Brain Inspired
Brain Inspired
BI 184 Peter Stratton: Synthesize Neural Principles

Support the show to get full episodes and join the Discord community.

Peter Stratton is a research scientist at Queensland University of Technology.

I was pointed toward Pete by a patreon supporter, who sent me a sort of perspective piece Pete wrote that is the main focus of our conversation, although we also talk about some of his work in particular – for example, he works with spiking neural networks, like my last guest, Dan Goodman.

What Pete argues for is what he calls a sideways-in approach. So a bottom-up approach is to build things like we find them in the brain, put them together, and voila, we’ll get cognition. A top-down approach, the current approach in AI, is to train a system to perform a task, give it some algorithms to run, and fiddle with the architecture and lower level details until you pass your favorite benchmark test. Pete is focused more on the principles of computation brains employ that current AI doesn’t. If you’re familiar with David Marr, this is akin to his so-called “algorithmic level”, but it’s between that and the “implementation level”, I’d say. Because Pete is focused on the synthesis of different kinds of brain operations – how they intermingle to perform computations and produce emergent properties. So he thinks more like a systems neuroscientist in that respect. Figuring that out is figuring out how to make better AI, Pete says. So we discuss a handful of those principles, all through the lens of how challenging a task it is to synthesize multiple principles into a coherent functioning whole (as opposed to a collection of parts). Buy, hey, evolution did it, so I’m sure we can, too, right?

0:00 – Intro
3:50 – AI background, neuroscience principles
8:00 – Overall view of modern AI
14:14 – Moravec’s paradox and robotics
20:50 -Understanding movement to understand cognition
30:01 – How close are we to understanding brains/minds?
32:17 – Pete’s goal
34:43 – Principles from neuroscience to build AI
42:39 – Levels of abstraction and implementation
49:57 – Mental disorders and robustness
55:58 – Function vs. implementation
1:04:04 – Spiking networks
1:07:57 – The roadmap
1:19:10 – AGI
1:23:48 – The terms AGI and AI
1:26:12 – Consciousness

[00:00:00] Peter: 

And that’s where I really started sort of solidifying my ideas around what you might call biological neural computation and how different it was to, or how different it is to, you know, standard artificial neural networks and machine learning.

How much of the brain can be understood in terms of these principles. Now, obviously, you can’t understand the principles individually. You do need to understand the principles individually. But if you want to look at how the brain works individually is not going to cut it.

Even your thoughts for the future are really just predictions, right? It’s a principle that should be understood from a brain building block perspective.

[00:00:50] Paul: This is brain inspired. I’m Paul. Peter Stratton is my guest today. Pete is a research scientist at Queensland University of Technology, and I was pointed his way by a Patreon supporter who sent me a sort of perspective piece that Pete wrote, which is the main focus of our conversation, although we also talk about some of his work in particular. For example, he works with spiking neural networks, like my last guest, Dan Goodman. What Pete argues for is what he calls a sideways in approach. So a bottom up approach is to build things like we find them in the brain, put them together, and voila, we have cognition. On the other hand, a top down approach, which is the current main approach in AI, for example, you train a system to perform a task, you give it some algorithms to run, and you fiddle with the architecture and the lower level details until you pass your favorite benchmark test. Pete is more focused on the principles of computation that brains employ that current AI doesn’t. So if you’re familiar with David Maher, this is akin to his so called algorithmic level, but it’s between that and the implementation level, I’d say, because Pete is focused on the synthesis of different kinds of brain operations, how they intermingle to perform computations and produce emergent properties. So he thinks more like a systems neuroscientist in that respect. Figuring out how to synthesize these neural principles, Pete says, will lead to better AI. So we discuss a handful of those principles all through the lens of how challenging a task it is to synthesize multiple principles into a coherent, functioning whole, as opposed to a collection of parts. But hey, evolution did it, so I’m sure we can, too, right? By the way, I’ll be in Lisbon, Portugal, at the end of February for the cosign conference. So if you’re around that conference and you want to say hi, please do. Thank you to all my Patreon supporters. If you want to be able to reach out to me and influence who I invite on the podcast, for example. Or you just want to show appreciation, consider supporting brain inspired through Patreon. Go to braininspired co to learn how. All right, thanks for listening. And here’s Peter Stratton.

It was actually a Patreon supporter listener who turned me on to a particular paper that you wrote. And I was going to say recently, but when I looked it up, it was actually four years ago. I think that when you actually wrote the manuscript of the paper called convolutionary evolutionary and revolutionary, what’s next for brains, bodies and AI? And this is, I recommend the paper because it’s short, but it’s like packed full of information and ideas. So it was like a really pleasant read. And so he sent this to me with a couple of questions. I thought, oh, I should have Pete on the podcast. So thanks for joining me here.

[00:03:48] Peter: Thank you for inviting me.

[00:03:50] Paul: One of the things I find interesting, and we’re going to get into a lot of the ideas that you’ve written about, but one of the things I found interesting is that from what I understand about your background, you have a computer science and artificial intelligence background, and yet what you write about and advocate for in the paper are a lot of principles to build into artificial intelligence that are from the neurosciences. And I wonder where that came from and how you came about deciding that maybe neural principles could be important.

[00:04:25] Peter: It started in my PhD. I think I was quite interested in machine learning way back, well, in the late 80s when backpropagation first appeared, and in the did my PhD, and I was really interested in neural networks, but not so much in gradient descent. Even back then I was thinking there’s probably other ways of doing this. Gradient descent seemed like a very unbiological way of doing things.

And so I was more interested in this idea of self organization, which has been looked at in the past and then kind of forgotten.

So then I think after the PhD, I then sort of spent some time in industry. Then I came back to research, and I spent about ten years at this institute called the Queensland Brain Institute. And literally I was more or less the data scientist, looking at all the data from neuroscience experiments.

And that’s where I really started sort of solidifying my ideas around what you might call biological neural computation and how different it was to, or how different it is to standard artificial neural networks and machine learning.

And when I look at the capabilities of even simple brains like flies, insects, things like that, the Drosophila, the Drosophila melatonin gaster, that they often do experiments on, I think, has about 100,000 neurons in its brain, and yet it has a massive behavioral and survival repertoire. That’s amazing. And these animals seem to sleep and they do all sorts of interesting things. And then the bee, the standard honeybee, about a million neurons and incredible behavioral repertoire. They communicate, they do the waggle dance to tell each other how far away sources of pollen are things like this. So these neural networks are tiny compared to our large language models and the biggest models that we currently work on, but immensely capable. And I guess the interesting question is, how do they do that? So that’s what I’m most interested in.

[00:06:40] Paul: I want to come back to this because the examples that you gave are all about behavior, and we’ll come back to this idea. But one of the first things you said was that you became interested, partly at least, because of backpropagation, but then you weren’t actually interested in stochastic gradient descent. But you said in the 80s when these algorithms were written down on paper, but it didn’t work so well back then. Right? But you were interested in it even. So, the story of backpropagation, right, is that it kind of worked a little bit, but it took like 20 years for enough data to scale up essentially in computational power. And then all of a sudden, back propagation really worked to train these networks. But you saw that algorithm, and that piqued your interest?

[00:07:25] Peter: It did, because it was so different from standard von Neumann computation, right? And it was using these things called units or cells or I guess neurons, if you could call them, these little processing units, all connected together in parallel. So it was doing a lot of this idea of parallel computation, which obviously is also what the brain is doing at a very kind of like, basic level. There’s some analogies there. So that’s what got me interested. It was like, well, okay, this is something along the lines of what the brain might be doing, but it’s also got to be very different.

[00:08:00] Paul: So what is your just overall take then, given the massive success, quote, unquote, of these deep learning networks these days that are essentially built off of relatively few neural principles and then a lot of computation and a few tricks and bells and whistles, are you not impressed by them like everyone else? Or are you going to say, of course I’m impressed by them?

[00:08:24] Peter: But, yeah, I think you called it. They’re very impressive in what they can do, right, especially these large language models. It just seems to be a step change in capability, and it’s got a lot of people interested in the general public and a lot of, let’s face it, a lot of scientists as well. No one really understands how these things work. We can build them, we can train them, but how do they actually work mechanistically inside is quite a mystery.

So they’re very interesting and honestly, yet also, I’ve got to say the capabilities of these networks has surpassed anything that I expected gradient descent to be able to do. And I think a lot of people are probably in the same boat as me in regards to that.

