BI 245 Dan Levenstein: Neuro-AI, Dynamics, and Model Systems

[00:00:03] Speaker A: One of the things about the brain is that it's a very multi level system, right? Both in terms of like spatiotemporal levels, from cells all the way up to behavior and maybe even interactions between organisms. Right? That's like a huge range of spatiotemporal scales. And then also in terms of like, from physical properties to like computation and functions. Like, those are very different levels of abstraction, right? The difference between what the hell is a computation. Like, let's even take Mars levels, right? Computational algorithmic implementation. And I think if you want to think about the brain, you need to be thinking about all of these levels. So you asked earlier, in what ways is theory easier or harder than experiments? That's one of the ways that theory is harder than experiments is like that decision that in itself is the art of modeling, is how do you set up your experiment? Which is, what are the assumptions that I build in and what are the things that I want to emerge? I'm working on the spontaneous activity part. [00:01:18] Speaker B: I know. Yeah, but you think the spontaneous activity is where it's at. [00:01:24] Speaker A: I think it's important that when I am not quote unquote, doing a task, my brain keeps doing stuff. Let's say these things are conscious. They're only conscious of the text that they receive when they receive it, and the text that they produce when they produce it. And then the lights are out. [00:01:47] Speaker B: No, man, they've got world models, right? That's where this, like a term like that is. This is brain inspired. Powered by the transmitter. Daniel Levenstein started his neuroai and dynamics lab at Yale University about a year ago. We briefly discuss what that was like, starting a lab, transitioning from a postdoc to a principal investigator, that is the head of a lab. But mostly we discuss his work and ideas. So Dan studies spontaneous neural activity during sleep, especially in brain areas like hippocampus and cortex, and how this internally generated spontaneous activity is related to learning and memory and navigation, really. He used to study those processes directly through experimental brain recording data sets. These days, he builds and studies models of those processes using AI models and seeing how their dynamics and functions match what we see in brains. I also loved that Dan called me out during a talk I was giving at his university when I was invited to give a talk about some recent work I've been doing. So that takes about three seconds of the podcast, but stood out in my mind. So Dan believes that sleep is important for brain processing. I look like I do not believe sleep is important, but I hope you're well I'm well. I hope you're well. You can find links to his information and the papers that we talk about in the show notes at brain inspired co podcast 245. Here's Dan. The first thing I want to know. Everyone has their origin story of what got them interested in neuroscience. And is yours sleep? [00:04:06] Speaker A: No, mine is more, I guess, brain dynamics in general. [00:04:12] Speaker B: Wow. [00:04:13] Speaker A: Yeah. Yeah. So my path to neuroscience is a little convoluted in that I started my undergrad as a music major, decided that wasn't worth my time or money, and switched to biochemistry. [00:04:32] Speaker B: Now, why biochemistry? So I started off in aerospace engineering because it seemed like, hard and it would be a good career. And then I switched to molecular biology in undergrad because it seemed like it would be hard. No really better reason for than that, you know. Yeah. [00:04:49] Speaker A: You know, I don't really remember. It seemed interesting. It might have been that. It might have been that simple. I think I was looking for things to do and that seemed like an interesting one. [00:05:02] Speaker B: Sure. [00:05:04] Speaker A: And so then as part of that, I ended up taking a bunch of physics classes and ended up with a minor in biophysics, started a PhD in biophysics, decided like two years through that that I wanted to switch to neuroscience, and then left with a master's and applied to neuroscience programs. [00:05:24] Speaker B: Wow. Okay. Yeah. [00:05:26] Speaker A: And like, basically what happened during those periods of time is I was reading a lot of, you know, neuroscience books, like Olaf Sporin's books about networks, Christoph Koch's books about consciousness, you know, the things that get people interested in the brain. Sure. And then I read Isakiewicz, and so I was doing some dynamical systems and was like, this is great. I want to use dynamical systems to study the brain. [00:05:59] Speaker B: Okay, well, when did your interest in sleep come about then? Because you do a lot of work in sleep. [00:06:06] Speaker A: Yeah. So when I started my rotation in Yuri's lab, there was. Yeah, just one of Jockey's lab. There was a postdoc, Brendan Watson, who had collected a big sleep data set of various frontal regions and was trying to figure out what to do with it. As one often does in the Boujacki lab, they collect a dataset and then figure out what to do with it. And Yuri said, why don't you go do your rotation with Brandon? And so that kind of got me hooked, really. I mean, you know, between seeing the interesting dynamics that kind of come about during sleep, you know, there's a whole bunch of internally organized crazy patterns that happen that are very sleep specific and then trying to connect that with the function which is not a trivial thing to do. [00:07:01] Speaker B: Well, was like complementary learning systems a big thing at that point? Like with and, you know, sharp wave ripples? Was that already a big story? [00:07:11] Speaker A: Yeah, absolutely. You know, I didn't end up doing too much with sharp wave ripples till like later in my Ph.D. it kind of like fell out of the model that I was working with. But, yeah, complementary learning systems was certainly a motivation. [00:07:26] Speaker B: Yeah. [00:07:26] Speaker A: And I think that was one of the papers that Yuri was basically like, all right, you have to read this paper if you want to do anything with sleep. [00:07:33] Speaker B: Oh, that's cool. That's cool. Just for the listeners. That's the idea essentially, that, like, you get these fast encoded memories in the hippocampus, and then the hippocampus, like sends the memories to the cortex to sort of put it in longer term memory and generalize, et cetera. That's a really poor man's version of the complementary learning systems story. [00:07:55] Speaker A: That's it in a nutshell. That's. That's the story we tell. [00:07:58] Speaker B: Yeah. Is it still the story we tell? Because I say that and then I think, oh, I'm sure that's been updated quite a bit. [00:08:05] Speaker A: It's the story everyone tells. And then everyone basically says, yes, but. And people have different yes buts that they tell. [00:08:14] Speaker B: Well, so we're going to talk about some of your work in that field in sleep and with maybe sharp wave ripples and the function of the hippocampus and the cortex. But your lab. Well, you know, let me back up. How long have you been now working at Yale? [00:08:29] Speaker A: Has it been one year? As of two days from now? [00:08:34] Speaker B: Okay. Yeah. [00:08:36] Speaker A: August 1st last year. [00:08:37] Speaker B: August 1st last year. Okay. When I visited Yale, I guess it was. I mean, it was a few months ago. So you were just a few months in. It looks like you have a couple more books on your desk in the back than you did at the time. [00:08:48] Speaker A: But no more art. [00:08:49] Speaker B: No more art. Okay. And then. I'm not sure if you want me to say this, but congratulations also on a recent wedding. Is that okay to say? [00:08:58] Speaker A: Yeah, absolutely. That's not. [00:09:00] Speaker B: Things are rolling for you. Things are just rolling forward. Something bad's gonna happen, right? [00:09:05] Speaker A: Oh, I hope not. Especially since I'm trying to write so many grants right now. [00:09:11] Speaker B: Is that what's happening right now? A lot of grant writing? [00:09:12] Speaker A: Yeah, absolutely. [00:09:13] Speaker B: Yeah. Okay, well, but so you're. The tagline of your lab is that you're a neuro AI and dynamics lab. So I know your heart is still very much in Dynamics. But you're also all in on the, on the neuro AI. So can you explain a little bit maybe just how neuro AI comes in to what you do? [00:09:35] Speaker A: Yeah, so I think after. And I guess this is good because it continues the story I started before. After my PhD, I was really interested in brain dynamics, but I felt like with just like dynamical systems, mean field type models, it's very hard to get at computation. [00:09:58] Speaker B: Right. [00:09:58] Speaker A: The these models can produce super interesting dynamics and can do absolutely nothing useful. And so the question was kind of if I want to study how brain circuits do anything useful, how do I do that? And it just seemed like artificial neural nets were really the only way to go, you know, because they can, they can compute. [00:10:23] Speaker B: So. Okay, now I immediately want to go off on a tangent about computation. I've been talking about computation a lot because, you know, whether it's dynamics or computation. So but I will leave that alone. But they can do useful things. How about that? [00:10:37] Speaker A: They can do useful things. They can do adaptive behavior that we can quantify their performance at, quantify how their performance improves and then study how it's done. [00:10:50] Speaker B: And so are you using neuro AI, like deep learning networks to study sleep as well? [00:10:56] Speaker A: Yeah. So that is one of the big things we're hoping to do and we're starting to do. [00:11:03] Speaker B: Well, you sent me a lot of things that you're sort of beginning to do. [00:11:08] Speaker A: One of the fun things about starting a lab is you kind of come in with an idea and then people show up and they're excited about. About certain things. And so you do the things they're excited about. [00:11:19] Speaker B: Also it goes in different directions. I do want to talk about what it's like becoming a PI. Maybe we'll save that sort of for the end. Ish. But so I don't know, maybe we can start off with just