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

The following is a conversation with Dalip 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, Dalips always sought to engineer intelligence the 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 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, Podcast support on Patreon, or connect with me on Twitter lexfriedman. 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 by Alexander Block that you'll start to understand if you sign up to Babel. Now. I say that you'll only start to understand this poem because Russian starts with the language and ends with the vodka. Now the latter part is definitely not endorsed or provided by Babel and will probably lose me the sponsorship. But once you graduate from Babel, 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@buyraycon.com flex 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 ultramarathon and listening to audiobooks on World War II. The sound is rich and really comes in clear. So again, get them@buyraycon.com Lex this show is sponsored by MasterClass. Sign up on 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, I really want to play a City Builder game. Carlos Santana on guitar, Garret Kasparov on chess, Daniel Negrano on poker and many more. Chris Hatfield 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@masterclass.com to get a discount and to support this podcast. And now, here's my conversation with Dalip George do you think we need to understand the brain in order to build it?

Yes. If you want to build the brain, we definitely need to understand how it works. Blue Brain, or Henry Markham's project, is trying to build a brain without understanding it. Just trying to put details of the brain from neuroscience experiments into a giant simulation by putting more and more neurons, more and more details. But that is not going to work. Because when it doesn't perform as what you expect it to do, then what do you do? You 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.

Build it at the functional level, understand? So can you actually linger on and describe the Blue Brain project? It's kind of a fascinating principle, an idea to try to simulate the brain we're talking about the human brain, Right?

Right. Human brains and rat brains or cat brains have lots in common that the cortex, the neocortex structure, is very similar. So initially, they were trying to just simulate a cat brain and to understand.

The nature of evil.

To understand the nature of evil, or as it happens in most of these simulations, you easily get one thing out, which is oscillations. If you simulate a large number of neurons, they oscillate. And you can adjust the parameters and say that, oh, oscillations match the rhythm that we see in the brain, et cetera.

I see. So the idea is. Is the simulation at the level of individual neurons?

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

But in your sense, these models of the way signal travels with the axons and all the basic models, that's. They're too crude.

Oh, well, actually, they are pretty detailed and pretty sophisticated. And they do replicate the neural dynamics. If you take a single neuron and you try to turn on the different channels, the calcium channels and the different receptors, and see what the effect of turning on or off those channels are in the neurons. Spike output. People have built pretty sophisticated models of that. And they are, I would say, in the regime of correct.

Well, see, the correctness. That's interesting because you mentioned it at several levels, the correctness is measured by looking at some kind of aggregate statistics.

It would be more the spiking dynamics.

Of a singular spiking dynamics of signaling.

And yeah, these models, because they are going to the level of mechanism. Right. So they are basically looking at, okay, what is the effect of turning on an ion channel? And you can model that using electric circuits. So it is not just a function fitting. People are looking at the mechanism underlying it. And putting that in terms of electric circuit theory, signal propagation theory, and modeling that so those models are sophisticated, but getting a single neurons model. 99% right. Does not still tell you how to. It would be the analog of getting a transistor model. Right. And now trying to build a microprocessor. And if you just observe, if you did not understand how a microprocessor works, but you say, oh, I now can model one transistor. Well, and now I will just try to interconnect the transistors according to whatever I could guess from the experiments and try to simulate it, then it is very unlikely that you will produce a functioning microprocessor. When you want to produce a functioning microprocessor, you want to understand Boolean logic. How does. How do the gates work all those things and then understand how do those gates get implemented using transistors.

Yeah, there's actually, I remember this reminds me, there's a paper, maybe you're familiar with it. I remember going through in a reading group that approaches microprocessor from a perspective of a neuroscientist. I think it basically it uses all the tools that we have of neuroscience to try to understand, like as if we just aliens showed up to study computers and to see if those tools can be used to get any kind of sense of how the microprocessor works. I think the final takeaway from at least this initial exploration is that we're screwed. There's no way that the tools of neuroscience would be able to get us to anything, not even Boolean logic. I mean, it's just any aspect of the architecture of the function of the processes involved, the clocks, the timing, all that. You can't figure that out from the tools of neuroscience.