So how has this been possible? I think it’s just a matter of these networks getting bigger and bigger and bigger. They’re ridiculously sized now and the amount of training data that you’re able to throw at them. I still believe that gradient descent is quite an inefficient way of learning. It’s just that we do have such huge amounts of data now and access to huge compute resources, data warehouses, which is what you need to actually train one of these models.

And I think that’s made it possible.

It’s not so much the technology or our understanding that it’s progressed. It’s literally just. It’s brute forcing the problem, I guess, without really understanding how we’re doing it.

[00:10:02] Paul: Yeah, I’ve been trying to avoid the term brute forcing because I think I’ve used it on the last few episodes. I was just thinking on the train ride home. No, you should use it, but I feel like I’m overusing it, so I’m glad that you used it.

[00:10:13] Peter: Great.

[00:10:14] Paul: So do you think something like gradient descent then, once we build in, these principles that we’re going to talk about in this episode will be seen as a capable but just inefficient way of doing the same kinds of things that future AI networks will do. Systems.

[00:10:34] Peter: Yeah, I mean, that’s a good question.

I would have said, I think, like I just said, it surprises me how capable these things are now and that this is even possible. So I’m kind of, like, loathed to say that we’ve reached the limit or we’ve hit the wall. Right, in terms of what gradient descent can actually do.

[00:10:52] Paul: Yeah, that’s. That’s a fool’s errand, it seems, these days.

[00:10:54] Peter: I think. So as long as the amount of data that we have access to keeps going up and our computers keep getting more powerful or bigger, bigger data centers, throwing more.

Bigger gpus. I mean, that’s really what’s driven this, then the potential is, I guess, theoretically limitless, but in practice, the fact that these models seem to be getting exponentially bigger. They’re basically growing by an order of magnitude in size every year. And the amount of power required to train them is also growing about the same, an order of magnitude every year.

And the cost is substantial. It’s got to the point now where it’s only the biggest, basically, tech companies can actually afford to train one of these models.

It’s cost in the order of 10 million to $100 million.

That’s 100 million real us dollars to train one of these large language models. Now, it’s an insane amount of money.

So how much bigger can these models get before people go, we just can’t afford to train them anymore? There’s not enough compute and or money in the world to actually do this.

Now there has to be a limit. I’m just not quite sure where that limit is. So other people have obviously recognized this, and they’re looking at ways of making training AI, or training artificial neural networks using gradient descent. They’re looking at ways of making that more efficient. And there’s certainly efficiencies to be gained in building some more specific purpose built hardware and things like that, but in terms of these multiple orders of magnitudes of improvement, I don’t see that happening. So, to come back to your original question, the world is an extremely complex place right now. Yes, learning how to generate text and even look at images and things like that, it’s achievable if you’ve got, say, $100 million in your pocket.

But the world is way more complex than sets of images and text. So if you’re looking at actually creating general purpose AI, for instance, general purpose robots and things like this, like I said, I’m not going to say it’s impossible using gradient descent, but I just feel like it’s going to be very difficult and extremely, exceedingly expensive to do that, at least with current technology or with anything that we have, sort of like even in the pipeline.

And the other problem with robots is actually gathering the training data.

There’s a huge amount of whatever petabytes of text data and images on the web, which you can basically have instant access to. But if you’re trying to gather training data for robots in the real world, it gets a lot slower. Right? And can you get enough training data for that to do? Real robotics in the real world is another kind of unknown at the moment. There’s a lot of people working on simulation environments, but again, simulation is never the same as the real world, and you can’t do that. You can’t simulate data as efficiently as you can collect it.

Not if you want it to be really real. So there’s a lot of impediments to what you might call AGI, or at least general purpose robots, if you kind of want to call them sort of the same thing.

There’s a lot of impediments to doing that with gradient descent that I currently don’t see easy ways around.


[00:14:31] Paul: This goes back to you mentioning the importance of the impressive behaviors of honeybees and fruit flies. And I know a behavior is important to you, and you’ve worked on robotics, in a sense. Do you think that Moravx paradox is even heightened now that, you know, so we have these, you know, more of X Paradox is the idea that it turns out that what we thought was hard to do is actually rather easy, like training ais to play good chess and go and stuff. And what we thought was easy to do, like catching a ball, is actually difficult to do. Perhaps that’s even more highlighted now than it ever has been. I don’t even know the current state of robotics, so maybe you can catch me up to speed on that. I just know that they have traditionally been lagging behind these systems that don’t actually need to behave in any way.

[00:15:23] Peter: Right. So there’s a few different angles to that.

So there’s even some work going on here in Australia, in the institute that I work at, where there are some researchers trying to make robotics move more fluidly, more like people. So trying to get.

If you look at a typical robot, if you give it the task of going to a table and picking up a can, say it moves, for want of a better word, it moves very robotically. It will approach the table, then it will stop, and then this lever will come out with a gripper on the end, and it will move slowly and in a straight line towards the can, and then it will stop, and then it will grasp the can, and it just looks robotic. Right. And so there’s clearly different ways that biology solve these problems, and there’s people here working on that. So now we’ve got quite an impressive demonstration that I wasn’t involved in at all, but basically, where a robot can be moving past a table quite rapidly and fluidly and dynamically, actually grasp something on the table as it goes past. Now, this is quite an impressive feat. And when you see the videos, it looks really different, it looks organic.

This has not been solved using terribly many neural networks or biological sort of, motion and muscle principles. It’s still going back to kind of like first principles. In robotics, but doing it in a very, very clever way.

[00:16:59] Paul: What are those, like, feedback loops?

[00:17:01] Peter: It has to be closed loop, for a start. Yeah, and you’re continuously sensing your position and your gripper position relative to the object you want to grasp, and it’s all about closed loop and dynamic updates and things like that, which is kind of like a no brainer, really, when you think about it. And when you see the demo, you go, yeah, why hasn’t it always been done that way?

And again, this is one of the first problems in robotics. It’s a lot about sort of breaking a problem down. It’s this, like divide and conquer, and it’s the stepwise refinement that engineers like to do to kind of, like, isolate problems and go, okay, well, the first step is to get to the table, and then the second step is to work out what I need to grasp, and the third step is to put my gripper close to the cup and things like this, and engineers like to break this down, and that’s not how biology does it. Again, it seems to be quite often the case that if you take inspiration from biology, you can actually solve these problems in a much more dynamic and robust fashion. And I think that’s what the team here has been trying to do. So that’s one of the issues, I guess, with robotics, it’s just not fluid. It’s too overly engineered, which basically means that it looks robotic, and the solutions are often brittle.

Now, obviously, the brain sort of uses the same principles of robustness and closed loop control, but the processing behind that is very different.

And that leads into the second thing about robotics and why I think robotics is probably useful for AI is because our intelligence is actually built on controlling our body in the world.

This whole capacity we have for things like abstract thought and language and all that stuff, the amazing high level abstractions that the human brain and other higher animals can do didn’t come out of nowhere, and it didn’t appear in an isolated sort of environment. It appeared in the world, in the physical world, in bodies that are actually needing to survive in the world.

And it seems like a lot of the principles, the computational principles that are used by much simpler animals are actually kind of, like, adopted and co opted into the higher level processing that our brain is capable of doing. So if we want to understand intelligence, whatever that actually means, I mean, it’s obviously very hard to define. Right. Thank you.

[00:19:44] Paul: Yeah.

[00:19:45] Peter: If we want to actually understand it and even come up with a good definition for it, right. Then we really need to be looking at simpler animals and how they process sensory stimuli and control the body in the world, and then how we can co opt some of those principles for higher level processing. And that’s exactly what our brains seem to be doing.

So that leads to this. Maybe you wanted to bring it up. I’m sorry, maybe later. But it’s the idea of the embodied Turing test that quite a few eminent AI researchers have kind of alluded to recently in a nature paper a couple of years ago, basically saying that if we want to understand intelligence, then we really need to understand how simple creatures survive in the world and then build up from there, rather than diving straight into these high level, sort of like symbolic reasoning processes that the human brain is capable of doing.

Start simple and see how that leads to our capabilities. Right. And the understanding of the nuts and bolts behind the high level reasoning that we’re capable of doing is what we’re currently lacking.

[00:20:50] Paul: So there’s a lot there to respond to and digest, actually. So I recently had my friend John Crackhouer on the podcast, and he’s a neuroscientist, and he actually argues directly against am.

