, the characters that you are trained on.

Lex Fridman 59:21 Yeah.

Dileep George 59:21 They are the causes. They, you are trying to explain the pixels-

Lex Fridman 59:25 Mm-hmm.

Dileep George 59:25 ... using the characters as the causes. The, uh, you know, the characters are the things that cause the pixels.

Lex Fridman 59:33 Yeah, so there's this causality thing. So, the reason you mention causality, I guess, is because there's a temporal aspect to this whole thing.

Dileep George 59:40 In this particular case, the temporal aspect is not important. It is more like when, if, if I turn the character on, the, the pixels will turn on. Uh, yeah, it will be after this, a little bit, but-

Lex Fridman 59:51 Okay.

Dileep George 59:51 Yeah.

Lex Fridman 59:51 So the, it's causality in the sense of, like, a logic causality, like-

Dileep George 59:55 Yeah.

Lex Fridman 59:55 Hence inference. Okay.

Dileep George 59:56 Right, right. The dynamics is that, uh, even though locally it will look like, okay, this is an A.

Lex Fridman 1:00:03 Mm-hmm.

Dileep George 1:00:04 Uh, b- and, and locally, just when I look at just that patch of the image, it looks like an A.

Lex Fridman 1:00:10 Mm-hmm.

Dileep George 1:00:10 But when I look at it in the context of all the other causes, it might not, you know, A is not something that makes sense. So, that is something you have to kind of rec- you know, recursively figure out.

Lex Fridman 1:00:20 Yeah.

Dileep George 1:00:20 Um...

Lex Fridman 1:00:21 So, okay, so, uh, and, uh, this thing performed pretty well on the CAPTCHAs?

Dileep George 1:00:26 Correct.

Lex Fridman 1:00:27 And, um, I mean, is there some kind of interesting intuition you can provide why it did well? Like, what did it look like? Is there visualizations that could be human interpretable to us humans?

Dileep George 1:00:39 Yes. Yeah. So, the, the good thing about the model is that it is extremely, um, so it is not just doing a classification, right? It is, it is, it is, it is providing a full explanation for the scene. So when, when it, when it, um, operates on a scene, it is coming it back and saying, "Look, this is the part is the A, and these are the pixels that turned on, uh, these are the pixels in the input that tells, makes me think that it is an A."

Lex Fridman 1:01:08 Mm-hmm.

Dileep George 1:01:08 "And also, these are the portions I hallucinated." It, it, you know, it, it provides a complete explanation of that form. And then ............................ These are the contours, these are, this is the interior, and this is in front of this other object. So, th- that, that's the kind of, um, explanation it, uh, the, the inference network provides. So, so that, that is useful and interpretable. Um, and, uh, um, then the kind of errors it makes are also g- I don't want to, um, read too much into it. But the kind of errors the network makes are, uh, very similar to the kinds of errors humans would make in a, in a similar situation.

Lex Fridman 1:01:50 So, there's something about the structure that's, uh, feels reminiscent of the way humans, uh, visual...

Dileep George 1:01:55 Yeah.

Lex Fridman 1:01:57 ... system works. Well, I mean, uh, how hard coded is this to the CAPTCHA problem, this idea?

Dileep George 1:02:04 Uh, not really hard coded, because it's the, uh, the assumptions, as I mentioned, are general, right? It is more, um, and, and those themselves can be applied in many situations which are-... natural signals. Um, so it's, it's the foreground versus, uh, background factorization and, uh, the factorization of the surfaces versus the contours. So, these are all generally applicable assumptions.

Lex Fridman 1:02:27 In- in all vision.

Dileep George 1:02:29 Yeah.

Lex Fridman 1:02:29 So, why, why capture- why attack the capture problem, which is quite unique in the computer vision context, versus, like, the traditional benchmarks of ImageNet and all those kinds of image classification, or even segmentation tests and all that kind of stuff? Do you feel like that's, uh... I mean, what- what's your thinking about those kinds of benchmarks in, um, in this- in this context?