Yeah, so I'm very familiar with this particular paper. I think it was called can a Neuroscientist understand a Microprocessor or something like that. Following the methodology in that paper, even electrical engineer would not understand microprocessors. So I don't think it is that bad in the sense of saying neuroscientists do find valuable things by observing the brain. They do find good insights, but those insight cannot be put together just as a simulation. You have to investigate what are the computational underpinnings of those findings. How do all of them fit together from an information processing perspective? Somebody has to painstakingly put those things together and build hypothesis. So I don't want to diss all of neuroscience as saying, oh, they are not finding anything. No, that paper almost went to that level of neuroscientists will never understand. No, that's not true. I think they do find lots of useful things, but it has to be put together in a computational framework.

Yeah, I mean, but just the AI systems will be listening to this podcast 100 years from now, and they will probably. There's some non zero probability they'll find your words laughable. It's like, I remember humans thought they understood something about the brain. They're totally clueless. There's a sense about neuroscience that we may be in the very, very early days of understanding the brain, but I mean, that's one perspective. In your perspective, how far are we into understanding any aspect of the brain? So the dynamics of the individual neuron communication to how when they, in a collective sense, how they're able to store information, transfer information, how intelligence then emerges, all that kind of stuff. Where are we on that timeline?

Yeah, so, you know, timelines are very, very hard to predict. And we can of course, be wrong, and it can be wrong on either side. You know, we know that when we look back, the first flight was in 1903. In 1900, there was a New York Times article on flying machines that do not fly and humans might not fly for another hundred years. That was what that article stated. But no, they flew three years after that. So it's very hard to.

And on that point, one of the Wright brothers, I think two years before said that like, he said, like some number, like 50 years, he has become convinced that.

It'S impossible, even during their experimentation.

Yeah, I mean, that's attributed to when that's like the entrepreneurial battle of depression, of going through just like thinking this is impossible. But yeah, there's something. Even the person that's in it is not able to see estimate correctly.

Exactly. But I can tell from the point of 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. So my way of understanding the brain would be to basically say, look at the insights neuroscientists have found. Understand that from a computational angle, information processing angle, build models using that, and then building that model which functions, which is a functional model, 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 trying to do what the brain is trying to do on the whole functional level. And building that model will help you fill in the missing pieces that biology just gives you the hints. And building the model fills in the rest of the pieces of the puzzle. And then you can go and connect that back to biology and say, okay, now in it makes sense that this part of the brain is doing this, or this layer in the cortical circuit is doing this, and then continue this iteratively because now that will inform new experiments in neuroscience. And of course, building the model and verifying that in the real world will also tell you more about does the model actually work? And you can refine the model, find better ways of putting these neuroscience insights together. So I would say neuroscientists alone, just from experimentation, will not be able to build a model of the brain or a functional model of the brain. There's lots of efforts which are very impressive efforts in collecting more and more connectivity data from the brain. How are the model microcircuits of the brain connected with each other?

Those are beautiful, by the way.

Those are beautiful. And at the same time, those do not itself by themselves convey the story of how does it work? And somebody has to understand, okay, why are they connected like that and what are those things doing? And we do that by building models in AI, using hints from neuroscience and, and repeat the cycle.

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. 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, functionally speaking, like you said, you find interesting?

Yeah, quite a lot of things. All right, so one is, if you look at the visual cortex, visual cortex is a large part of the brain. I forget the exact fraction, but a huge part of our brain area is occupied by just vision. So visual cortex is not just a feed forward cascade of neurons. There are a lot more feedback connections in the brain compared to the feedforward connections. And it is surprising to the level of detail, neuroscientists have actually studied this. If you go into neuroscience literature and poke around and ask have they studied what will be the effect of poking a neuron in level it in Level V1? And have they studied that? And you'll say, yes, they have studied that.