how you see neuro AI broadly. First of all, do you like that term? Because it's still kind of a new term and I'm not. Did Tony Zador, I don't remember who coined the term, but yeah, I don't know. [00:11:44] Speaker A: I don't even know if, if Blake likes it, but we use it. [00:11:49] Speaker B: Oh yeah. Sort of like the term artificial intelligence. [00:11:52] Speaker A: Yeah, yeah, exactly. Like, I feel like you can either be in the game of like trying to police terms and come up with terms or you can spend your time doing things. And so, you know, I'm fine with the term neuro AI. It captures that it's uses some stuff from AI and some stuff from Neuroscience, and none of us really agree on exactly what it means, but we all draw that stupid loop from the brain to AI and from AI back to the brain. And we say this is neuro AI. [00:12:21] Speaker B: Yeah, it's mandatory. The loop is mandatory. But I mean, there's something to it as well, right? I mean, that's. We wouldn't be using it if we didn't really believe in the loop. [00:12:29] Speaker A: Yeah, yeah, absolutely. And I think kind of what it comes down to is if you're interested in how the brain does intelligent things, then the tools from AI are very useful for. For studying that. And, you know, I'll say idealizing that process. I'm not going to say that the brain is doing the same thing as modern AI systems and maybe even less so as they get more modern, but they're a good kind of model organism almost for mirroring what the brain does or being a standard. [00:13:08] Speaker B: But do you think that when you, okay, you train a model to perform a task and then you look at the model's dynamics, slash computations, do you think then you've solved it or like, how far do you think that gets us in terms of understanding? [00:13:24] Speaker A: Understanding what? Understanding how the brain does it or understanding how that system does it? [00:13:29] Speaker B: Understanding how the brain does it. That's a good clarification. Yeah, [00:13:36] Speaker A: I think that, Lee, you know, that's an open question, and you need to go back to the brain and look at how things compare. Maybe you compare how the representations work. Maybe you know how the representations look in their structure. Maybe you are able to do perturbations of the brain, either, you know, of the brain itself through optogenetics of the stimuli or of the behavior, and see how that matches. I think understanding how a neural network does something on its own doesn't tell you how the brain does it, but it does tell you how a model system that you've set up with hopefully some properties that you've designed based on the brain does it. And now at least you have a working system. [00:14:17] Speaker B: Right, Right. [00:14:19] Speaker A: I mean, kind of the same way that when we understand how a mouse model of Alzheimer's does whatever it does and has disrupted memory, that doesn't mean we know how memory is disrupted in patients with Alzheimer's. It means we have a model system that we have better access to. [00:14:41] Speaker B: Dan, Dan, you're sharing too much with the public here. [00:14:44] Speaker A: Right? [00:14:45] Speaker B: Yes, that does mean we understand Alzheimer's. I mean, have you then, like, because you made that clarification of whether I meant understanding the brain or understanding the Models. But is understanding the models something you're also interested in? [00:14:59] Speaker A: Yes, because I think they're interesting systems, but I wouldn't say that that's my primary motivation. Right. Like at the end of the day, I'm interested in the brain as a physical system and how it interacts with the world to produce adaptive behavior. And I think that artificial neural networks are a good model organism for that, for certain aspects of that. [00:15:24] Speaker B: Okay. [00:15:25] Speaker A: Yeah, I would say that they are the best for the problems that I'm interested in, but not necessarily because any aspect of them is really good. Right. So, but because they kind of span the whole gamut from. Right, right. So one of the things about the brain is that it's a very multi level system. Right. Both in terms of like spatiotemporal levels from cells all the way up to behavior and maybe even interactions between organisms. Right. That's like a huge range of spatiotemporal scales. And then also in terms of like, from physical properties to like computation and functions. Like those are very different levels of abstraction. Right. The difference between what the hell is a computation. Let's even take Mars levels, computational, algorithmic implementation. And I think if you want to think about the brain, you need to be thinking about all of these levels. And neural networks are the only model in town that interface quantitatively with each of those levels. And their interface with each of them is kind of cartoonish. It's very much a spherical cow. Right. Neurons and neural networks are a spherical cow of neurons in the brain layers in neural networks are spherical cows of brain regions. The way they connect are spherical cows of the interaction between, say, cortical regions or cortex, hippocampus, whatever. But the fact that it has all of them is what makes them so interesting to me. [00:17:09] Speaker B: And so do you think moving forward that they're just going to continue to be the main. Well, I was going to say the main modeling tool. I don't know if they're the main modeling tool across neuroscience yet, but it sure seems like it. [00:17:28] Speaker A: I mean, I hope not. Science goes well when there's different people doing different things. Certainly I see them being a big player. The fact that you can learn stuff from studying the brain to improve neural networks and that can lead to more applications than just, you know, it really kind of expands the space of potential applications for our research, which even if we're not doing directly, I think is good for the field, both because it results in an influx of interest, influx of funding, you know, whatever it is. And at the end of the Day, I think the time when the stuff that we do gets tested is when we have to put our money where our mouth is and make something that works. And even if we're not doing that ourselves, the field has to do that. [00:18:22] Speaker B: I forget, are you collecting data these days or is it all. It's purely modeling in theory. [00:18:27] Speaker A: Yeah. I haven't collected any data since my master's. That was one of the reasons that I, that I left my biophysics program is that I was doing like single ion channel patch clamp, electrophysiology. [00:18:40] Speaker B: Yeah, that's hard. [00:18:41] Speaker A: And that. Yeah, no, that's awful. And you know that. And like cell culture, protein structure function stuff and the nature of doing experiments is not something I love. [00:18:59] Speaker B: It's not for everyone, but it is for some particular kinds of people. But I was always. So I'm still doing experimental stuff. I'm doing a lot more computational stuff and modeling stuff now, but I still am on the sort of experimental side. In the lab I'm in, we still collect data. And I've always been kind of envious of people who just go pure modeling because somehow it seems easier. And some modelers say, yes, it is easier, and some say, no, it's not. It's just a different set of problems. How would you characterize it? [00:19:32] Speaker A: I think some aspects of it are easier and some aspects of it are harder. That's kind of a wishy washy answer. Right? But when you're collecting data, the fact that you've collected a data set, which hopefully you've planned to be unique in some way, means that just the fact that you've collected that data set, you have a publishable entity you can spend. I mean, you have to do stuff with it. You can't just say, here's my data set. But that is work that you've put in that has, like, produced something so you don't get scooped, whatever it is. But you can spend a whole bunch of time doing theory and just come out with nothing or like, feel like you've done nothing. [00:20:14] Speaker B: Okay, well, how do you. How do you. [00:20:16] Speaker A: Sometimes it's almost unclear as to like, oh, what it is we're actually trying to do until you've done it and then you're like, oh, that's the problem I was trying to solve this whole time. [00:20:29] Speaker B: Oh, that's interesting. Do you have an example of something like that that has come out of the lab? [00:20:36] Speaker A: Well, I mean, this was also a data analysis paper, but the inner spike intervals paper, that was the end of my PhD and now six years later is finally really published. [00:20:49] Speaker B: There was a good ground state stuff. [00:20:51] Speaker A: This is the ground state paper. Yeah. [00:20:54] Speaker B: We have to describe what that is though. [00:20:56] Speaker A: Yeah. I mean, there was a period of time, a long period of time where I had no idea what that project was about. [00:21:03] Speaker B: Oh, oh, that's so interesting. Yeah. [00:21:06] Speaker A: That project did not start from like a question or, you know, a specific problem. That project really started like from an observation that I was doing in some data analysis that Yuri was basically like, huh, that's funny. What was the observation? Figure out what? [00:21:25] Speaker B: Yeah, so was that or. Yeah, I'm sorry, sorry. Because I know this is like six years now and you're like, oh my God, I just want to be done with this thing. Right? [00:21:33] Speaker A: Yeah. No, I mean, that's the funny thing about papers that take a long time at the end is you're just like, it's your past thinking about ancient history and you're just so over it. So the observation was. No, no, no, no, no, it's fine. We should. Because it would suck if that paper, you know, disappeared into the dustbin of things that happened in the past even. [00:21:56] Speaker B: I mean, we don't have to go in depth about the actual findings, you know, but maybe a broad overview. But I'm actually more interested in the process of how you got there. Right. Because it feels if it came as a sort of theoryless, ish development. [00:22:12] Speaker A: So the observation, we were looking at interspike interval distributions, which are the time periods between spikes, and looking at how they vary. [00:22:21] Speaker B: Because that's what one