Dileep George 1:02:53 I mean, those benchmarks are useful for deep learning kind of algorithms where you- you know- so the- the settings, uh, that deep learning works in are... here is my huge training set, and here is my test set. So the- the- the training set is almost, uh, you know, 100x, 1,000x bigger than, uh, the test set, in many- m- many cases. Uh, what we wanted to do was invert that. E- either the training set is very smaller than the- the test set.

Lex Fridman 1:03:21 Yes.

Dileep George 1:03:22 Uh, and, uh- uh, and, you know, uh, capture is a problem that is, by definition, hard for computers, and it has these good properties of strong generalizations, strong out-of-training distribution generalization. If you are interested in studying that, uh, and putting- uh, having your model have that property, then it's a- it's a good data set to tackle.

Lex Fridman 1:03:46 So is there- have you attempted to... which I think- I believe there's quite a growing body of work on looking at MNIST and ImageNet without training. So like taking- like the basic challenge is how- what tiny fraction of the training set can we take in order to do a reasonable job of the-

Dileep George 1:04:09 Right.

Lex Fridman 1:04:09 ... classification task? Have- have you explored that angle on these classic benchmarks?

Dileep George 1:04:15 Yes. So- so we did do MNIST. So, uh, you know, so it's not just capture.

Lex Fridman 1:04:18 Mm-hmm.

Dileep George 1:04:19 We, uh... So there was, uh, also, uh, uh, versions of- multiple versions of MNIST, including the- the standard version which- where we inverted the problem, which is basically saying rather than train on 60,000, uh, training data, uh, you know, how, uh, quickly can you get, uh, to high level accuracy with very little training data?

Lex Fridman 1:04:39 Was- is there some, uh, performance that you remember? Like how well- how well did it do? How many examples did it need?

Dileep George 1:04:47 Yeah. I- I- like, you know, I remember that it was, you know, uh, on the order of, uh, tens or hundreds of examples to get into, uh, 95% accuracy. And it was- it was definitely better than the systems- other systems out there at that time.

Lex Fridman 1:05:04 At that time?

Dileep George 1:05:04 Yeah.

Lex Fridman 1:05:04 Yeah. They're really pushing that. I think that's a really interesting space actually. Uh, I- I think there's an actual name for MNIST that, uh- like there's different names for the different sizes of training sets. I mean, people are like attacking this problem.

Dileep George 1:05:21 Yeah.

Lex Fridman 1:05:21 I think it's super interesting.

Dileep George 1:05:22 Yeah.

Lex Fridman 1:05:22 It's funny how like the MNIST will probably be with us all the way to AGI.

Dileep George 1:05:28 Yes. (laughs)

Lex Fridman 1:05:29 (laughs) It's a data set that just sticks by. It- it is, it's a clean, simple, uh, data set to, uh, to study the fundamentals of learning with. Just like captures. It's interesting-

Dileep George 1:05:40 Got it.

Lex Fridman 1:05:40 ... not enough people... I don't know, maybe you can correct me, but I feel like captures don't show up as often in papers as they probably should.

Dileep George 1:05:48 That's correct. Yeah. Because you know, um, usually these things have a momentum, uh, you know, once- once, uh, something gets established as a standard benchmark.

Lex Fridman 1:05:59 Yeah.

Dileep George 1:05:59 There is a- there is a- uh, there is a dynamics of, uh, how graduate students operate and how, uh, academ- academic system works that, uh, pushes people to track that, uh, benchmark. So-

Lex Fridman 1:06:11 Yeah. To fo- (laughs)

Dileep George 1:06:12 Yeah. So that-

Lex Fridman 1:06:12 Nobody wants to think outside the box. Okay.

Dileep George 1:06:15 Yeah. (laughs)

Lex Fridman 1:06:16 Uh, (laughs) okay. So good performance on the captures. What else is there interesting, um, on the RCN side before we talk about the cortical microstr-

Dileep George 1:06:25 Yeah. So the same model- so the- the- the important part of the model was that it trains very quickly with very little training data, and it's, uh, you know, quite robust to out-of-distribution, uh, perturbations. Um, and- and we are using that, uh, very, uh, fruitfully and uh, vicariously in many of the robotics tasks we are solving.