I love the idea of thinking of higher cognition and thoughts as analogous to internalized movements. Right? And it’s interesting that we started off talking about feedback loops and robotics. Actually, I had a guest on many moons ago now, Henry Yen, who thinks that neuroscience essentially needs a revolution, like a paradigmatic revolution in the cooneyan sense, in terms of thinking of cognition, as just nested hierarchical feedback loops, essentially all the way up into our cognition, which is in line with that idea of thought as being analogous to internal kinds of movements or actions. Right? But John argues that he thinks it’s ironic, for example, that the latest great AI is built on not on things like that. Not on things that lower, lower animals, other smaller mammals, et cetera, are capable of doing, but are built on what we as humans are capable of doing, and that there might very well be. And he argues that there is a break between understanding how things move in the world and those principles and the principles of higher level cognition, that they’re just worlds apart to him. And he doesn’t agree with the idea of the embodied inactive. The four e approach to studying as a gateway to studying higher cognition, essentially.

Do you think that there’s, like, just a stepwise way of going from what you. From simpler animals, that’s a better way to say it. And studying how they survive and move in the world to higher cognition? Or will there need to be different principles that we’ll have to grapple with and understand.

[00:22:57] Peter: It’s interesting. I find it interesting that a neuroscientist will actually have that point of view to me, when you look at the brain. Right. And like I said, I spent ten years in a brain institute, this idea that what you might call thought is really just internalized and abstracted action. Right. Which is something that you just said, and it’s also something I mentioned in the paper that we’re discussing that just makes sense from.

There’s a lot of evidence that that is kind of how the brain works, and it’s how we behave as well.

So in terms of how the brain works, if you look at the brain, the frontal lobes, where basically most of our actions are represented, you could say at a very course level, the front of the brain, the frontal lobe, is more for action and movement, and the back of the brain is more for perception, sensory perception. So that’s obviously a very coarse generalization, but you can start with that. And then if you look at just forward of the.

What’s it called? The central sulcus of lost central sulcus. Yeah. That’s the line that goes this way across your brain that basically divides the front half from the back half of the brain. So directly forward of that is basically your muscle representation. You have a representation of all the muscles in your body. And again, it’s like the sensory homunculus. It’s basically laid out in the shape of a body in your brain and triggering those neurons. If you stimulate one of those neurons with an electrode or something, it’ll cause a muscle twitch somewhere in your body. So you can map out this body map in the brain, the homunculus. Right. So as you go forward from there, you go from motor. That’s called motor region m one, as you go to m two. Basically, these are combinations of muscle actions that occur over time, say, so repetitive, easily repeated, well memorized movements are represented there.

And these neurons tend to fire, like, a few hundred milliseconds before the actual motor neurons fire when you’re about to do an action. And then as you move further and further forward in your frontal lobe, basically what seems to happen is you’re getting longer and longer sort of temporal representations, so representations in time of movements that you are potentially going to do further in the future. So you potentially have an I’m going to go shopping kind of like neuron somewhere in your brain. And obviously it’s not a grandmother neuron, and it’s not that simple. But there is some representation in your frontal lobes of doing the action of going shopping or driving a car or things like this. Some very high level neurons in your brain and all these things are basically, you could call them abstracted actions or delayed actions or compound actions or something like that. You could also call them, in a way, concepts, right? And this is where a lot of concepts are actually tied to what that means for us as an embodied agent in the world.

And these high level concepts really stem from just delayed and abstracted actions. And you can see that even, like, phylogenetically and evolutionarily, in the way that brains have developed over time.

The simpler vertebrates, for example, have very small frontal cortices, or almost no frontal cortex at all.

And then the more ability you have for abstract thought, it’s this frontal lobe that’s really exploded in humans, right? And it literally just builds off the simpler actions that are represented in the brains of simpler animals and also in our brain, in the lower level regions of the frontal cortices.

So that’s the evidence from the neuroscience perspective, in any case, that this is what’s going on. Everything stems from low level actions, in terms of the structure of the brain, from low level concepts and low level actions, up to high level concepts and high level actions. So if you think about it, everything we do and everything we think is really just ultimately leading to some sort of action, right?

It can be action like, we can make plans for years or decades in the future, and that’s our brain. Simpler animals don’t have the capacity to think that far ahead.

And the very simplest ones are simply like stimulus response, like mollusks and things like this, right? They have very simple nervous systems, but no ability to abstract into the future.

So there’s definitely evidence from neuroscience that kind of shows this. And also when you look at, even at the structure of the brain across the cortex, from these low level motor regions up to the very high level regions, the structure is also really, really strikingly similar across the entire motor or frontal cortex. But also even in the sensory cortices, the actual structure of the brain in the cortex itself is remarkably similar. Whether that part of the brain is devoted primarily to sensory processing or primarily to coordinating muscle movements and motor actions and abstract thought, it’s very difficult to tell, unless you’re an expert, by looking at a slice of cortex and by looking at sensory cortex or motor cortex, it’s often quite hard to tell. So there’s really, really similar processes going on in all these different parts of the brain.

So if you’re saying that there’s something very particular that humans can do in terms of abstract thought, you need to explain why that would be the case when the actual substrate that seems to carry that out is the same across all brains.

[00:28:59] Paul: Yeah. I’m not saying that. He’s saying that he has the solution, for example, and we don’t need to harp on this for long, but the example that he regularly gives is, if I tell you to go to imagine standing outside your own home and then walking through the door and going to the kitchen to get a butter knife, then opening the fridge, you can imagine all that without moving at all. And that there’s simply no evidence that any other species can do this. Right.

That’s the most parsimonious explanation. And so we need to be able to explain those higher level kind of processes. And he doesn’t believe that there’s a continuous path from simple movement to high level thought. Right.

[00:29:43] Peter: Okay. Yeah. I mean, that’s an interesting point of view. And obviously, I wouldn’t call it contentious, but this is undecided. Right.

And anyone who says that it is decided, obviously, I would disagree with. Right. That’s the one thing I can say for sure.

So, yeah, there’s a lot of stuff that we just don’t know about the brain, and I think this disagreement that I would have with this guy would obviously be just an indication of that. So it’s no problem. It’s just the way science works.

[00:30:13] Paul: Right. Yeah.

Quickly, I haven’t asked this of anyone in a while. Where are we, do you think, in neuroscience and understanding the brain? Are we at the beginning? Are we in the middle? Are we close to it?

[00:30:25] Peter: That’s a difficult question. I would say it’s a bit of a cliche to say, but it’s. The more we understand about the brain, the more we know that we don’t know. Right.

[00:30:35] Paul: God, I said this exact phrase. That again. That’s another thing I’m trying to not say very frequently, so I’m glad you’re saying it.

[00:30:42] Peter: I got to apologize. I knew that was a bit of a cliche, but it’s kind of true, right?

[00:30:48] Paul: Yeah.

[00:30:48] Peter: So my kids have asked me the same question. How far are you to creating agi or whatever? And it’s along the lines of don’t really know, but it’s somewhere between probably 1000th of a percent and 1%, probably more like a thousandth of a percent.

And our actual understanding of the brain is probably similar. I would say we understand so much about the components of the brain, down to basically the molecular scale, the gates and the neurotransmitters.

The chemicals, all the neuromodulators, and all this sort of thing, we understand exactly, almost exactly how they work. We can even simulate the molecular movements on a supercomputer to show how the gates work and things like this. But how it all holds together. Where we really fall down is in the emergent dynamics and the emergent properties of the brain, and that’s what we don’t understand. And is that neuroscience even, or is that computation or engineering or physics? There’s a lot of physics involved in this as well. Physics is a lot about emergent properties and emergent dynamics even.

So, in the end, it might not even be a neuroscience problem, or you’d have to redefine what you mean by neuroscience. It’s more like what I’m interested in, which is neuroscience and neurocomputation, I guess, and neuroai. But it’s really kind of like a separate field to neuroscience, because we take what we know about the brain from neuroscience, and then we try to construct models and understand the emergence that’s going on.

[00:32:30] Paul: Before we get into. Because I wanted to just jump right into the ideas in your paper. But before we get into that, how would you describe your overall goal?

[00:32:43] Peter: Well, that’s a good question, too, I guess. On what level?