does. [00:22:23] Speaker A: Because that's what one does when you have a data set. I don't remember why it was a plot that I made. Because you make so many plots that you never show anyone. That's right. And so the plot was just. Oh, because you're interested in firing rates. Yuri has this long standing obsession with why neurons have different firing rates and why the shape of the distribution of their average firing rates is log normal. And you know, if you want to know more about that, you can read his multiple reviews on, on that. [00:22:54] Speaker B: He is the log normal guy. [00:22:56] Speaker A: Yeah. And so we're looking at interspike interval distributions. We're looking at the distribution of log inner spike intervals, which is nice because you can see the different time scales. Right. Usually when you look at inter spike intervals, you have to pick basically your bounds. Right. Because that tail is really, really long, multiple seconds. And if no matter where you look, it looks kind of exponential, maybe with some funny shape and obviously no very short inner spike intervals. Because you have the refractory period. And so what we found is, or the observation was basically that when we look at the log inner spike intervals, you see basically discrete bumps of density at different characteristic timescale at 100 milliseconds, 10 milliseconds, 1 second, whatever. And that when you sort the distributions by the cell's firing rates, the shorter timescale modes, the shorter timescale bumps in the distribution are basically the same across all of the neurons in the same recording, which happened to be in the same brain region. But the long inner spike interval modes basically followed the cell's mean firing rate. And so that was the observation. And years later, we, you know, over years of kind of poking at this, we, we came up with a story to kind of wrap around that. [00:24:29] Speaker B: And do you think so now, looking back, sort of reflecting on that, just as an approach to science, is that worth doing? As an approach? Right. Just toying around. I mean, I love exploratory science because it leads to things like that, but boy, it takes a long time and it's frustrating. [00:24:48] Speaker A: So I think it's worth doing either if you're old and your lab is established and you have enough people, or you're in your PhD and you have time to kind of mess around. And it's not your only project. [00:25:04] Speaker B: Right. Okay. [00:25:06] Speaker A: I mean, I, I think, you know, the Bujaki lab, this is one of the main things that they do. It's like Yuri basically says, you can't stick a wire in the brain without finding something. [00:25:17] Speaker B: Finding something. Well, that's the inside out approach that he advocates. Right. [00:25:21] Speaker A: But like, if you're starting your lab, it's a terrible idea [00:25:27] Speaker B: because you know [00:25:29] Speaker A: what you're going to find or how long it's going to take you or how important anyone is going to find it. [00:25:36] Speaker B: So what, what is your alternative approach now? How are you starting your lab now? [00:25:41] Speaker A: Yeah, good question. I mean, I guess the, the baseline approach is when, and you know, I'm still figuring that out, but when people come to talk to me about wanting to be in the lab, I try to kind of feel out what they're interested in, try to assess how it relates to kind of the weird grab bag of interesting ideas I've been playing around with, you know, at varying degrees of well formed and try to throw spaghetti at the wall and see what sticks. You know, if someone gets excited about an idea, then they're going to roll with the project and it's going to do it. They're going to do what they want. And then your job is just to kind of channel them, suggest things, etc. [00:26:39] Speaker B: But is it ever the case that the idea that we go into something with is the idea that is eventually what we're working on? [00:26:49] Speaker A: Very rarely. Usually when you start something, you, you know, you find something and then you switch over to something else and the project goes where it goes, I think. And again, still figuring out how the hell to do this job. Right. But sure, I think part of the job of the PI is to do that kind of direction and figure out how that works best with each student or, you know, each, each trainee. And you're, you're trying to also figure out how to channel those things into what the lab is good at, what you know, you're comfortable with, what your skills are, what the complimentary skills are in the lab, and like the subjects of interest and balance that with the things that make them exciting or make them excited. [00:27:47] Speaker B: All right, we're just going to go ahead and talk about this. Is it a hard transition to go from? I mean, are you doing science still or are you a manager? [00:27:57] Speaker A: Yes. So it's a huge transition and everyone kind of does it differently. Right. I know some young PIs who are still in the lab, still very hands on, doing experiments. I got some advice before I started, which is don't be a super postdoc. I've also heard the complete opposite advice, which is you have to be a super postdoc until basically you have things up and running. Right. So personally, I feel like, well, you wear a bunch of hats, but in terms of the science, like that's getting done through meeting with people, talking about what they're doing, suggesting new things, and I guess it's almost kind of done vicariously. So in that sense, you are more of a manager, but you're doing the science with them. You're just not the one coding. [00:29:02] Speaker B: You have the high level vision and constraints sort of that help channel the science. [00:29:08] Speaker A: Absolutely. And that comes in a lot of ways. It comes in through how you choose to direct your lab meetings. It comes in through what papers you put in the lab slack. It comes in through what you suggest people do, you know, for next week when you meet with them. Yeah. And then. Yeah, no, you go ahead and on side of it, you need to write grants as if you have, like, very clear things you're doing. And then, you know, see what happens with that. [00:29:47] Speaker B: Yeah. Okay. [00:29:48] Speaker A: For the record, this might be a terrible way to run a lab. [00:29:52] Speaker B: Well, that's the problem is, you know, it's I think it's called the Peter Principle, where people get, you get, what is it called when you get, you [00:30:01] Speaker A: go above the job you've been doing and trained for and good at, and [00:30:05] Speaker B: now you're in it and you end up as a PI. Right. And I guess that some people wear that differently than others. So, anyway, well, I hope, you know, I, I, I wish you the best [00:30:14] Speaker A: time, at least for me. When I was a postdoc, I was kind of gradually like, I, I was still working on my main project, but I was also mentoring like a number of graduate students and undergrads. So, you know, I don't know if that was, you know, Blake and Adrian doing that on purpose, but it certainly felt like it was a soft transition, at least on the how to do science through mentorship, basically. [00:30:52] Speaker B: And again, I don't want to put you on the spot because I know that you're toying with a lot of different ideas and being a theory driven, modeling driven lab, especially these days, you really could go in lots of different directions. But is there something in particular that you are super interested in solving or if you picture yourself 30 years from now, that's something that you'd be really proud of that you did some great work on? [00:31:21] Speaker A: Sure. One of my main interests is really how internally generated activity supports memory and navigation. I think that's pretty broad. And when I say an internally generated activity, spontaneous activity mostly during sleep, but also during rest, whatever. And so something that I would be very happy to feel like I had, I don't want to say solved, but, you know, had made a contribution to and had come out of my lab, was a system that I felt didn't, When I say a system, probably some series of artificial neural networks that do that, similarly to how the hippocampus and cortex do that. Oh, okay. And where I can say quantitatively this does it similarly because it produces data that looks like the hippocampus and the cortex, I manipulated in ways that I can or my collaborators can manipulate the hippocampus and produce predictable changes in neural activity and behavior. And I think the connection there between the computations that the neural activity is doing and the actual behavioral observations, if I could make a system that could do that in virtual environments, let's say that would be, I would be quite happy. [00:32:51] Speaker B: Well, you've already done a lot of work toward knowing why brains need sleep. Why do brains need sleep? What do we know about, what do we know and what do we not know about that internally generated activity, you know, during periods of Rest and sleep. [00:33:08] Speaker A: Yeah. So I think there's not one clear answer to that. Partly because you have different people that think different things, and partly because the brain probably uses sleep for a bunch of different things. I think this question of, like, what is the function of sleep is a. Is a terrible question because that's like saying, what's the function of wake? Well, you know, it's to get food, get away from predators, find. Find them, you know, like all the stuff that you do when you're awake. [00:33:38] Speaker B: But can you say of a brain area, then can you say what's the function of the hippocampus? Sorry, that's to derail. [00:33:47] Speaker A: I don't know if that's a. That's a dangerous pit of snakes. I mean, I would say no. I would say the. The hippocampus. In fact. No, I would say no, you can't. I would say the hippocampus has many functions. Right. And probably at different levels of abstraction, too. [00:34:04] Speaker B: Right. [00:34:04] Speaker A: Like, the hippocampus is involved in memory. Part of its function is to support memory. I wouldn't say the hippocampus does memory. I would say the hippocampus does things that support the function of memory. Hippocampus does things that support the function of navigation. [00:34:19] Speaker B: You're tiptoeing around this really well. [00:34:21] Speaker