Dileep George 1:06:48 Uh-

Lex Fridman 1:06:48 Solving now. Well, let me ask you this kind of touchy question. I have to- I- I've spoken with, uh, your friend, colleague, Jeff Hawkins too. I mean, he's, uh... I ha- I have to kind of ask, there is a bit of... whenever you have brain-inspired stuff-

Dileep George 1:07:04 Yeah.

Lex Fridman 1:07:04 ... and you make big claims-

Dileep George 1:07:06 Yeah.

Lex Fridman 1:07:06 ... uh, big sexy claims-

Dileep George 1:07:08 Yeah.

Lex Fridman 1:07:08 ... there's a, you know, uh, there's critics. I mean, machine learning subreddit.

Dileep George 1:07:14 (laughs)

Lex Fridman 1:07:14 Don't get me started on those people. Uh, they're- they're har- I mean, criticism is good, but they're a bit- uh, they're a bit over the top. Um, there is quite a bit of sort of skepticism and criticism, you know, does this work really as good as it promises to be?

Dileep George 1:07:30 Yeah.

Lex Fridman 1:07:30 What... Do you have thoughts on that kind of skepticism? Do you have comments on the kind of criticism you might have received, uh, about, you know, is this approach legit? Is this- is this a promising approach?

Dileep George 1:07:44 Yeah.

Lex Fridman 1:07:44 Or at least as promising as it seems to be s- you know, advertised as?

Dileep George 1:07:49 Yeah. I- I can comment on it. Um, so you know, our- our RCN paper is published in Science, which I would argue is- is a very high quality journal, very hard to, uh, publish in. And us- you know, usually it is indicative of the- of the quality of the work. And, um, uh, I can- I can- I- I am very, very certain that the ideas that we brought together in that paper, uh, in terms of the importance of feedback connections, uh, recursive inference, lateral connections, uh, coming to best explanation of the scene as the problem to solve, trying to solve, um, recognition segmentation, uh, all jointly.... in a way that is compatible with higher level cognition, top-down attention, all those ideas that we brought together into something, you know, coherent and workable in the, uh, in the world, and solving a challenging, tackling a challenging problem. I think that will, that will stay and that, that contribution, I stand by, right?

Lex Fridman 1:08:44 Yeah.

Dileep George 1:08:44 Now, uh, I can, I can sh- uh, tell you a story, uh, which is funny in the, in the context of this, right? Um, so if you read the abstract of the paper and, like, you know, the argument we are putting in, we- you know, we are putting in, "Look, current deep learning systems take a lot of training data. Uh, they don't use these insights. And here is our new model, which is not a deep neural network, it's a graphical model. It does inference." This is wha- how the paper is, right? Now, once the paper was accepted and everything, um, it went to the press department in, in Science, you know, to play as science office. We, we didn't do any press release when it was published.

Lex Fridman 1:09:17 Yeah.

Dileep George 1:09:17 It was... it went to the press department. What did the, what was the press release that they wrote up? "A new deep learning model (laughs) so-

Lex Fridman 1:09:25 Solves CAPTCHAs.

Dileep George 1:09:26 Solves CAPTCHAs. And, uh, so, so you can see where was, you know, what, what was being hyped, uh, in that, uh, thing, right? So, so it's like, um, there is the, there is a dynamic in the, uh, in the community of, you know, so, uh, um, that's especially happens when there are lots of new people coming into the field and they get attracted to one thing, and some people are trying to think different, uh, compared to that. So there is, there is some, uh... I, I think skepticism in science is important and it is, eh, eh, um, you know, very much, uh, required, but it's also, it's not, uh, skepticism usually, it's mostly bandwagon effect that is happening, rather than inform-

Lex Fridman 1:10:05 Well, well, but that's not even that. I mean, I'll tell you what they react to, which is like, uh, I'm sensitive too as well. If you look, if you look at just companies, OpenAI, DeepMind...

Dileep George 1:10:15 Yeah.