[00:32:51] Paul: Well, the reason why I ask is tied to the way that you present all these principles in the paper, and it’s tied to what you just said about how all these things hang together, and that’s what we’re going to get into in a minute. So then I wonder, like, well, it’s an ambitious goal to synthesize all these principles, and we’re not going to go through all the principles that you talk about in the paper. But I thought, well, is that your goal, to kind of step by step incorporate these things until you have them all incorporated into the. And build these emergent systems and emergent computations?

[00:33:30] Peter: Look, given enough time, but perhaps like a semi infinite life, I guess that would be my goal. Yeah, but given that I’m probably not going to have time to do all of that, I guess the goal would be to kick start that process and to get people aware of what we do know about the brain and how that we can actually leverage that into building better AI, which is what this paper was all about.

So at one level, my goal is to just enjoy what I do and follow my curiosity, which is kind of like, I guess, the ultimate goal of pure science. But then I guess you also need to have concrete goals in order to keep yourself honest and actually do something useful. So in terms of that, yes, it would be to take these principles and start putting them together and then understand it. You know that each principle alone doesn’t do much. Right. It’s only when you combine them that you start getting kind of like potentially building useful emergent systems. And so you need to try to understand that emergence. And like I mentioned before, that’s where we really fall down in these complex systems. You can’t write an equation that actually solves these things.

You need to build it and understand it from an emergent point of view.

[00:34:55] Paul: Okay, well, like I said, it’s ambitious, and I’m about to read 13 different principles that you list toward the end of your paper at the risk of losing the listeners. But I do it just because to highlight the ambition and the difficulty of going down that road and trying to incorporate all these things so that they hang together. Okay, so sparse spike time coding, self organization, short term plasticity, reward learning, homeostasis feedback, predictive circuits, conduction delays, that one’s not often mentioned. Oscillations, innate dynamics, stochastic sampling, multiscale inhibition, k, winner take all, and embodied coupling. So I don’t even know where to begin.

[00:35:41] Peter: What does it all?

[00:35:43] Paul: Well, yeah, well, it’s all in the paper, so I’ll refer people to the paper where you discuss all these principles in more detail. But the idea of. So you just got done saying that we understand so little of the brain, but then there are these 13 principles that you find are probably super important to be able to build not separately, but together, and have everything work in a synthetic fashion to generate these emergent properties of complexity.

So maybe we know more than we think we do if we can point to these principles.

[00:36:21] Peter: Personally, I would like to think so.

I tried to make the list exhaustive, but I think that’s rather conceded to actually believe that it would be right. I think there are principles that we don’t currently understand, that we have no idea of the potential existence of or the requirement for. So I imagine that list is going to get bigger at the moment, I think, yeah, I’ve tried to basically capture everything, all the high level, kind of like mechanistic principles that the brain seems to be using to perform computation.

And these are the ones, I’m obviously not the first one to talk about these principles.

Each one has been independently researched. And like I say in the paper, some of them are quite extensively researched. I might be the first one to put them all together, but again, that’s probably not true.

So it’s just a matter of how much of the brain can be understood in terms of these principles. Now, obviously, you can’t understand the principles individually. You do need to understand the principles individually. But if you want to look at how the brain works individually is not going to cut it. We’re going to need to combine them and have a look at the resulting.

I keep using the word emergent, so I’m overusing that now. But look at the resulting properties of the system as a whole. And that’s the only way to actually understand, to understand neural computation when we do that. So when we do start putting them all together, and this is one thing that I do say in the paper, there haven’t been very many attempts to actually put these principles together, even though individually they’ve been, some of them are reasonably well understood.

The real value in understanding these as mechanisms is when we try to unite them, as the brain has done quite successfully. So when we start doing that one, I think it’s going to be quite likely that other principles will emerge, or at least the requirement for other principles will emerge, because we’ll put these things together and the models still won’t be working like the brain, something will go awry, or something will just not work at all. And in which case, we have the ability then to look at the model and go, well, what seems to be missing?

And then we can look back to neuroscience. And this is quite often what happens. You build a model and, for instance, a model of STDP, this spike timing dependent plasticity, and then you realize that, okay, STDP is pretty much unbounded. These synaptic connections are more or less unbounded in pure theoretical sTDP. And that can’t happen in the brain. There has to be some sort of, like, normalization process going on. And of course, we have normalization in our deep networks, right? And that comes about purely through an actual practical requirement for it.

And the same happens in the brain. So then you start looking at things like, okay, there must be homeostatic processes in the brain, and it turns out there are, there are a heap of them that do basically the normalization. They do it in a more biologically realistic way, obviously, rather than just summing weights and dividing by the number of weights and whatever, doing l one or l two normalization, the brain is not doing that, but it has a similar ultimate functional goal of normalization in artificial neural networks. So that’s where the homeostasis comes in. You realize you need the mechanism so you can look to biology to go, okay, how does biology accomplish this? And quite often the information is there already in neuroscience, you just need to basically digest and absorb that into your model, and there will be other things that, other principles that we’re going to need to incorporate. I can be almost certain about that.

[00:40:10] Paul: The ones that I just listed off also are they do sound like a bottom up approach, right? They sound like kind of low level mechanistic processes. But you write in the paper that what you’re taking is a side in approach. Yeah, sideways in approach. So maybe you just describe that because it does sound like a list of lower level processes.

[00:40:31] Peter: Well, they’re sort of low level in as much as you can name them. And the homeostasis in some cases is like normalization, but mechanistically, that can actually embody quite a complex biological process underneath that. And that doesn’t necessarily need to be modeled from a bottom up point of view.

So the functional goals of some of these mechanisms are potentially implemented with very complex biological mechanisms. You don’t need to actually do that in a bottom up faithful sort of a fashion. You can sort of abstract the functional goal out and just do that in the simplest computational way possible. So that’s kind of what I mean by the sideways in. So when it comes to the modeling system called neuron that models physical neurons and channel densities and this sort of thing, ion channels, ion channels and dendrites and axons in actual their full 3d glory, I don’t think we need to do that to understand at least the basics of neural computation.

There’s obviously nuances there. And I think if you wanted to actually reproduce a specific brain, for example, well, then yes, you’re going to need to model those things. But if you just want to model neural computation at a core scale, you probably don’t need to do that level of biophysical detail, I would say. So that’s what I mean by not doing the bottom up. Some of these processes, like I said, very complex biologically, but we can actually just abstract them away.

And as long as you get that, the really important thing then is what level of abstraction is sufficient and what level of abstraction is too much?

Because at some point, if you abstract too much away, then you start losing some of these emergent properties, and you don’t want to do that.

[00:42:39] Paul: Yeah, well, that was maybe my next question is you look at a list like this and these are all biological processes, and then you think, okay, I’m going to abstract away what’s unnecessary from them all and then somehow piece them together. But this is something that evolution has done over eons and eons and eons, and it just seems like a tall order for us to, and maybe a bit of hubris for us to be able to say, okay, well, I’ll just take all of these things and I can just program oscillations in this one program, and then I’m going to merge it with homeostasis in another program and reward learning with another program at an abstract level that feels comfortable to me without even knowing whether the biological processes, what levels are important for their integration. Right.

[00:43:28] Peter: Yeah, that’s obviously a very good point.

And that comes back to probably other mechanisms will be required.

Probably some of these mechanisms will need to be broken down into simpler submechanisms when it comes to it. And I think it all just comes down to the fact that we don’t know what these models will do until we build them, right? Because there is complexity science trying to understand the science of complexity and emergence is really in its infancy, right? So it’s just one of those things where you’ve got to build it and see and, well, what’s it not doing, what’s not working? So when I say oscillations, I think oscillations are themselves an emergent property of neurons and networks and spikes and inhibitory GABA expressing neurons in combination with the excitatory amper expressing neurons and NMDA. And I think all these things, oscillations themselves are an emergent property. I wouldn’t want to be building oscillations of a given frequency into my model. I’d want that to be an emergent property, because I think that’s really important for this communication through coherence idea, where basically what that means is that different brain regions will oscillate in synchrony when they need to communicate, and they’ll go out of phase or oscillate at different frequencies when they don’t need to communicate.

And it’s a very, very powerful idea. And this is one of those ideas that I think hasn’t really been used in AI at all.

We can simulate the dynamics of these processes in the brain, but in terms of utilizing them for computation, as far as I know, no one’s ever tried to do that, at least not at more than a very sort of basic level. So I wouldn’t want to program the oscillations in. Oscillations are something that I want to have emerge because you never quite know when two brain regions are going to couple through synchronous oscillations or phase shifted oscillations as they need to do. So these sorts of things are just things that you need to build and see how it works.