A: It's a delicate thing to do. [00:34:23] Speaker B: No, no. It's a difficult problem to talk the function of something because you don't want to pin your, you know, oh, that's the function of the thing, you know, and. Yeah, I like how you said it. I wouldn't say it does memory, but for short shorthand, you could say it does memory. [00:34:40] Speaker A: Yeah. So for sleep, I think there are really. To actually answer your question. [00:34:44] Speaker B: Sure. [00:34:44] Speaker A: Not to tiptoe around it. There. There are really three main. Different theories, but I don't want to say competing because I think they're probably all happening. But one of them I would call policy learning, and this really comes from reinforcement learning. And they use replay buffers to learn policy functions. And they need to. Because it's very hard to get enough real experience to do reinforcement learning. So you need to supplement your real experience with fake experience. And so you use replay for that. And so I would say that's like one bag of ideas of what's happening during sleep is you're using the hippocampus to generate fake experience so your action and decision making networks can learn how to act. [00:35:34] Speaker B: So you're learning during sleep. [00:35:36] Speaker A: Yep, absolutely. I would say the other bag of theories Is memory consolidation, schema, learning, whatever you want to call it. That's the two stage model that what the hippocampus is doing is storing a bag of memories that are then used to train the cortex. Because the hippocampus is a slower hippocampus is a fast learner, the cortex is a slow learner. And if you want to learn the statistical structure of the world, you need to learn from many memories that you've experienced at maybe even different times and kind of extract higher level schema, we can call them, from those individual experiences. And that's the memory consolidation view of sleep. The third view is what I would call the homeostasis view, which is that Giulio Tononi has done a lot of this, but there are other homeostatic theories and basically that activity during sleep is regulating something about neural circuits such that they can get reset so you can perform well later. So maybe this is synaptic scaling. You know, the, the, the synaptic scaling hypothesis is that synapses are downscaled basically globally during sleep. And then learning selectively upscales them. But you need to kind of combat that with downscaling. But there are, there are kind of other flavors of homeostatic theories which all involve some form of, you know, some form of healthy state that sleep needs to bring you back to so you can do okay and wake. Right. So one of them, another one is about criticality. This is a. Which I think we talked about when you visited. Right. So this is. [00:37:22] Speaker B: Yeah, yeah. You're the jerk in the audience who, who said, like, what, what is criticality really? [00:37:26] Speaker A: It's critical now what? Yeah, yeah, sorry about that. I mean, I ask it because it's a, it's, it's the question I have when I think about criticality, which I think is interesting. Right. But it's always all right, it's critical. I get it. Criticality is good now what. But I think this is a great theory of what sleep is doing and it seems like there's good evidence for it that sleep returns neural circuits to a critical state and that during wake they kind of diverge from this critical state and they return to it during sleep. And some sort of internal processes, you know, the interaction of dynamics and plasticity contribute to that in some way. [00:38:06] Speaker B: We should just say that criticality is this idea that it's not too ordered and it's not too chaotic, it's not too active, it's not below being too active. It's just at this just right level of activities that maximizes computational Capacity and information processing, et cetera. [00:38:25] Speaker A: How many times have you had to do that to say that, Explain to the audience what criticality is? [00:38:31] Speaker B: Did it sound pretty rote? [00:38:33] Speaker A: No, no, it's just it. It's a theme you had a bunch of people on recently or over the past, I guess, what year or two? [00:38:42] Speaker B: Yeah, well, there's just something. [00:38:43] Speaker A: Criticality. [00:38:44] Speaker B: Yeah, well, I'm working on criticality in my own research as well, so it's of high interest to me. But hopefully it won't take me six years to get something out, but it sure feels like it is. [00:38:57] Speaker A: Yeah. So, I mean, when we were talking about things that got me interested in the brain, criticality was one of them. [00:39:03] Speaker B: Oh, really? Yeah, yeah. [00:39:05] Speaker A: You know, it's one of these interesting things where when you come from biophysics, it's like, oh, that's an interesting thing. Looks like the brain is doing that. Cool, let's go study that. [00:39:15] Speaker B: Well, one. This is a total sideshow now, but one interesting thing I find about it is that I have a different story of how I would imagine results would turn out under different experimental conditions than the. Than my advisor does. Right. We have sort of competing, like, well, if it's. It needs to be critical in this condition, and he thinks it needs to be critical in a different condition. And so now we have to analyze the data and find out and stuff. [00:39:45] Speaker A: But yeah, so it's good when you have two competing hypotheses and now you have to go analyze it to find out. [00:39:51] Speaker B: Yeah, yeah, but the hypotheses are always being developed. Okay, but so back to your, your interest in sleep and, and wanting to develop models. I mean, you shared with me that. That you're either working already on spiking neural network models that have both the homeostasis aspect and that do the learning. And are you like, trying to combine all three of these camps into one kind of model as we speak, or is that something in the future? [00:40:22] Speaker A: Yeah, so that that model doesn't sleep yet. So to back up, this is an example of a project that I had not planned to do, but a student showed up who was interested in some related stuff, and that's kind of where the project went. And actually it's turning out to be really interesting. And so it's actually in some ways a follow up to the, the Crown State paper we were talking about. [00:40:53] Speaker B: Well, is that, is that why you're doing spiking instead of like a firing rate model? Yeah. Why spiking? [00:40:59] Speaker A: Yeah. Why. Why did we end up working with spiking networks. [00:41:03] Speaker B: I mean, because that's how you could get the inter spike interval distributions that you. [00:41:07] Speaker A: Yeah, yeah. I don't know if that was the original motivation. That was a decision that was made in like our first one or two conversations and. And then we went with it. [00:41:23] Speaker B: But spiking networks are notoriously more difficult to train. And [00:41:28] Speaker A: so I think they're getting easier and easier. And actually this, I think Dan Goodman posted this on Blue sky like a few weeks ago and there was a discussion about why it's a great time to work with spiking networks. And it's because the methods to train them are actually pretty good now. And so it feels like I did do some spiking, balanced spiking network stuff during my PhD and just felt like I'm never going to get these things to do computation, so I might as well switch over to rate based networks. But now it kind of feels like, oh, I can actually train these things to do useful stuff. So the project is basically we are training spiking networks using an algorithm called eprop. And actually EPROP has a local version, like a purely local learning version. We're training to do various reinforcement learning tasks. So there's the standard RL tasks like cart pole and another one where the thing has to balance something in the right way. Which are, which are basically the MNIST of reinforcement learning. But we're also training it to do a foraging task in grid world where there's a reward and the agent needs to move around and collect the reward, and then a new reward appears and it needs to go find another one. So spiking networks can solve these kind of naturalistic tasks. And so at the same time as using E prop to train it to do the task, we're also training the network to regulate its own firing rate and comparing different ways to do that. Right. So you could imagine just trying to minimize firing rate. That's like a standard machine learning thing. And what we're doing is instead having a loss, which is try to maintain your activity at a firing rate set point, right? So if neurons fire too little, bring them up. If neurons fire too much, bring them down. And so what we find is that one, this helps training and performance a lot. So networks are much more stable when they do this. They learn faster, they're less likely to crash. So in reinforcement learning you have this issue where like some networks just randomly crash, you don't know why. And this basically prevents that problem. And what comes out of this is inter spike interval distributions and firing rate distributions that look an awful lot like what we see in the brain. And so for me, this makes this really nice connection between a homeostatic function. Right. Just maintain your spiking at some level. We can call it a healthy level of spiking, which helps computation in a way that we can quantitatively say like learning is faster, performance is faster, or performance is better at criterion, you know, at the end of training. And networks are more stable, like their training is more stable. I would like to do this during sleep. And I keep suggesting that the student do this. We haven't quite. [00:44:53] Speaker B: Okay, yeah. I mean, how would you. Well, how would you implement sleep in the model? [00:44:59] Speaker A: Yeah, so. And actually. Oh, this is, this is where the project started, was actually trying to work with spontaneous activity. [00:45:07] Speaker B: Right. [00:45:07] Speaker A: Most, most spiking networks, like classically are, I would actually say a loose analogy to sleep in that there's no like sensory input. Right. [00:45:20] Speaker B: They're. [00:45:20] Speaker A: They're just balancing their activity to, to spontaneous activity, let's say. So I would start