Lex Fridman 1:10:15 Um, Vicarious. I mean, they just, uh, there's a little bit of a race to the top and hype, right?

Dileep George 1:10:25 Right.

Lex Fridman 1:10:25 It's, it's like it doesn't pay off to be humble.

Dileep George 1:10:29 (laughs) Right. (laughs)

Lex Fridman 1:10:31 We... so like, uh... and, and the press is just, uh, irresponsible often. They, they just... I mean, don't get me started on the state of journalism today. Like, it seems like the people who write articles about these things, they literally have not even spent an hour on the Wikipedia article about what is neural networks. Like-

Dileep George 1:10:51 Yeah.

Lex Fridman 1:10:51 ... they haven't, like, invested just even the language, the laziness. It's like, uh, robots beat humans. Like, they, they write this kind of stuff-

Dileep George 1:11:04 Yeah.

Lex Fridman 1:11:05 ... that just, uh... and then, uh, and then, of course, the researchers are quite sensitive to that, uh, because it gets a lot of attention. They're like, "Why did this word get so much attention?" Uh, you know, that's, that's over the top. And people get really sensitive, you know, the same kind of criticism with, um... OpenAI did work with the Rubik's Cube, with the robot.

Dileep George 1:11:24 Yeah.

Lex Fridman 1:11:24 The people criticized. Uh, same with GPT-2 and 3, they criticize. Uh, same thing with, uh, DeepMinds with AlphaZero. I mean, uh, y- yeah. I, I'm sensitive to it. Um, but... and of course with your work, you mentioned deep learning, but there's something super sexy to the public about brain-inspired.

Dileep George 1:11:45 Correct.

Lex Fridman 1:11:45 I mean, that immediately grabs people's imagination. Not even like neural networks, but like really brain-inspired. Like-

Dileep George 1:11:53 Got it.

Lex Fridman 1:11:54 ... neur- like, brain-like neural networks.

Dileep George 1:11:57 Right, right.

Lex Fridman 1:11:57 That seems really compelling to people, and, um, to me as well, to, to, to the world as, as a narrative. And so, uh, people hook up, hook on to that, and, uh, sometimes you, uh, the skepticism engine turns on in the research community, and they're-

Dileep George 1:12:14 Correct.

Lex Fridman 1:12:14 ... skeptical. But I think putting aside the ideas of the actual performance on CAPTCHAs or performance on any dataset, I mean, to me, all these datasets are useless anyway. It's nice to have them. Uh, but in the grand scheme of things, they're silly toy examples. The point is, is there intuition about the f- the ideas, just like you mentioned, bringing the ideas together in a unique way? Is there something there? Is there some value there? And is it gonna stand the test of time?

Dileep George 1:12:45 Yes.

Lex Fridman 1:12:46 And that's the hope.

Dileep George 1:12:47 Yes.

Lex Fridman 1:12:47 That's the hope here.

Dileep George 1:12:47 Uh, I, I am... my confidence there is very high. I, you know, I don't treat brain-inspired as a marketing term.

Lex Fridman 1:12:55 Mm-hmm.

Dileep George 1:12:55 Uh, you know, I am loo- looking into the details of biology and, and puzzling over, uh, those things. And I am, I am grappling with those things. And so it is, it is not a marketing term at all. It... you know, you can use it as a marketing term and, and people often use it, and you can get combined with them. And when, when people don't understand how we are approaching the problem, it is, it is easy to be, uh, misunderstood and, and, uh, think of it as, you know, purely, uh, marketing. But that's not the way, uh, we are-

Lex Fridman 1:13:26 So, so you really... I mean, as a scientist-

Dileep George 1:13:29 Yeah.

Lex Fridman 1:13:30 ... you believe that if we kinda just stick to really understanding the brain, that's going to, that's the right... like, you, you should constantly meditate on the, "How does the brain do this?" Because that's going to be really helpful for engineering intelligence systems.

Dileep George 1:13:46 Yes. You need to... so I think it is, it's one input and it is, it is helpful, but you, you should know when to deviate from it, too. Um, so an example is convolutional neural networks, right?