[00:45:37] Paul: But so oscillations is a good example also, where there’s plenty of work looking into the potential causal properties of oscillations. So yes, they’re an emergent property, but then they can have a top down effect on these lower level components, as well as being generated by the lower level components. So there’s this interlevel play that again, it seems like it’s going to be daunting to find that right excitatory inhibitory balance within a layer, coupled with the homeostasis, and for those oscillations to have the right effects when they need to have the right effects, for the honeybee to do the waggle dance.

[00:46:17] Peter: Right. Good point.

I think this is actually something I allude to towards the end of the paper, is that there is this idea that the more principles that you start incorporating into your models, the more complex the model becomes and the harder it is to tune or to get to do anything.

And that’s certainly a relevant concern. But the opposite might also be true if you do incorporate the right principles. For instance, some of this homeostasis at potentially multiple different levels, because you mentioned levels as well. And that’s another very interesting point. But if you do incorporate these principles, sometimes they become synergistic and they just work when you get the combination just right.

And the homeostasis is a good one.

In the models I’m building right now, I do have some simple homeostatic processes going on, and it’s amazing.

[00:47:15] Paul: These are the spiking neural network model.

[00:47:18] Peter: Network model that I’m currently working on.

And it’s amazing how far you can push some of the other parameters in the model in terms of things like synaptic weights, for example, or the number of connections each neuron has, you can change some of these parameters that should wildly affect the dynamics, but in the end, ultimately they don’t. The network just finds a happy operating point somewhere in the parameter space due to the interplay between things like the homeostasis and the causal strengthening of synapses between neurons. This spike timing dependent plasticity, it will simply find a happy operating point due to the interaction of these mechanisms. And if you take one of the mechanisms out, the model then actually becomes quite unstable.

So it’s actually, when you put all these. When you get the right combination of mechanisms in there, it actually all just seems to work. And I really believe this is exactly what’s happening in the brain. The brain is at this sort of like. And this is where a lot of the physics comes in. And this is one reason why there are some eminent physicists who have gone into neuroscience, because the brain dynamically seems to be operating at a critical point in its dynamics, at a phase transition, basically somewhere between seizure and random noise, seems to be where the brain is operating, and our brains thread that needle really well. And that’s why.

[00:48:49] Paul: Robustly.

[00:48:50] Peter: Very robustly, yeah. Due to, it would seem, a lot of these homeostatic mechanisms and things need to be pushed way out of the operating region in order for diseases like epilepsy, for example, to manifest, because the brain is just so good at maintaining its dynamics in this critical region of phase space. And so the reason it’s able to do this is through these homeostatic mechanisms. And if you can identify the right mechanisms, I think it will all just self adjust. Right. So the reason that spiking networks have been thought of as being difficult, impossible to work with, and everyone, basically everyone who. A lot of people who’ve had a go at modeling spiking networks ultimately kind of, like, give up and go back to either deep learning or doing gradient descent with spiking networks, for example, to try to shoehorn that learning in there.

And the reason being that they haven’t actually just got the right combination of homeostatic principles and some of these other mechanisms. When you get it right, it just works.

And that seems to be a fundamental principle of the brain that we really need to be cogent of. I think you get that right, and then suddenly things start falling into place.

[00:50:07] Paul: This is a bit of an aside, but I’m sure some listeners are thinking, well, what about all of the psychological difficulties people are having, especially these days? If it’s so robust and it just falls into place, why is psychosis up more than ever? Why does everyone have ADHD? Why are we all depressed?

I’m not sure if that’s a fair question to ask you at this point, because I know you’re not a psychologist, but one does wonder, thinking of higher cognition, and is there a continuum, or is there a break and it’s an emergent sort of property? And then, oh, are we really that robust? Or have we passed some point that now we’re skating on thin ice, et cetera?

[00:50:48] Peter: Yeah. Okay, so now we’re kind of getting into, I guess you might call it social neuroscience. I’m certainly interested in all the different neuroscience fields. I don’t know terribly much about social neuroscience, but I think, yeah, we are pushing the limits of what our brain has sort of evolved and is capable of adapting to.

The world has changed so much in the last, even 20 years and definitely incredibly unrecognizable from, say, 150 years ago in terms of the stress and strain we’re putting on our brain with always on and always connected and the social media and things like this, it’s just not a natural environment. It’s not what the brain.

[00:51:33] Paul: Maybe it’s too much leisure time now. Maybe we have it, too. Things are too good. They’re better than they’ve ever been, so we don’t have to think about where to get food, except for the shopping list.

[00:51:43] Peter: Right. Yeah, I think there is definitely evidence for that, too, in terms of. Yeah, the incidence of, say, depression and anxiety is way lower in developing nations. Right. When you are actually concerned about your day to day survival, you tend to not get depressed. You can be living in horrid conditions.

[00:51:59] Paul: Right.

[00:52:00] Peter: But because you actually have a literal goal to survive each day, you literally just don’t have time to, I guess, to get depressed, more or less. You’re too busy living. And it’s only when you have sort of like, this time to ruminate, self reflect, and be on your screen and get bored watching TikTok clips and things like this, that’s where these kind of, like, psychological problems seem to start creeping in. And I would still go back to the fact that it’s simply because our brain has not evolved in that sort of an environment. We are not adapted to live in a high tech world.

I love getting out to nature, especially like the beach and swimming in water and things like that. Right. It’s so calming and so relaxing. You just feel good doing it. And I think there’s a lot to be said. I mean, you could get into the paraphysical interpretations and things like that that a lot of people do, but I don’t do that. I think it’s just because it’s where we grew up as a species, and I think it just feels comfortable being there.

[00:53:10] Paul: Yeah. I’m not going to pretend to be a well being guru or anything, because there are, like, debilitating psychological disorders that have nothing to do with TikTok and social media. So I’m not trying to shortchange psychological disorders as being all about our modern leisure, lackadaisical lifestyles or anything, but there’s probably something there.

[00:53:36] Peter: Yeah, I agree. And there’s also a lot of, as you say, there’s a lot of psychological problems that have been around forever. As well. And again, I think there’s a reason for that. If everyone was the same, if there was no kind of like, we need a distribution of traits, of phenotypes, basically, in any species. Because if everyone is specialized to excel in the current environment, right, if everyone is perfect for whatever it is you currently do as a species, then if things change, if there’s some sort of, like, a catastrophe or a new predator appears or whatever, then there’ll be no one that is kind of, like, adept at dealing with that within your species, and your species is likely to go extinct. So you do need, for instance, an obvious one, is you need people who are extroverted and bold and fearless, right? Because when times are good, they’re the sort of people who are going to do really well in an environment, right? The ones who take risks.

Whereas if things are not looking good for you, like environmentally speaking or there’s a new pathogen or something, what you really want are these fearful, introverted type people who keep to themselves and don’t like socializing. Say there’s a new pandemic, for instance, back in neolithic days or something like that.

You want people who basically are very fearful of this in order to. Because they will probably, in that case, be the ones who survive. So you need a broad range of phenotypes, and it’s a very clinical and scientific way of looking at it. But that is one reason why you get such a diverse range of people. And unfortunately, sometimes these phenotypes just push the limits so far that fear leads to anxiety.

A constructive fear actually leads to more or less full time anxiety and phobias and debilitating psychological trauma and things like that. And it’s simply a consequence, I think, of. Like I said, it’s very scientific and clinical, but it’s a consequence of having to maintain diverse phenotypes within a population.

[00:55:44] Paul: So, Pete, I have you on record now saying that introverts are fearful and extroverts are want. I don’t want that to be the.

[00:55:54] Peter: Not entirely true.

[00:55:55] Paul: Obviously, I’m a bold introvert myself. I’ll say, yeah, we could go down a very dangerous road here and get both of us in trouble, I suppose.

But speaking of psychology, you’re taking this sideways in approach.

Another approach is this kind of top down approach, the functionalist approach. Right. Which is, in large part, how large language models and current deep learning have had the successes that they’ve had because they’ve been trained to do particular tasks. And it’s through the learning of those tasks that their weights all get set in just the right way to perform the task very well. Right. And so in that respect, it’s more about the psychological. I’ll use that term in air quotes function that a network, let’s say, is tasked to perform.

I’m wondering if. Because I don’t think that you really talk about using that kind of functionalist behavior, top down psychological function approach to sort of constrain models. Right. A lot of what you talk about are these principles of neural computation. So are you more interested in the computational capabilities or the actual functional outputs? Functional capabilities.