with just spontaneous activity and then I would try to induce sleep. Like dynamics, maybe up and down states, maybe. I mean, spindles are tough, you need a thalamus for spindles. But you can get local up and down states and you can get local sharp wave ripple type things quite easily [00:45:46] Speaker B: in spiking networks by generating spontaneous internal activity. Yeah. [00:45:50] Speaker A: So usually it's a matter of. At least the way I've modeled it in the past, it's just a matter of adding a slow adaptation current. Right. So you have some source of flow, negative feedback on neural activity or depolarization, and that generates dynamics that look an awful lot like slow waves in the right regime. Right. It can generate other kinds of dynamics too. But there is a dynamical regime that looks a lot like slow waves. [00:46:23] Speaker B: Yeah. Sorry. So you are interested. Like you wouldn't be satisfied if the model performed the function, but the dynamics did not look similar to brain dynamics. [00:46:36] Speaker A: I would be partly satisfied, but not wholly satisfied. Right. I think that in itself is interesting. Right. Just getting spontaneous activity to do some function. And then there's another layer which is spontaneous activity that looks like brain to do some function. I think these are steps in the same direction. [00:46:55] Speaker B: Yeah. I was going to ask you how. So there, you know, like when you do modeling, I mean, this is kind of a question, but I'm going to like. It's gonna sound like a statement. Right. So you. Some things, you think, okay, well, how. Like you were saying, how do we get the. How do we implement homeostasis? Right. In the Model. And then so you toyed around. Maybe we minimize the activity, maybe we make it just right. Maybe near criticality. No, just kidding. [00:47:22] Speaker A: But it's on the to do list. [00:47:25] Speaker B: Yeah, but, but so some things you, you need to implement and you have to figure out how to implement them. And some things you want to emerge from the model itself, like the dynamics. Right. So the homeostasis thing, for example, I mean, is that a cheat, implementing a, you know, a set firing rate? Or are you, are you happy with, with implementing that if what you get out is something that is reminiscent of what you would see via homeostasis? [00:47:55] Speaker A: So you asked earlier, in what ways is theory easier or harder than experiments? That's one of the ways that theory is harder than experiments is like that decision. That in itself is the art of modeling, is how do you set up your experiment? Which is what are the assumptions that I build in and what are the things that I want to emerge? And the things that you want to emerge are the things that are explained by. Is that the explananda or. The explananda? Yeah, I always forget. Anyway, that's the thing. [00:48:31] Speaker B: Every single time I forget which is which I do. It's awful. [00:48:35] Speaker A: Yeah. Anyway, so the things that you assume and build in are the things that are explaining the things that emerge. [00:48:43] Speaker B: Right. [00:48:43] Speaker A: And so in this case I would say that I don't know how homeostasis is being implemented. Probably it's through, you know, some sort of feedback from calcium and who knows what. But what I do know is that its net effect is to return activity to a set point and it adjusts excitability, synapses, maybe just inhibitory, you know, whatever it is, you can pick that. And then the question is, if I have a system that does that, what does that buy me? And so picking that is. Have you heard the saying that like picking and setting up your problem is like 90% of the way there or something? That's basically picking. Yeah, Whatever it is, that's. The take home message. Right. Like that is picking your problem and figuring out how to set it up in a modeling experiment. [00:49:45] Speaker B: I mean, I guess it depends on what you want to explain and what you're satisfied not needing to explain. [00:49:51] Speaker A: Yeah. [00:49:52] Speaker B: So you have a spiking neural network that can solve this grid world task. And it has some properties that look like brain dynamics, et cetera is the idea. And you want to implement sleep in the model, are you going to need, I mean, are you going to need to have a thalamus model and a cortex model? And a hippocampal model and like the different models talking to each other. How does, how would that look? [00:50:20] Speaker A: So in this case, the next step would be we're just going to worry about the cortex and we're just going to worry about slow waves because the cortex is sufficient to produce slow wave, like, activity. [00:50:34] Speaker B: I mean, you say that like it's a small thing, but that's a. That's a huge. [00:50:38] Speaker A: Exactly. So, you know, you need to. You need to think. I always tell my students to think about n +1, where n is like what you have and what you're comfortable with. M plus one is the thing that you're adding. [00:50:51] Speaker B: I see. [00:50:52] Speaker A: And you never want to do n +5. So n +5 would be adding a thalamus and a hippocampus and maybe a basal ganglia. I don't know. [00:51:02] Speaker B: Is that why cognitive architectures fail? Because. [00:51:08] Speaker A: Well, if you want to research cognitive architectures, you should probably not use spiking networks. You should use rate networks. Right. Because those are much easier to set up with multiple layers that interact in different ways. And people have done that. And so you can take that as N and figure out what your plus one is. Well, I'm going to redesign this architecture to look a little bit more like dilemma cortical interactions. So to answer your question, what that looks like in the near future is having a phase in which the network is doing the task, having a phase in which the network is just doing spontaneous activity. During that spontaneous activity, varying the dynamics. Because maybe you don't just have slow waves. You have a space of different dynamical regimes you can do. And looking at what is the slow wave state by you. [00:52:04] Speaker B: Okay. I mean, yeah, that sounds excellent. Good luck with. [00:52:09] Speaker A: It's easy to say. [00:52:11] Speaker B: Yeah, yeah, it's easy to say. Well, you also shared with me, like, so you're interested just in general, about the possibilities that NeuroAI affords to what has become like sort of a boom in neuroscience. And you said the word earlier with air quotes. Naturalistic. So nascent naturalistic neuroscience has become a term. What is it? Like, ecological neuroscience? Yeah. Anyway, like, naturalistic neuroscience is fine, but because we have, like, so much more data now, we have such, you know, better technology where we can record everything and look at everything and set up every, you know, record all the variables, et cetera. And so there's this push to get out of, like, lab reduced experimental preparations and put organisms or models right in much more, quote, unquote, naturalistic, meaning more complex environments, complex tasks to solve etc. So what is your interest with in Neuro AI toward naturalistic neuroscience? [00:53:15] Speaker A: Neural networks are a platform that let us build model organisms that can do things like naturalistic tasks. I think that's really the interest. Right. And I think if we think that experimental neuroscience has been impoverished, then theoretical neuroscience has been really impoverished. Let's take the Monte and Cecillo model. That is a one layer rnn that solves a cognitive task. But the cognitive task is that its sensory input is either a one or a zero of like two on two channels. And it's noisy. Right. So it's fluctuating and it has a context queue which is either a one or a, you know, a one or a negative one, indicating is it the motion task or the color context? Yeah, I. [00:54:09] Speaker B: That's right. Yeah, yeah, right. [00:54:12] Speaker A: And so that. And then its behavior is to basically select left or right. Right. And that's impoverished. [00:54:20] Speaker B: You're mocking Monte and Solo. [00:54:21] Speaker A: No, no, absolutely not. This is like a foundational study. And actually this is like one of my. I recommend everyone read this paper, but that is way more impoverished than the monkey that had to solve this task because at least the monkey had to do it with visual stimuli and motor behavior. Right, right. And the idea, the reason to do this, which is a good reason, is that you're thinking only really about the prefrontal cortex. And at the time you couldn't get this thing to work with, with visual stimuli and motor behavior because you would need to model probably the motor system and the visual system. But neural networks now are at the point where you can model the visual system and the motor system and you can put them in virtual environments where they, you know, they have different states, let's call them hunger, thirst, sleepiness. And they have to achieve different goals based on what those states are. But you don't tell them that, you just reward them, you know, if one of those states. Right. Again, this is going towards, it's not a perfect thing, but it's going towards models that have to behave in environments with egocentric stimuli that are visual like and egocentric behavior. And then you can study how they work in these environments. [00:55:49] Speaker B: And that's not N plus 5, that's N plus 1. [00:55:53] Speaker A: No, that's N plus 1. Absolutely. So we have a project in the lab right now doing that. Right. So there's a lot of systems you could do with this or do this with. The system we're working with is a model based reinforcement learning system called Dreamer, Dreamer V3, to be specific, because V4 uses transformers. And I, I don't really care about touching Transformers, but. Right, so Dreamer basically has a convolutional neural network. That's the visual system. It has a recurrent network that is a world model. And then it has a mlp, a multi layer perceptron. Well, two of them. One is the value model, you know, the value, and one is the policy. And this thing can behave, and it can behave in Minecraft. Right. And so we're using a 2D version of Minecraft called Crafter, but the principle is basically the same. And the agent has states of hunger and thirst, and it needs to satisfy these things