Lex Fridman 1:13:59 Yeah.

Dileep George 1:13:59 Uh, convolution is not an operation brain in, uh, implements. Eh, eh, the visual cortex is not convolutional. Visual cortex has local receptive fields, local connectivity.

Lex Fridman 1:14:11 Mm-hmm.

Dileep George 1:14:11 But they... you know, the, um... there i- there is no translation in invariants in the, um, uh, the network weights, um, in, in the visual cortex. That is a, uh, computational trick, which is a very good engineering trick that we use for sharing the training between the different, uh, nodes. Um, so-... a- and, and that trick will be with us for some time. It will go away when we have, um, uh, uh, robots with eyes and heads that move. Uh, and, so then the- that trick will go away. It will not be, uh, useful at that time. So-

Lex Fridman 1:14:48 Oh, so the brain doesn't... So the brain doesn't have translational invariants, it has the focal point, like it has the thing it focuses on.

Dileep George 1:14:56 Correct. It does- it does have fovea and-

Lex Fridman 1:14:58 Fovea.

Dileep George 1:14:58 ... and because of the fovea, um, the, the receptive fields are not, like, uh, the copying of the waves.

Lex Fridman 1:15:04 Yeah.

Dileep George 1:15:04 Like, the, the waves in the center are very different from the waves in the periphery.

Lex Fridman 1:15:07 Yes. At the peripher- I mean, I, I did this, uh, actually wrote a paper and just gotten the chance to really study peripheral- peripheral vision, which is a fascinating thing.

Dileep George 1:15:19 Right.

Lex Fridman 1:15:19 Very under-understood thing of what the br- uh, you know, at the every level, the brain does with the periphery. It does some funky stuff.

Dileep George 1:15:29 Yeah.

Lex Fridman 1:15:29 So it's a- it's another kind of trick than, uh, convolutional. Like, it does, it, it, uh, it's, uh, you know, convulsion your- convolution in your own networks is a trick to- for efficiency, is efficiency trick.

Dileep George 1:15:45 Yes.

Lex Fridman 1:15:45 And the brain does a whole another kind of thing, I guess.

Dileep George 1:15:47 Correct. Correct. So, so you need to understand the principles of processing so that you can still apply engineering tricks-

Lex Fridman 1:15:55 Yeah.

Dileep George 1:15:55 ... when- where you want it to. You don't want to be slavishly mimicking all the things of the brain. Um, and, and so, yeah, so it should be one input, and I think it is extremely helpful. Uh, but you- it should be the point of really understanding so that you know when to deviate from it.

Lex Fridman 1:16:10 So, okay. That's really cool, that that's work from, uh, a few years ago. So, you'd, uh, you did work in Numenta with Jeff Hawkins-

Dileep George 1:16:19 Yeah.

Lex Fridman 1:16:20 ... uh, with, um, hierarchical temporal memory. How has your just- if you could just b- give a brief history, how has your view of the way the models of the brain changed over the past few years leading up to, to now? Is there some interesting aspects where there was an adjustment to your understanding of the brain? Or is it all just building on top of each other?

Dileep George 1:16:44 In terms of the higher level ideas, uh, especially the ones Jeff wrote about in the book, if you, if you blur out, right, you know, if you-

Lex Fridman 1:16:51 Uh, yeah. On intelligence. That's-

Dileep George 1:16:52 Right, on intelligence. If you, if you blur out the details, and, and if you just zoom out under the higher level idea, uh, things are, I would say, consistent with what he wrote about. But, but many things will be consistent with that because it is- it's a blur, you know, when you, when you, uh, you know. Deep learning systems w- are also h- you know-

Lex Fridman 1:17:09 Right.

Dileep George 1:17:09 ... multi-level, hierarchical, all of those things. Right? So, so at the... But, um, in terms of the detail, a lot of things are different. Uh, and, and those details matter a lot. Um, so, so one point of difference I had with Jeff, uh, uh, was, uh, how to approach, you know, how much of biological plausibility and realism do you want in the learning algorithms?