Another way to say that is, do the psychological functional capabilities inform or guide you at all, like, in thinking about neural computation?

[00:57:26] Peter: To answer that, I could potentially go back to this idea of the embodied Turing test.

[00:57:34] Paul: Sorry, could you just explain the embodied Turing test? I don’t think that. You explained it earlier.

[00:57:38] Peter: Right, sorry. So it’s this idea that. Yeah, the basic idea is just that to understand intelligence, we need to understand the building blocks of the brain. And to understand the building blocks of the brain, we really need to go back to what simple creatures do from, say, whatever the simplest, even simplest animal with a nervous system. Like a sea slug or something like that. Or sea elegance. Yeah.

All the way up to us. But everything that’s in between, because there are many things that are conserved from the simplest nervous system up to the most complex.

[00:58:19] Paul: But what’s the Turing test part of that?

[00:58:21] Peter: The Turing test is basically, the idea is you build robots that can do what animals can do. Right.

[00:58:30] Paul: The Turing part of it is that I can’t tell that octopus robot from a real octopus or something.

[00:58:35] Peter: Correct. In terms of its behavior.

That’s right. So the way they frame it in the paper is more or less like animals have these competencies that we cannot currently replicate, even with our best technology.

And obviously, these competencies are a function of their brains. Right. But also their bodies. And the fact that the brain is coupled with the body. And, of course, this is a fairly pervasive idea.

Everybody knows about it, but very few people actually pay it much heed or use it in their research. Right. And so what this paper was calling for was, okay, we actually need to start doing this in order to understand how the brain works, you really need to start simple and put the brain in a body and see what it can do when it interacts with the environment. And so it becomes this entire. The feedback loop, the closed loop between the brain, the body, and the environment, back to the brain. Resensing the brain does perform some action in the world, it changes the next perception. The brain then processes that perception, then decides on the next action and so on. Obviously, it’s not that simple. And there are multiple loops, abstracted loops, going all the way up, and they all interphase and influence each other, top down, bottom up, and sideways in. Right. And that’s what we need to understand in terms of understanding intelligence. So it’s a really completely different approach, as you said, to something like a large language model, where it’s basically just, well, let’s just throw all this text at it and get it to predict the next word.

And that’s all that the large language models are doing, but they do it extremely well. And it turns out that in doing so, they’re formulating concepts like humor and even like a concept of awe and inspiration and all these things that are very, very hard to define and pin down. But if you tell a large language model to make up a joke, it will. And occasionally, they’re even a little bit funny.

So they’ve got some sort of idea of what humor actually is.

[01:00:45] Paul: Right.

[01:00:45] Peter: And they know what humor is when they see it, and they’re able to even generate it, novel instances of it, which is amazing. It’s amazing what these things can do. But I think that’s only.

It’s a small fraction of what the brain can do. If you want to put the parallel there between. Okay, what part of the brain is a large language model actually emulating? I would say it is literally just the feed forward connections in the cortex.

[01:01:13] Paul: Wow. Okay, right.

[01:01:14] Peter: And I think, now, how would I justify that? Well, large language models are typically there. They’re all just feed forward. There’s a huge amount of feedback in the brain as well, which is completely ignored in almost all our AI. We have very few effective, recurrent models, I guess, in AI.

So these large language models are very good at extracting high level features from their input.

And I think that’s exactly what the cortex is doing, both sensory and motor. So it’s perceptual feature extraction is what the sensory cortices are doing. So you take individual edges at low level, say visual cortices, and they’re combined into things like maybe like simple shapes, and then shapes are combined into things like faces and cars. And you have representations of these objects in your sensory cortices. And I think exactly the same thing is happening in your motor cortex. It more like temporal extraction, temporal feature extraction, which also happens in. You have temporal, which means features that occur over time, like moving representations of moving objects and things. Like that, you get those in sensory cortices, in your motor cortices, you have representations of actions over time, and it’s doing motor feature abstraction. That’s sort of what I’ve called it.

So you get perceptual feature extraction in sensory cortices and motor feature abstraction in your motor cortices, and it’s the same process just applied to different inputs, basically.

So if we want to understand the brain, we really need to understand those processes, and if we want to understand computation in the brain, so neural computation, we need to understand those processes. And you’re not going to do that just by throwing a huge text corpus at a feed forward network. You’ll get a part of the brain, you’ll get, like, I think, the feature extraction, the feed forward feature extraction, that’s what you’re getting in a large language model, but you’re not getting anything else. Feedback connections outnumber feed forward connections in the brain. So the feedback is obviously at least as important as the feed forward, possibly more important than the feed forward connections. Then you have all these subcortical connections as well, and then you have the body that the brain isn’t tightly innately coupled with. All these things influence and control and actually dictate what computation in the brain actually is. And a large language model has none of that. So you need to come at it from, and this is what I’m advocating for, you need to come at it from a holistic perspective. Right.

[01:03:56] Paul: We’re not good at that.

I’m not good at.

[01:04:02] Peter: It’s difficult. It’s very difficult.

[01:04:04] Paul: It’s not out yet, but I just had a researcher, Dan Goodman, on the podcast, and he specializes in working with spiking neural networks. And he actually thinks that now is the time to be excited about spiking neural networks because of a recent development in training them, which kind of mimics backpropagation called surrogate gradient descent. But it is a global learning rule. Right. And it still uses gradient descent, but he’s found a lot of success in how it works. But would you agree that now would be the time to be excited about spiking neural networks? Or do you think, because you mentioned earlier, a lot of people shy away from them or dabble in them and then go back to what works because they’re tricky and can be difficult and prickly, et cetera. But is that coming around? Are we going to see more and more spiking neural networks these days?

[01:04:55] Peter: They’re still a very niche kind of research project.

The problem is that for specific tasks.

Deep networks, deep learning, and gradient descent works really, really well. There’s no argument there. And it’s almost like a black box.

You can throw a problem at it. And because there’s been so many really smart people at Google and various other places working on these python packages for doing deep learning, it really is. Literally, a novice can almost come along now and construct a deep learning model.

Spiking networks are not at that stage yet. It’s not that they can’t be. It’s just that the research effort hasn’t been applied to them as yet.

[01:05:43] Paul: Well, one of the things he says is they’re still really difficult to scale up because they’re computationally really super expensive still. But maybe that’s using that gradient descent learning rule and maybe the local learning rules that you’re implementing reduce that computational.

Computational requirements.

[01:06:00] Peter: Yeah, I mean, I think that’s certainly part of it. It also comes down to just not understanding all the principles well enough. Or at least there’s a lot of little tweaks and heuristics in deep learning, like the dropout and things like that, that just seem to work in practice, and they’re well implemented and you don’t need to worry about it anymore.

That sort of overall tweaking and heuristics and polishing to the nth degree of the spiking network packages, they barely even exist now, let alone having been tweaked to the nth degree. They’re just not, and they’re not out there. So that’s a big impediment, I think, to further research. And the problem is that the longer this status quo is maintained, the further behind spiking networks will get, potentially. So it’s more or less a labor of love, and you’ve really got to embrace the vision and believe the vision of what spiking networks can be before you leave and spend time actually working on them.

And there’s probably just not a lot of people who really believe it that deeply, I think, to actually take a gamble on their career and go into spiking nets.

They are potentially. It’s been shown theoretically, that they’re more powerful.

It’s rigorous. They’re rigorously more powerful than artificial hill networks, but people think they’re finicky and difficult. And yeah, sure, theoretically, they might be more powerful, but in practice, they just don’t seem to work. And this is what people will say. So that’s, I guess the medium term goal is to kind of, like, leapfrog these chasms in functionality and understanding and get spiking networks to the point where, okay, right. They do seem to work and they do seem to have some advantages. And when it gets to that point, then I think people will start embracing them a little bit more easily and more often.

[01:07:58] Paul: And more often, I want to hear your kind of own, and I’m sorry to put you on the spot if I’m putting you on the spot, but your own roadmap in your head, right? So now you’ve got like a pretty good spiking neural network model. And so are you going to take that and then start adding other things more to it, or are you going to then build a different kind of model with, say, innate dynamics or something, the vision that you have to implement some innate dynamics or some self organization and then try to marry them? Or do you see a roadmap for yourself moving forward? What’s the low hanging fruit for you? What’s the really difficult thing that maybe the light at the end of the tunnel, do you see that vision?