So every possible combination, I mean, it's.

Not a random exploration at all. It's very hypothesis driven. Right. They are very experimental. Neuroscientists are very, very systematic in how they probe the brain because experiments are very costly to conduct. They take a lot of preparation, they need a lot of control. So they, they are very hypothesis driven in how they probe the brain. And often what I find is that when we have a question in AI about has anybody 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 they have hypothesis about how those lateral connections are supposedly contributing to do visual processing. But of course they haven't built very functional, detailed models of it.

By the way, how do studies. Sorry to interrupt. Do they stimulate a neuron in one particular area of the visual cortex and then see how the signal travels kind of thing.

Very, very fascinating experiments. So I can give you one example I was impressed with. So before going to that, let me give you overview of how the layers in the cortex are organized. Right. Visual cortex is organized into roughly four hierarchical levels. Okay, so V1, V2, V4, it. And in V1, what happened to V3? Well, yeah, there's another pathway. Okay, so I'm talking about just object recognition pathway.

All right, cool.

And then in V1 itself. So it's. There is a very detailed microcircuit. In V1 itself, there is organization within a level itself. The cortical sheet is organized into multiple layers. And there are columnar structure. And this layer wise and columnar structure is repeated in V1, V2, V4IT, all of them. And the connections between these layers within a level. In V1 itself, there are six layers, roughly, and the connections between them there is a particular structure to them. So one example of an experiment people did is when you present a stimulus which, let's say requires separating the foreground from the background of an object. So it's a textured triangle on a textured background. And you can check does the surface settle first or does the contour settle first? Settle, settle in the sense that when you finally form the percept of the triangle, you understand where the contours of the triangle are and you also know where the inside of the triangle is. That's when you form the final perception. Now you can ask what is the dynamics of forming that final percept? Do the neurons first find the edges and converge on where the edges are and then they find the inner surfaces, or does it go the other way around?

The other way around. So what's the answer?

In this case, it turns out that it first starts. Settles on the edges. It converges on the edge hypothesis first, and then the surfaces are filled in from the edges to the inside.

That's fascinating.

And the detail to which you can study this. It's amazing that you can actually not only find the temporal dynamics of when this happens, and then you can also find which layer in v1 which layer is encoding the edges, which layer is encoding the surfaces, and which layer is encoding the feedback, which layer is encoding the feed forward and what's the combination of them that produces the final percent. And these kinds of experiments stand out when you try to explain illusions. One example of a favorite illusion of mine is the Kanitsa triangle. I don't know whether you are familiar with this one. So this is, this is an example where it's a triangle, but only the corners of the triangle are shown in the stimulus. So they look like kind of Pac man.

Oh, the black Pac man.

Exactly. And then you start to see your visual system hallucinates the edges. And when you look at it, you will see a faint edge. And you can go inside the brain and look. Do actually neurons signal the presence of this edge? If they signal, how do they do it? Because they are not receiving anything from the input. The input is blank for those neurons. How do they signal it? When does the signaling happen? If a real contour is present in the input, then the neurons immediately signal, okay, there is an edge here. When it is an illusory edge, it is clearly not in the input, it is coming from the context. So those neurons fire later. And you can say that, okay, it's the feedback connections that is causing them to fire and they happen later. And you can find the dynamics of them. So these studies are pretty impressive and very detailed.

So by the way, just step back. You said that there may be more feedback connections than feed forward connections.

Yeah.

First of all, if it's just for like a machine learning folks.

Yeah.

I mean that, that's crazy that there's all these feedback connections. I mean, we often think about. Thank. Thanks to deep learning, you start to think about the human brain as a kind of feed forward mechanism.

Right.

So what the heck are these feedback connections? Yeah, what's the dynamics? What are we supposed to think about them?