by eating cows and drinking water and avoiding skeletons and zombies. And this already exists. We didn't have to develop any of this. So the N1 is just poking around at it to see how it works, how it actually solves these tasks, making slight modifications to make it slightly more brain like in certain ways that we're interested in, and then trying to understand how those brain like changes support computation and behavior. [00:57:29] Speaker B: And what right now is the biggest challenge in that setup [00:57:36] Speaker A: at this point? The biggest challenge in that setup is figuring out which of the interesting things we've talked about doing with it are most worth doing next. [00:57:49] Speaker B: Okay, so it's the oldest setting up. The problem is 90%. Yeah. [00:57:54] Speaker A: So we, yeah, absolutely. Like, we have tons of ideas of things we could do. The question is, which of them are worth doing next? [00:58:06] Speaker B: Can you share a few of them? [00:58:08] Speaker A: So the way Dreamer works is it actually learned. The reason it's called Dreamer is because it works through dreams, which are basically internal simulations in the world model. And so what we, what we're basically doing, what we've decided is the thing to do next is to look at those dreams, try to decode what they're representing. Right. Like you would do with hippocampal replay. Decode which position is the. The agent thinking it's in. Trying to see when that looks like hippocampal replay, when it doesn't, when is it producing coherent trajectories through the environment? And how can we play around with that? One, to make it more hippocampus like, and two, to use it to learn better. [00:59:03] Speaker B: Ah, okay, cool. [00:59:05] Speaker A: Yeah. You know, another set of things would be looking at how is the representation of state of the. Of the agent. Right. Hungry, thirsty, sleepy, low health. How does that translate into representations in the world model that are then used by the policy network? [00:59:24] Speaker B: Right. [00:59:24] Speaker A: Because it turns out the hippocampus doesn't just represent place, it also represents basically which task you're doing, which is a very artificial thing and for an animal task really means like, what am I trying to do right now based on my body state and the state of the world? And so the question is, do you get similar task specific representations in the world model of Dreamer that we see in the hippocampus? [00:59:53] Speaker B: Does the term cognitive map encompass enough of what we think of the hippocampus as doing to? I mean, do you like that term? Does it not do it because it encompasses sort of the abstract space that people like Tim Behrens have worked on and it encompasses obviously spatial navigation, etc. And maybe memory spaces. I don't know. [01:00:15] Speaker A: Yeah, yeah. I don't actually know what a cognitive map is in a like strict definitional sense. I know in the reviews for our paper we had a reviewer really pushing back against our use of the word cognitive map. And actually we've had now two reviewers push back in different ways. One of them saying wasn't cognitive enough. And we were really talking about the representation which is a neural manifold that maps the environment. And we were saying that's the cognitive map. And then another reviewer saying that there's very, you know, less evidence that the hippocampus actually supports these cognitive things that people claim the cognitive map does. So these were like complete opposite things. At the end of the day, a cognitive map is some kind of usable representation of the states of the world and how they relate to each other and how I can get between them. Let's say that. But that's a high level thing. You can imagine a bunch of ways of doing that. Another term I've been liking lately is world model, but that is similarly ambiguous. [01:01:38] Speaker B: Okay, so I was going to ask you because this term world model has become seemingly quite popular. There's been like a recent issue, like completely devoted to world models. And I realized I don't know what a world model is. [01:01:52] Speaker A: So it's going to be at ccn. There's going to be a gak about this, of course. [01:01:57] Speaker B: Oh really? About what they are or if they're useful? [01:02:01] Speaker A: No, about if networks that learn to predict develop world models or something like that. [01:02:06] Speaker B: Okay, that's assuming. [01:02:09] Speaker A: Yeah, that depends on what you mean by world model and also like which predictive algorithm you use because we at least found that like different kinds of prediction result in different kinds of representation, which, you know. Of course. Of course, yeah. So I loosely take world model to mean, I guess it's similar to cognitive maps, some sort of representation of the environment. One thing A world model, I think gets at or like really tries to emphasize that a cognitive map doesn't is that a world model can be used to simulate plausible experiences, plausible trajectories through the environment. At least that's how it's used in reinforcement learning and model based reinforcement learning. But that's just because that's what they use these networks for. Usually it's a network that is trained, let's say, to predict the next state given the current state and action. And so once you have a system like that, you can use that to simulate, here's my state, I take this action. Now let me take my predicted state and put the action back in and predict the next state. And so you can use that to simulate rollouts of things that could happen. [01:03:27] Speaker B: Okay, that helps me because I've been seeing this term and realizing I didn't know what it was, but I've been resistant to learning about it. I've almost been like eye rolly about it because, oh, here's another term that is going to be important somehow, but not meaningful or what, you know, am I just old and cranky or what? [01:03:48] Speaker A: I think different people use it different ways. And like, like, I think part of this, I think part of the debate at this gak is going to be about like, are they world models because they can predict some things and not others and they use heuristics to do this. And I don't know if that makes it a world model or not. I don't know. A term I like for things like this, which I remember seeing on Twitter from Neuropolar Bear, who. I don't actually know who it is, but they are. Okay, they are someone who was on Twitter who was, was very kind of active on Twitter when science Twitter was a more active place. Is the term weasel words, right? So these are words that we use that aren't really well defined, but we all kind of use them and they're kind of weaselly. And this thread, which for whatever reason really stuck with me was about how and why weasel words are actually really useful because they let us make progress without before we're ready to define our terms, when we're all kind of like circling around some idea that we haven't developed well enough to define, but it lets us continue to communicate. [01:05:14] Speaker B: This is exactly, I think this is exactly what Vicente Raja means by the term motif. So he uses the term motif and he's written about this in papers. And it's exactly that. It's like something that we can kind of gather around it as a scientific community and say it and use it to make progress without really knowing what we're talking about. Which is an interesting concept. I like weasel words too. [01:05:41] Speaker A: It sticks with you. Yeah, yeah, yeah. I mean, I guess neuro AI is one of these, right? But maybe, I mean, I, I guess the hope is that over time they will either become less weaselly as we, you know, communicate and interact and develop or become replaced with more specific words. [01:06:09] Speaker B: But isn't. Yeah, but isn't what generally happens. Actually what happens is they get reified and mistaken for like become. They're like a real thing and then they're just assumed and there. People don't realize that they're talking about a metaphor anymore. They're talking about a real thing. Isn't that generally the way that it happens? [01:06:29] Speaker A: Yeah, maybe that's pessimistic. These are two ends of the spectrum from like things working well and things working oh no, the wrong end. I, I mean, who's to say if that's even things working poorly or not at the end of the day, right? The, this is all internal science stuff and the stuff that gets out is the stuff that people can use. And like, maybe people can't use our weasel words, but they were very useful for us developing the things that they could use. And like sometimes we get a little wrapped up in thinking that like the stuff that science, like we scientists care about is like the reason for doing science. And it turns out that maybe that's just like the internal fumes that we need to do to make the true product. I don't know. [01:07:28] Speaker B: Yeah, I mean, but there is like a, there's a whole socio cultural dynamic, funding dynamic on like now I'll start using world model and it'll increase my chances of getting a paper out or something, you know, or like, you know, everyone has to throw a mechanism in their paper and that increases the chances, you know, so there, there's that sort of dynamic, which I'm not fond of. But you, you said so. But anyway, that is very helpful to me to realize that part of the definition, let's say, of a world model is the ability to simulate possibilities. [01:08:02] Speaker A: That helps me, at least for me. Yeah. For personal reasons, because I'm interested in replay. The ability to simulate plausible experiences is the most important part of the world model. [01:08:18] Speaker B: Okay, fair enough. But you said you don't want to touch transformers. But I thought the brain is now a transformer. Why don't you want to touch transformers? [01:08:28] Speaker A: Maybe because I'm old and biased. I don't know. No, I, I have trouble. So I really like RNNs because I am comfortable with the mapping between RNNs and brain circuits and it's a loose mapping, but I don't quite have a good handle on the mapping between transformers and brain circuits. And, and maybe, you know, at a higher level, like an algorithmic level, there's a nice mapping and maybe there is even a good mapping at the level of cells, but I'm not sure what it is. And the