[01:08:39] Peter: So I think as a scientist, you tend to oscillate, right. Sometimes you’re looking at the big picture and you try to build this big roadmap for yourself. And then other times you need to really bury yourself in the details, right.

To make notable progress on any one of those particular problems. Right.

[01:09:00] Paul: You really need to say reality slaps you in the face then.

[01:09:03] Peter: Right. So this paper, as you said, I think I wrote it five years ago. It was only published last year.

I didn’t submit it for years. It was just sort of like lying around.

So that paper five years ago was my attempt, obviously, at the grand vision. Now I’m kind of buried in the details. So when you ask a question like, what’s the big roadmap?

[01:09:27] Paul: Well, what’s tomorrow bring answer.

[01:09:32] Peter: Yeah, exactly. It’s a difficult question. I think a lot of it is spelled out in the paper in terms of incorporating all these mechanisms, 13 of them, I’d never even counted them. So that’s interesting.

[01:09:43] Paul: Yeah, well, I actually enumerated them because they’re listed, but the order of adding them. Right. So that’s why it’s so ambitious, because how do I even begin? What is the easiest thing to begin with? What are the two things that I think I could get to integrate well, in a model to get them working?

It’s difficult to know what’s most important on the list.

[01:10:09] Peter: Right.

[01:10:10] Paul: What’s easiest, et cetera.

[01:10:12] Peter: The models I’m currently building are using sparse spike coding, spike timing, dependent plasticity and homeostasis so firing rate, homeostasis. So each neuron has a set firing point, that number of spikes over time it would like to emit, and you have a range of firing rates across the network.

There’s also weight hermeostasis, but it’s different to just like l one or l two normalization. It’s basically trying to balance. These networks also have an excitatory inhibitory balance, which again, seems to be another principle that might have been missed in machine learning community. It’s well known amongst neuroscientists, and in the cortex, there is this EI balance, it’s called.

And it turns out it’s not there just for looks or for complex dynamics, right, which the cortex seems to generate. It’s even required in feed forward networks to maintain the spike amplitudes or the net input into each neuron. As information propagates through a multilayer network, you need this balance in order to actually maintain, more or less, call it dynamics, the feed forward dynamics in a functional range. Right. And that really hasn’t been recognized in the past. So what I’m doing now is I’m incorporating all of these principles into these networks. And right now, just doing it on mnist. Right. I know mnist is a solved problem. A deep network can get something like, what’s the record now? 99.9% correct or something. I don’t.

Basically, it’s better than humans do. And again you go, how is that even possible? These are handwritten digits, and now we’ve got ais that can read them better than people. So it’s solved in terms of gradient descent and convolutional neural networks. A simple convolutional CNN can solve mist, or at least a battery of them, what they call the forest, or more or less of CNN can solve it very well.

But in spiking networks, it’s definitely not a solved problem. So we’re trying mnist more or less as a checkpoint for ourselves. And on MNIST, we’re getting better than any other non convolutional architecture spiking network has ever achieved on MNIST. So that’s progress as far as spiking networks are concerned.

The next step would be probably to try it on a bigger image type problem. Because even people who I’ve personally spoken to, people who’ve been in spiking networks, and they say, yeah, they work for MNIST, that’s fine, but they don’t work for bigger problems. They literally just fail. So that’s tomorrow for me. Tomorrow is okay, let’s chuck Imagenet at a spiking net and see how it performs. And can we get it close to gradient descent for a similar network architecture?

That’s pretty much my goal is to look at the performance of gradient descent and get spiking networks to approach. They’ll never be as good because there’s no error propagation, there’s no indication to the network of exactly what the output should be. Right. But the brain gets away with that.

Getting into the idea, is there gradient descent happening anywhere in the brain or something like backpropagation? There probably is in specific regions and for specific tasks. But in general, the cortex learns through self organization. I think that’s reasonably well accepted. There’ll be people who disagree, but they’ll be in the minority, I think. So that would be the next task for spiking networks is just throw more complex problems at it that are already solved in the gradient descent field. But the more complex a problem we can solve, the more people will start taking notice of spiking networks, I think.

[01:14:06] Paul: Well, okay, so I was going to ask you if you feel bound to attack these benchmarks that have been set for what deep learning models have been successful at, because that’s not robotics and it’s not what brains are for.

[01:14:24] Peter: Right.

[01:14:24] Paul: So it might be different kinds of. And this goes back to me asking you about the functional psychological functions and cognitive functions that are more related to what humans do and whether that might be, if that’s a step too far to start just asking about a different avenue of cognitive function rather than tackling these things that have already been done. But you just said you do need people to pay attention to it, and people will pay more attention to it when they start tackling these same benchmark tests.

[01:14:53] Peter: I suppose that’s exactly right, for one thing.

Well, there’s three things. Yes, people will pay attention, right? And these are known benchmarks. Right? And that’s the second thing. These are known benchmarks. So it’s easy to. You’re basically comparing apples with apples, right? But you’re totally correct. The whole point and where spiking networks and more biological AI really comes into its own, is when you do start looking at interactions with the body and with the environment, because that’s where the dynamics of a spiking network actually come to the fore.

And I mentioned this briefly in the paper. As you say, it’s very dense, and I don’t expand on a lot of the ideas. Maybe that’s another paper down the track, and that’ll be the 100 page paper rather than the six page paper. But the dynamics of spiking networks are always transient, right? There’s no stable state in the brain because a spike happens and then it’s gone. It is actually just a moment in time. It’s a unitary event. And so the way that this processing actually unfolds is completely different to an artificial neural network. And when you think about it, our bodies, we survive in time, in the world we perceive in time. Everything is trans. It’s all dynamics, right? And so this is where spiking networks will really, I think, become particularly useful. And this is where gradient descent and artificial neural networks may be starting to hit a wall. I’m low to actually proclaim that, definitely. But in terms of interacting with the real world, the complexity of the real world, gathering enough data to train a gradient descent model in the real world, these are all quite prohibitive at the moment. Whereas the dynamics of a spiking network and the rapid local learning that you get using rules like spike timing, dependent plasticity, they seem to be perfectly cut out for doing that job.

So, yes, it’s actually a question that I kind of grapple with. Should we just keep trying to go to the next benchmark and at least get close to matching what gradient descent is capable of doing on each of these benchmarks? Or should we just dive right into the robotics side of things and show what spiking networks are potentially capable of in the real world?

[01:17:23] Paul: There’s a risk, right, that if you follow the benchmarks that you’re actually going to, it’ll lead you to the incorrect solutions, because you’re going to be tackling problems that you won’t eventually be tackling when you’re in an embodied system, for example.

[01:17:36] Peter: That’s entirely true, yeah.

And that’s one of the reasons that I took this job in a robotics institute, right? Even though I’m not actually working with the robots myself, I’m surrounded by them all the time. So I keep them basically front of my mind. And every time a robot trumbles past my desk, it’s like, okay, what’s it doing? And how could I get a spiking network to do that? And I’ve got this idea kind of like ticking over in the back of my mind all the time.

But it is a leap. It’s a leap, and it’s a leap into the unknown. And it’s probably a bit too ambitious and a bit too much of a chasm at the moment to really bridge. I think I need to understand, we all need to understand the dynamics and the capabilities of spiking networks a little bit better before we make that leap. And so really, the best way to do that is just on current benchmarks. I guess as trivial and mundane as it may seem, every time I try a spiking network on a new data set, I learn something new.

[01:18:36] Paul: Right.

[01:18:37] Peter: I think that’s the point. When I stop learning anything new by applying it to these canned data sets and standard benchmarks, that’s when I’ll stop doing it.

[01:18:48] Paul: All right, very good. So I’m going to zoom us out and ask you kind of a larger question, and then I actually have a guest question from the person who sent me your paper in the first place before we wrap up here. So the zoom out. So this will be like the end of the public version, and then the rest will be, like a little bit extra for Patreon supporters.

So what we’ll end on here in the public version is you don’t write the acronym AGI, or the phrase artificial general intelligence does not appear in that paper. But you did mention it earlier. And part of the context that you mentioned it is that your kid has asked you, when are you going to figure out AGI? Or is that what you’re working on? I don’t remember what the exact question was, but you don’t mention it. But looking at a list of the principles that we’ve been talking about, and this also has to do with me asking you what your goal was, because I was going to guess maybe it’s AGI. So here’s the question. Do you think that building these principles in is going to lead to AGI? Do you think that we’re on our way to AGI, or do you even acknowledge AGi as a thing? Because I don’t really, depending on how you define it. But where are we in terms of that, if you think that we’re headed that.