Yeah. So this fits into a very beautiful picture about how the brain works. Right. So the beautiful picture of how the brain works is that our brain is building a model of the world. So our visual system is building a model of how objects behave in the world. And we are constantly projecting that model back onto the world. So what we are seeing is not just a feedforward thing that just gets interpreted in a feed forward part. We are constantly projecting our expectations onto the world. And what the final percept is a combination of what we project onto the world combined with what the actual sensory input is.

Almost like trying to calculate the difference and then trying to interpret the difference.

Yeah, I wouldn't put it as calculating the difference. It's more like what is the best explanation for the input stimulus based on the model of the world? I have. Got it.

Got it. And that's where all the illusions come in. But that's an incredibly efficient process. So the feedback mechanism, it just helps you constantly. Yeah. So hallucinate how the world should be based on your world model. And then just looking at if there's novelty, like trying to explain it. 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 cortex. You're saying this happens at the lowest level? At the highest level.

Yes. Yeah. In fact, feedback connections are more prevalent in everywhere in the cortex. And so one way to think about it, and there's a lot of evidence for this, is inference. So, you know, so basically, if you have a model of the world and when some evidence comes in, what you are doing is inference. Right. You are trying to now explain this evidence using your model of the world. And this inference includes projecting your model onto the evidence and taking the evidence back into the model and doing an iterative procedure. And this iterative procedure is what happens using the feedForward feedback propagation. And feedback affects what you see in the world, and it also affects feedforward propagation. And examples are everywhere. We see these kinds of things everywhere. The idea that there can be multiple competing hypothesis 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.

Wait, what? So you have competing models of the world that try to explain. What do you mean by explain away?

So this is a classic example in graphical models, probabilistic models.

What are those?

Okay.

I think it's useful to mention because we'll talk about them more.

Yeah, yeah. So neural networks are one class of machine learning models. You have distributed set of nodes which are called the neurons. Each one is doing a dot product, and you can approximate any function using this multilevel network of neurons. So that's a class of models which are useful for function approximation. There is another class of models in machine learning called probabilistic graphical models. And you can think of them as each node in that model is variable, which is talking about something. It can be a variable representing Is an edge present in the input or not? And at the top of the network a node can be representing is there an object present in the world or not? So it is another way of encoding knowledge. And then once you encode the knowledge, you can do inference in the right way. What is the best way to explain some set of evidence using this model that you encoded? So when you encode the model, you are encoding the relationship between these different variables. How is the edge connected to the model of the object? How is the surface connected to the model of the object? Then of course this is a very distributed, complicated model. And inference is how do you explain a piece of evidence 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 square present in the image? So this is the process of inference. So one example of inference is having this expanding away effect between multiple causes. So graphical models can be used to represent causality in the world. So let's say your alarm at home can be triggered by a burglar getting into your house, or it can be triggered by an earthquake. Both can be causes of the alarm going off. So now you're in your office, you heard burglar alarm going off, you are heading home thinking that there's a burglar. Got it. But while driving home, if you hear on the radio that there was an earthquake in the vicinity, now your strength of evidence for a burglar getting into their house is diminished because now that piece of evidence is explained by the earthquake being present. So if you think about these two causes explaining at lower level variable, which is alarm, now what we are seeing is that increasing the evidence for some cause, there is evidence coming in from below for alarm being present. And initially it was flowing to a burglar being present. But now since there is side evidence for this other cause, it explains away this evidence and evidence will now flow to the other cause. This is two competing causal things trying to explain the same evidence.

And the brain has a similar kind of mechanism for doing so. That's kind of interesting. And that how's that all encoded in the brain? Like where's the storage of information? Are we talking just maybe to get it a little bit more specific? Is it in the hardware of the actual connections? Is it in chemical communication? Is it electrical communication?

Do so this is a paper that we are bringing out soon.

Which one is this?