thing that I really don't like about them, and this was a debate that, or a discussion that I often had with a few people in Blake's lab because. [01:09:24] Speaker B: Blake Richards. [01:09:25] Speaker A: Blake Richards Lab. [01:09:26] Speaker B: Yeah, sorry, you just gotta name drop these famous names, you know. So you've been around some pretty large names. So. [01:09:36] Speaker A: So the thing I really don't like is that when you want to do anything in time, right? So navigation, whatever it is, even, even languages in time, there are things that are earlier in sentence and later in the sentence. Transformers can attend to anywhere in their context window and that context window can be causal, which basically just means you can only attend to things in the past, but they're still non local in time, right? So in the real world, if you're thinking about things like a physicist, the state of a system at time T can only be affected by the state of the system at time t +1 and any external factors that influence the system at that time. There's no way for the system to have direct access to things that happened five seconds in the past. You have to go through states. And so RNNs capture that really? Well, the state of an RNN at time T is a function of time t plus one. Maybe that's discrete, maybe that's continuous, whatever. And the input for a transformer, it's, it's I just. That mapping and maybe it's just me, maybe this is like a failure of imagination on my part, but the ability of a neuron now to have direct access to what happened five seconds ago. And I mean, you know, direct access, right? Obviously indirect access through the activity of neuron, you know, that's fine, it doesn't fly with me. And maybe that's just a taste thing. [01:11:19] Speaker B: I mean you're talking about time, but is it time that you're talking about or sequence? Because what I want to ask you is whether neuro AI, let's say writ large, is paying enough attention to time. Because in the real world everything we do is we are bound by time constraints, right? And that's why robotics is so hard, etc. But I don't know if is that is time something that is not being paid enough attention to? [01:11:55] Speaker A: So I don't know how to say enough or not enough attention. Certainly there's a lot of work in neuro AI that's not paying attention to time. Really, really thinking about input output, where maybe the input is images and the output is image category. I mean, very few work is on that anymore. [01:12:13] Speaker B: But, but, but even like the dynamics work in like recurrent neural networks, which they're so good at, right? You, you mentioned Montaigne Sicilo. That was like the big thing there. It's like, oh my God, that's got like these internal dynamics that look, that look great, you know, but even that is more. We say dynamics, but it's not in time. It's just sequential activity, which we could run that thing over 12 years and get the same quote, unquote dynamics as we ran it in 12 seconds. [01:12:42] Speaker A: So I think there's two questions here. One is, is the field paying it enough attention? And two, does like the method of using ANNs allow you to do it? Right. So one is like a question of what are people doing? And I don't know. I do think ANNs do allow you to think about time, and I would like for people to use them more to think about time. But now kind of a question for you is like, what's the difference between the sequence in sequential activity or the sequence in Monte and Cecilo and thinking about time, like, what's the difference between sequence and time? [01:13:26] Speaker B: Yeah, so I, I don't know is, I think that's a philosophical question first of all. But the difference is life or death when you know, you're being chased by a lion or whatever, you know, like whatever the normal go to example is. But I don't know if that's important to understand cognition, but it sure seems important because like you were saying, we are made up of all of these different layers, spatio temporal layers that all have to work well together. And yeah, there is a lot of degeneracy and a lot of room at every level, you know, for, for different ways of doing that. But in some sense, you know, even homeostasis is all about time. It has to happen within a given time or you die, you know, so it's very important for our biological bodies and brains. But I, I'm not sure if it's important for understanding cognition. I would assume that it is. [01:14:20] Speaker A: Yeah, I would also assume that it is. And I Guess the question is, like, if you were to bring it back to how would you tackle this under a NEUROAI framework? It's basically, how would you build this into your network? You know, if I want to channel Blake again, architecture loss, function learning rules. And then the fourth one, which is not in that paper, but which should be, is data set. Right. These are the ingredients for a NEUROAI model. And so how would you build that question into that framework? And maybe it's a matter of your task environment, which is really your data set, instead of having a trial based structure where this network only exists in the context of behavioral trials, so it knows when they start and when they end. That would be Monte Ancillo. There's a behavioral trial. It starts with a queue, it starts with a context, a queue, and then it produces a behavior and. And then that network, you know, ceases to exist. And then there's another. Another trial. Maybe it's just formulating your experimental setup such that it has continuous time. And these trials, you know, you have ongoing spontaneous activity and every once in a while there's a trial that interrupts it. There's nothing to stop you from doing that other than, you know, the constraints of writing your writing, your network structure like that. And the other constraint is compute, which is that the longer your trials are, compute gets harder. Or like the environment set up. Right. Like. So in the naturalistic case that I was. Naturalistic case I was talking about before, the crafter environment. Right. I think that's a great setup to. Because that thing does have continuous time. It has predators that it has to avoid. Right. So now your trials are very long sequences in this world. And in fact, in that environment there's day and night. And so you kind of can start to build in time through the environment and maybe also in the network setup too. That's a separate question. [01:16:59] Speaker B: I don't know. Is the convolutional neural network used in dreamer? Is that a recurrent. [01:17:04] Speaker A: Yeah. [01:17:05] Speaker B: Okay. [01:17:05] Speaker A: So the sensory system is a convolutional [01:17:08] Speaker B: net, but it also has recurrent dynamics as well. Recurrence. [01:17:11] Speaker A: No, not in the. Yeah, so the only part of DREAMER that has recurrence is the world model. This is another piece we've talked about changing and seeing what it gets you. [01:17:23] Speaker B: I think one of my interests and like, why I asked you about time and thinking about, like, what's important biologically, I'm sort of surprising myself because, you know, when I got into neuroscience, it was all about, you know, compute computations. Right. And that's sort of how I was brought up and thinking about everything in terms of computations. And that's why neural networks are so damn cool because they can do so much stuff. But I'm surprising myself lately, I think as a, almost as a reaction to people who are worried about like AI becoming conscious, right. Or you know, that they have these minds that we need to worry about their welfare and stuff. And I'm sort of like, whoa, whoa. Like, but then if we buy into like this is what cognition is, maybe we do need to be worried about that. And so my intuition, and this is terrible as a scientist, but my intuition is just so far removed from worrying about any AI being having a mind to worry about. But then I need to sort of prove to myself that that's the case. And so I think, well, what's important? And so I'm getting on, you know, into things like the organization of biological organisms and brains, et cetera, and why even metabolism might be important and time. And so I'm sort of struggling in this sea of like what I even think is important for cognition. And I'm not sure if I'm too far out at sea sometimes. [01:18:51] Speaker A: There's a lot there. So I agree with you, right, that neural networks are missing this kind of aspect of the dynamics. But I don't think that means that they like it couldn't be built in. It just means that they haven't been used in that way. And same with like biological organization and autopoiesis and all this stuff, right? Like, I don't think that that's incompatible with the, a neural network based framework. I just think it hasn't been used in that way before. It's funny that you mentioned that because I, on my honeymoon actually, I was reading was it Principles of Biological Organization? Varela? [01:19:53] Speaker B: Yeah, yeah, yeah. [01:19:56] Speaker A: Which is great, right? Yeah. Well, they just re released. It's been on my reading list for a while and then they re released this updated version with like commentary and stuff. I didn't, I didn't finish it. I love to start books get like halfway through them and then they sit on my bookshelf like half finished. But you know, the, the, The idea of like organizational closure and stuff like that, like this could all be built into a neural networks framework. There's nothing to stop you from doing that. [01:20:32] Speaker B: No, it can't though, because the whole point is that it's, it builds itself, right? So like in your modeling, right, you're going to put, you're going to clamp spiking rates to like Some medium level. And that is externally done in an autopoietic system or the whole point of organizational closure is that the system builds its own constraints and the constraints build the system and they are mutually. Have to have each other, mutually dependent on each other. And so that's like sort of the magic of life. Right. And so you can't build, you can't build it unless you have children. Right. That's not happening in your life yet, right? [01:21:11] Speaker A: Yeah, yeah. I mean, I certainly don't think we are doing it. I would push back against the idea that you can't. And I think it's more of a challenge of thinking