[01:20:11] Peter: Right, right. So, I mean, that is the rub, isn’t it? How do you define AGI? It’s really difficult to even define just intelligence in terms of, well, what is intelligence for us, or what is intelligence for biology and intelligence in the environment? I mean, it can mean so many different things.

So that’s one of the reasons I left it out of the paper, I guess, because it is just such a loaded.

[01:20:34] Paul: That was intentional. Was it intentional that you didn’t.

[01:20:37] Peter: I think it was intentional.

[01:20:38] Paul: Maybe it’s so dense you couldn’t even.

[01:20:39] Peter: Fit it in those three letters.

I think it was probably at the back of my mind. I just thought, well, it’s a loaded term. And in some respects, it’s a little bit. Unless you’re specifically talking about that, it’s a little bit unscientific to bring it up, even, because it’s just so open to interpretation. So just for the sake of, I guess, trying to write the most rigorous paper that I could on this topic, I just didn’t mention it. But, of course, I think everyone who works in AI has that in mind. Right? Whether it’s their goal or not, they can kind of see that the field is probably heading in that direction.

So what do we mean, though? And this is the problem, what do we mean by Agi? Now, for me, before the idea of this embodied Turing test, which is really looking at animals up to humans, but starting with animals, there was this idea of, can we build a robot that can just walk into an unknown house, find the kitchen, make a cup of coffee for you?

A person could do that, right? Pretty much any person, at least in the western world, can walk into almost any house in the world. They know what a kitchen is. They know what taps are. They know where to get water. They know that the kettle is going to be on the shelf or in a cupboard somewhere. If they can’t see it, they’ll search for it. They’ll find the coffee in a cupboard somewhere, and they’ll be able to make a cup of coffee in a house you’ve never seen before. So this was a test that was proposed, I think, decades ago. Can we build a robot that can actually do this?

And like I said, that was kind of like a precursor to the embodied Turing test. Now, would you call a robot that could do that, would you call that an Agi?

I guess that’s just arguing, really, semantics.

We definitely don’t have a robot that can do that right now. And if we could build a robot that could do that, it would be incredibly impressive. But more to the point, it’d be incredibly helpful. Right.

[01:22:43] Paul: Especially because I’m too lazy and depressed to go make myself a cup of coffee. No, I’m just kidding.

[01:22:48] Peter: Right. So just in terms of if it can do that, it can do pretty much any manual labor that we want it to, and that’s going to be incredibly useful and game changing for society. And that’s what Elon Musk is trying to do with the Tesla bot optimus, or whatever it’s called.

He’s trying to change the face of labor globally.

Of course. I think that’s very misguided, because it’s not just a matter of building the hardware. The real problem is, the real issue is the brain. It’s the AI that controls the robot, and that’s a long way off. We can build robots that are physically capable, sure, but not mentally capable of actually carrying out the computation required to do these things.

So coming back to your AGI question, is that AGI, would you call that AGI for robot if we build a robot that can do those things?

I don’t know.

[01:23:43] Paul: Semantics.

[01:23:44] Peter: Semantics, exactly. So in terms of the bullshit term.

[01:23:50] Paul: I’ve come to hate the term artificial intelligence. And was it John McCarthy, one of the earliest AI folks in the Dartmouth summer conference, when they were going to figure out mean, I think it was John McCarthy who coined the phrase AI. I’m artificial intelligence and I’m butchering the history. But one of those guys, one of those researchers, really hated the term, and I really have come to dislike it myself.

[01:24:17] Peter: Yeah, I think it’s a loaded term and it’s probably overused. Do you have any other reasons for disliking it?

[01:24:26] Paul: Because it reifies intelligence. It sort of equates intelligence between biological organisms and not biological organisms, as if they’re the same thing. And like we just said, there’s all sorts of intelligences across species, different types of intelligences. So to say it’s an artificial intelligence, as if it’s one thing, is strange for me to conceptualize at this point.

[01:24:51] Peter: So I guess what you’re saying, it comes down in part to just our inability to even define what intelligence is, right? Let alone artificial intelligence.

[01:24:58] Paul: Yeah, I mean, we have to operationalize everything, which is fine.

[01:25:02] Peter: Okay. No, I get that.

We’re certainly heading towards more capable models, let’s put it that way.

[01:25:12] Paul: Well said.

[01:25:14] Peter: Whatever you want to call it, we’re getting towards more capable models. At some point they’re actually going to be useful, and that’s when it’s really going to make a difference.

Now, how do we make these things useful? You can argue that, I guess, Chat GPT is already useful because people are actually using it.

Kids are using it for their assignment, their school assignments. But more to the point, people are using it in workplaces to write emails and, I don’t know, draft up ideas, brainstorming. They’re used in all sorts of different fields, right? Really diverse fields.

Architects are using generative AI to brainstorm new ideas for buildings. So it’s becoming useful already, but still, it’s only in the digital. Useful in the digital domain.

So when it becomes useful in the real world, I guess I’m talking robotics, that’s really going to be even more game changing, I would say.

[01:26:11] Paul: Okay, so we’re winding down, and I have one last question for you, and then I promise I’ll let you go. And that is, do you see a role for consciousness in any of this? Is consciousness important? Is it a byproduct? What are your thoughts on that? And if we build in all of. If we integrate all these things, is consciousness going to emerge? If it’s a dynamical enough system, et cetera.

[01:26:39] Peter: Right.

Well, if you think AI or AGI is a loaded term, well, then consciousness is definitely.

[01:26:46] Paul: Sure, yeah. Okay.

[01:26:47] Peter: Awareness.

[01:26:48] Paul: I don’t know, what do we call it? I’m not sure what to call it.

[01:26:51] Peter: No, I just mean anything that’s representative of that concept.

[01:26:55] Paul: Look.

[01:26:58] Peter: Yeah, it’s very hard to kind of sum up.

I’ve had, like, hour long discussions with journalists and things about this and barely scratched the surface.

I don’t think there’s anything special about the human brain, right. Or any animal brain that leads to consciousness.

That’s almost like saying, if you believe that, then you kind of have to believe that there is like, I guess you’re separating the mind and matter, right? If you do that.

So can we emulate consciousness in a machine? I guess.

[01:27:40] Paul: Would it be useful to do so also?

[01:27:42] Peter: Well, would it be useful, but would it be avoidable? Or is there some level of intelligence where consciousness just appears?

So in other words, if we do build the loaded term agi or something that understands the world the way that we do, is it necessarily going to be conscious?

I tend to lean towards the answer being, yes, it will. It will be conscious simply because there’s just nothing special about the human brain. It is just a whole heap of biochemical processes going on, right? It’s chemistry and physics. That’s what the brain is. Somehow that leads to this innate subjective feeling of awareness that I have and that I presume you have, although I can never know.

But there’s nothing special about the brain. It is purely a physical thing that leads to a purely mental state of awareness. So I think if you build something that computes like the brain does, but you build it in silicon and it uses pure electrical spikes rather than electrochemical spikes that are used in the brain, will it be conscious?

I’d say there’s a good chance that it will, yes.

I’ve just got a warning about my battery. That’s okay. I think I’ve still got plenty.

[01:29:05] Paul: That’s all right.

[01:29:07] Peter: So I think yes. The answer is yes. Machines will be conscious at some point. When we do get to that point, exactly what that point is, I don’t think anyone knows.

[01:29:18] Paul: Yeah, all right, that’s a ridiculous and great place to end. So, Pete, thank you for staying along with me and going down many roads. And I’ll point people, obviously, to the paper that we most discussed, but some of the other work where you’ve started implementing some of these things as well in the show notes. So thanks for coming on.

[01:29:35] Peter: Thanks for having me. It’s been an absolute pleasure. Thanks, Paul.

[01:29:52] Paul: I alone produce brain inspired. If you value this podcast, consider supporting it through Patreon to access full versions of all the episodes and to join our discord community. Or if you want to learn more about the intersection of neuroscience and AI, consider signing up for my online course, Neuroai the quest to explain intelligence. Go to Braininspired Co to learn more. To get in touch with me, email Paul at Braininspired Co. You’re hearing music by the new year. Find them at the newyear. Net. Thank you. Thank you for your support. See you next time.