Lex Fridman 1:17:36 Mm.

Dileep George 1:17:36 Um, so, uh, when I was there, uh, this was, you know, almost 10 years ago now, so, uh, I, I- (laughs)

Lex Fridman 1:17:43 Time flies when you're having fun. (laughs)

Dileep George 1:17:44 (laughs) I don't know- I don't know what Jeff thinks now, but 10 years ago, uh, the difference was that I did not want to be so constrained on saying, "Uh, my learning algorithms want- need to be biologically plausible," uh, based on some filter of biological plausibility available at that time. To me, that is a dangerous cut to make, because we are, you know, discovering more and more things about the brain all the time. New biophysical mechanisms, new channels, uh, are being discovered all the time. So I don't want to upfront kill off an, uh, a learning algorithm, just-

Lex Fridman 1:18:20 Because we

Lex Fridman 1:18:21 don't really understand the full d- uh, the full, uh, v- biophysics or whatever of how the brain learns.

Dileep George 1:18:27 E- exactly. Exactly. So-

Lex Fridman 1:18:28 But l- let me ask, and I'm sorry to interrupt. Like, eh, d- what's up- what's your sense, what's our best understanding of how the brain learns?

Dileep George 1:18:36 So things like backpropagation, credit assignment. So, so many of these algorithms-

Lex Fridman 1:18:42 Yeah.

Dileep George 1:18:42 ... have- learning algorithms have things in common, right? It is- uh, yeah, backpropagation is one. We have credit assignment. There is another algorithm called expectation maximization, which is, you know, a- another b- weight adjustment algorithm.

Lex Fridman 1:18:56 But is th- your sense the brain does something like this?

Dileep George 1:18:59 Has to. There is no way around it, in the sense of saying that you do have to adjust the- the connections.

Lex Fridman 1:19:06 You- so, yeah, and you're saying credit assignment, you have to reward the connections that were useful in making a correct prediction, and not... Yeah, I guess, what el- but, yeah, it doesn't have to be (laughs) differentiable. I mean... (laughs)

Dileep George 1:19:17 Yeah, it doesn't have to be differentiable.

Lex Fridman 1:19:19 Yeah.

Dileep George 1:19:19 Yeah. But you have to have a, uh, you know, you have a model that you start with, you ve- you have data comes in, and you have to have a way of adjusting the model such that it better fits the data.

Lex Fridman 1:19:31 Yeah.

Dileep George 1:19:32 So that, that is all of learning. Right? And-

Lex Fridman 1:19:34 Yep.

Dileep George 1:19:34 ... and some of them can be using backprop to do that. Some of it can be using, uh, you know, very local, uh, g- graph changes to do that. Uh, there can b- you know, uh, many of these learning algorithms have similar update properties locally, i- in terms of what the neurons need to do locally.

Lex Fridman 1:19:57 I wonder if small differences in learning algorithms can have huge differences in the actual effect. So the dynamics of, I mean, uh, sort of the rev- the reverse like spiking, like, the, uh, i- if credit assignment is like, uh, a lightning versus like, uh, a rainstorm or something. Like, uh, whether, whether there's a f- like a looping local type of situation with the credit assignment.

Dileep George 1:20:24 Yeah.

Lex Fridman 1:20:25 Uh, whether there is, uh, like regularization, like how, how, um, how it injects robustness into the whole thing.... like, whether it's chemical or electrical or mechanical-

Dileep George 1:20:40 Yeah.

Lex Fridman 1:20:41 ... uh, all those kinds of things.

Dileep George 1:20:42 Yes. (laughs)

Lex Fridman 1:20:43 Like that, I, I feel like it, it that, yeah, I feel like those differences could be essential, right?

Dileep George 1:20:49 It could be. It's just that you don't know enough to, on the learning side, you don't know, uh, enough to say, "That is definitely not the way the brain does it."

Lex Fridman 1:21:01 Right. Got it.

Dileep George 1:21:01 Um-

Lex Fridman 1:21:01 So you don't wanna be stuck to it.

Dileep George 1:21:03 Right.