This is the cortical Microcircuits paper that I sent you a draft of. Of course, this is a lot of. It is still hypothesis one hypothesis that you can think of a cortical column as encoding a concept. A concept. You know, think of it as a. An example of a concept. Is, is an edge present or not? Or is an object present or not? So 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. And then there are connections within the cortical column. Each cortical column is implemented using multiple layers of neurons with very, very, very rich structure. There. There are thousands of neurons in a.

Cortical column, but that structure is similar across the different cortical columns, correct? Yeah, correct.

And also these cortical columns connect to a substructure called thalamus. So all cortical columns pass through this substructure. So our hypothesis is that the connections between the cortical columns implement this. That's where the knowledge is stored about how these different concepts connect to each other. And then the neurons inside this cortical column and in thalamus in combination implement these actual computations in there for inference, which includes explaining away and competing between the different hypotheses. And it is all very. So what is amazing is that neuroscientists have actually done experiments to the tune of showing these things. They might not be putting it in the overall inference framework, but they will show things like if I poke this higher level neuron, it will inhibit through this complicated loop through the thalamus, it will inhibit this other column. So they will do such experiments.

But do they use terminology of concepts? For example, Is it something where it's easy to anthropomorphize and think about concepts? Like you start moving into logic based kind of reasoning systems. So I would just think of concepts in that kind of way, or is it a lot messier, a lot more gray area, you know, even more gray, even more messy than the artificial neural network kinds of abstractions.

Easiest way to think of it as a variable, right? It's a binary variable which is showing the presence or absence of something.

But I guess what I'm asking is, is that something that we're supposed to think of, something that's human, interpretable of that something?

It doesn't need to be it doesn't need to be human interpretable. There's no need for it to be human interpretable. But it's almost like you will be able to find some interpretation of it because it is connected to the other things that you know about.

And the point is, it's useful somehow. Yeah, it's useful as an entity in the graph, in connecting to the other entities that are, let's call them, concepts.

Right.

Okay. So by the way, are these the cortical microcircuits?

Correct. These are the cortical microcircuits. You know, that's what neuroscientists use to talk about the circuits within a level of the cortex. So you can think of, you know, let's think of a neural network, artificial neural network terms. People talk about the architecture of how many layers they build, what is the fan in, fan out, et cetera. That is the macro architecture. And then within a layer of the neural network, the cortical neural network is much more structured within a level. There is a lot more intricate structure there. But even within an artificial neural network, you can think of feature detection plus pooling as one level. And so that is kind of a micro circuit. It's much more complex in the real brain. So within a level, whatever is that circuitry within a column of the cortex and between the layers of the cortex, that's the micro circuitry.

I love that terminology. Machine learning people don't use the same circuit terminology. Right, but they should. It's nice. So, okay, okay. So that's. That's the cortical microcircuit. So what's interesting about, what can we say, what is the paper that you're working on proposed about the ideas around these cortical microcircuits?

So this is a fully functional model for the microcircuits of the visual cortex.

So, so the paper focuses, and your idea in our discussions now is focusing on vision. Yeah, the visual cortex. Okay, so this is a model. This is a full model. This is how vision works.

Well, this is a hypothesis.

Hypothesis.

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

And that again, is a vision system.

That's a vision system.

Computer vision system.

That's a computer vision system, takes in.

Images and outputs what on one side.

It outputs the class of the image and also segments the image. And you can also ask it further queries. Where is the edge of the object? Where is the interior of the object? So 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 segmentation, et cetera. It's a joint model that can do multiple things. 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? So all those things went into the model. The model actually uses feedback connections.

All these ideas from neuroscience.

Yeah.

So what the heck is a recursive cortical network? What are the architecture approaches, interesting aspects here, which is essentially a brain inspired approach to computer vision.

Yeah. So there are multiple layers to this question. I can go from the very, very top and then zoom in. Okay, so one important thing.