Lex Fridman 1:21:03 So that, yeah, so you, you've been open-minded on that side of things.

Dileep George 1:21:07 Correct. On the inference side, on the recognition side, I am much more, uh, amenable to being constrained, because I- it's much easier to do experiments because, you know, it's like, okay, here is a stimulus, you know, wha- how many steps did it get to take the answer?

Lex Fridman 1:21:19 Right.

Dileep George 1:21:19 I can trace it back. I can, I can understand the speed of that computation-

Lex Fridman 1:21:24 Got it.

Dileep George 1:21:24 ... et cetera, much more, uh, readily on the inference side.

Lex Fridman 1:21:28 Got it. And then you can't do good experiments on the learning side.

Dileep George 1:21:30 Correct.

Lex Fridman 1:21:33 So let, let's, let's go right into cortical microcircuits right back. So what, uh, what are these ideas beyond recursive cortical network that, uh, you're looking at now?

Dileep George 1:21:45 So we have made a, uh, you know, pass through all, you know, multiple of the steps that we, uh, you know, as I, as I mentioned earlier, you know, we were looking at perception from the angle of cognition, right? It was not just perception for perception's sake. How do you, how do you connect it to cognition? Uh, how do you learn concepts and, uh, how do you learn abstract reasoning? Uh, similar to some of the things Francois, uh, uh, talked about, right? Um, so, um, so we have, uh, taken one pass through it basically saying, "What is the basic cognitive architecture that you need to have, which has a perceptual system, which has a system that learns dynamics of the world, and then has something like a routine program learning system on top of it, to learn concepts?" So we have, we have built one, the, you know, the version 0.1 of that system. Uh, this was another, uh, science robotics paper. Uh, it is, it's the title of that paper was, you know, something like Cognitive Programs. How do you build cognitive programs? Uh, I'm-

Lex Fridman 1:22:48 And the application there was on m- uh, manipulation, robotic manipulation?

Dileep George 1:22:53 It was, it was, um, so think of it like this. Suppose you, uh, wanted to tell, uh, a new person, uh, that you met, you don't know the language, uh, uh, that person uses. You want to communicate to that person, uh, to achieve some task.

Lex Fridman 1:23:08 Mm-hmm.

Dileep George 1:23:08 Right? So I want to say, "Hey, um, you need to pick up all the, the red cups from the kitchen counter and put it here."

Lex Fridman 1:23:16 Mm-hmm.

Dileep George 1:23:16 Right? Uh, how do you communicate that, right? You can show pictures. You can basically say, "Look, this is the starting state. Uh, the, the, the things are here, this is the ending state." And, and what does the person need to understand from that? The, the person need to understand what conceptually happened in those pictures from the input to the output, right?

Lex Fridman 1:23:35 Yeah. Right.

Dileep George 1:23:35 Um, so, um, so we are looking at pre-verbal conceptual understanding. Without language, how do you, how do you have a set of concepts that you can manipulate in your head? Uh, and from a s- you know, set of images of input and output, can you infer what is happening in those images?

Lex Fridman 1:23:57 Got it. With concepts that are pre-language. Okay.

Dileep George 1:23:59 Yeah.

Lex Fridman 1:23:59 So wha- what's it mean to, for a concept to be pre-language? Like-

Dileep George 1:24:03 Yeah.

Lex Fridman 1:24:04 ... why, why is so, why (laughs) why is language, uh, so important here?

Dileep George 1:24:10 So I, I want to make a distinction between concepts that are, are just learned from text-

Lex Fridman 1:24:17 Ah.

Dileep George 1:24:17 ... by, by just, just feeding brute force text. Uh, you can, you can start extracting things. Like, okay, a cow is likely to be on grass, uh, in a, in a-

Lex Fridman 1:24:28 Got it.

Dileep George 1:24:28 Uh, so those kinds of things you can extract purely from text. Um, uh, but that's kind of a simple association, uh, thing, rather than a concept as an abstraction of something that happens in the real world, uh, in a, in a, in a grounded way that I can, I can simulate it in my mind and connect