how would, within this framework you do that? And I don't know the answer. Certainly it's not giving it a target set point. Right. It's the question of how the target set points emerge. [01:21:34] Speaker B: Right. [01:21:40] Speaker A: And it's an interesting problem, you know, and maybe it's, maybe it's too much trying to fit a square peg in a circle hole or whatever to like try to adapt a ANN type framework to this. I, I don't know. [01:21:54] Speaker B: I don't want an ANN to be alive either. [01:21:56] Speaker A: Yeah, sure, that's. That's a separate issue. I also think it would be a terrible idea to have it. [01:22:04] Speaker B: Right. [01:22:04] Speaker A: And be alive, you know, and. Yeah, and we, but we could talk about why. I think it's like why we think it's ridiculous the idea that these things are conscious. But I think that's. [01:22:17] Speaker B: Why do we. Why do we think it's ridiculous? Yeah, let's talk about that. [01:22:20] Speaker A: I mean. Oh, for me, I think it's because they don't have spontaneous activity and they don't have intrinsic goals. [01:22:28] Speaker B: Okay. That's where the biological. That's where organizational closure comes in. Right. Because that's where things get its meaning and its goals. All internally generated. [01:22:37] Speaker A: Yeah, absolutely. [01:22:39] Speaker B: I don't want them to have their own goals. [01:22:43] Speaker A: I'm working on the spontaneous activity part. [01:22:45] Speaker B: I know, yeah. But you think the spontaneous activity is where it's at. [01:22:52] Speaker A: I think it's important that when I am not, quote unquote, doing a task, my brain keeps doing stuff. Right. Like these things are only, let's, let's say these things are conscious. They're only conscious of the text that they receive when they receive it and the text that they produce when they produce it. And then the lights are out and [01:23:24] Speaker B: they've got world models. Right. That's where this, like a term like that is going to like, bite us because, oh my God, they've got all [01:23:30] Speaker A: this paper you read, either they do or they don't have world models depending [01:23:35] Speaker B: they have cognitive maps. So. Okay, so you're not worried about. [01:23:43] Speaker A: I'm worried about much more with these things than if they're conscious. I think there are much more worrisome things around the development of AI and LLM based technology. Then are these things conscious? [01:23:58] Speaker B: I mean, what does it mean though? That there are very, very bright people who really do believe that that is what conscious. You know, Hinton. Right. Even Godel, Escher, Bach, what the hell's his name? I can't. Hofstadter. Even Hofstadter is like worried about it now, which, you know, very bright people. So it's just odd to me. [01:24:24] Speaker A: Maybe it means that conscious is a weasel word. [01:24:27] Speaker B: Oh my God, there it is. There it is. Yeah. [01:24:30] Speaker A: I don't know. [01:24:32] Speaker B: You also have interest in the role of theory, theory itself. So can we talk about that a little bit? What is. And you have kind of a pluralistic pro. Pluralistic mindset about this. And you think of theories and just correct me if I'm wrong, you think of theories almost as heuristics. [01:24:56] Speaker A: Yeah. So I would say two things. One is that models are very well defined things. Right. You set down a structure and you state the mapping of the elements of that structure to things in the world you can measure and pieces of your theory. Theories are very fluffy things. And I guess the, the, and the, you know, philosophy of science has lots of different views on this, that it's like a collection of sentences or a collection of, of logical statements and things like that. You know, I think, for me, I think of a theory as a big bag of assumptions. [01:25:48] Speaker B: Okay. [01:25:48] Speaker A: And maybe that's heuristics, but you know, a big bag of assumptions you kind of carry around that kind of play nice together that you can apply to questions you have or problems you have. Right. So if you have a theory, let's say the complementary learning systems is a theory. Well, that's like a, a large body of different assumptions about how the brain works, how memory works, stuff that happens during sleep, whatever it is that I can. Then when I have a question or see something I don't explain, I can pull from those assumptions and see if they can account for the thing I can explain. But I also want to preface that by saying that's a pretty loose definition and I am not a card carrying philosopher of science. [01:26:38] Speaker B: Well done. [01:26:41] Speaker A: One of the big take homes from that paper is that models are well defined things and because of that they can kind of play this bridge between theories which we can't even agree on what they are. Like big picture, what is a theory in general? As well as like little picture, what is any one theory? Like if I were to ask five people what is complementary learning systems theory? They might give me five different answers. But if I were to ask what is this specific model? Well, you can write down the equations and talk about how things. And so models in that sense play this bridge between data which again you can quantify and theories. [01:27:26] Speaker B: And so we're back to the virtuous circle theory, modeling data and throw experimentation in there, [01:27:35] Speaker A: which you can do experiments with models too. [01:27:37] Speaker B: But that's right, but it is nice too because a lot of people do conflate a model with the theory. So it's nice to make the distinction. [01:27:48] Speaker A: Often a theory will have many different model instantiations of it. Right. Like there are a bunch of different models that are different kind of specific instantiations of complementary learning systems theory. One might be based on hopfield type networks. One might be based on spiking networks. One might be kind of a very high level model of just vectors that correspond to memories. I don't know. But yeah. And each of those you can kind of really write it down and study its properties. [01:28:29] Speaker B: What is the problem ladenness of theory? [01:28:35] Speaker A: Yeah, so this is the idea that. Well one, it's just kind of a tongue in cheek title for a paper because there's this idea of the theory ladenness of experiment which is that you always come to experiments with preconceived notions, a set of theories. And there's no such thing as a theory free experiment. And so the idea of problem ladedness is that you always design theories with problems in mind. Theories are made usually because there's a set of problems, scientific problems, things that we don't understand, things we're trying to explain or practical problems, things that we're trying to do in the world. And so we use our theories to explain those things to solve those problems. We develop them for that, they are inspired by that. And so it's the idea that any theory is laden with the problems it was designed to solve. [01:29:34] Speaker B: Yeah. [01:29:35] Speaker A: And that actually we, we, we judge theories by the problems that can solve. We design them with problems in mind, we compare them based on their ability to solve certain problems. Yeah. So it's the idea that theories, there's no such thing as a problem free theory. [01:29:56] Speaker B: And that also helps you embrace the neuro AI approach I would imagine because it's like a specific set of tools for specific problems with background theoretical assumptions in mind. [01:30:12] Speaker A: Yeah, yeah, absolutely. And in fact, I think it. So in. In that paper, we introduced this idea called the problem space, which is the. The space of problems that, well, either a field is interested or that a theory can cover. And so I would say that neuro AI really expands the problem space of neuroscience to computational problems, to the development of neural networks, to things like that, which I think is a good thing for neuroscience. But I think it's not healthy for the field if the only problems we are trying to solve with it are, you know, biomedical problems, for example. [01:31:09] Speaker B: Right, right. I run this complexity discussion group where we're going through all the quote, unquote foundational papers in complexity sciences that the Santa Fe Institute put together. And one of the things we recently. [01:31:22] Speaker A: I'm actually in that discord, and I still get notifications, even though I think I only attended, like, the first one. Yeah, yeah. [01:31:29] Speaker B: Okay. [01:31:29] Speaker A: Turns out it's hard to find time when. When you start your lab. [01:31:33] Speaker B: No, it's hard for me to do it. It's hard for me to find time because I'm. Yeah, Anyway, that's a whole separate thing. But anyway. We recently studied what began the morphospaces sort of approach to thing, which was shell coiling and the possibilities. So this is Raup R A U P I think is the author. But so, you know, you think of like a state space in neuroscience. Right. The possibility of firing rates. Well, this was done with shell coiling. And what it was found is like shells can only occupy like a very small fraction of that possibility space. And so I've been thinking about this in terms of how to apply that in terms of neuroscience. I've been thinking more and more about morphospaces because everything's a state space, which is essentially a morphospace. But in terms of where the state can visit, I feel like. And I guess that's. That would be like topological approaches, et cetera. Anyway, I don't know. This is. [01:32:36] Speaker A: This is related to. What is it? Waddington's developmental landscape. [01:32:42] Speaker B: Yeah, the canalization. Yeah, Back. Back to. Right. And I don't really know where I was going with that. I just. Something that you said made me. Made me think of that. [01:32:51] Speaker A: Something about problem space and expanding the problem space. [01:32:54] Speaker B: Oh, yeah, that was it. Yeah. Because you. So the other thing about state spaces is, like, in a biological organism, it's an open system, so your state space can actually change to the adjacent possible or whatever. And I wonder if neuro AI has sort of changed the state space of the problems that we can even study. Is that a good way to think of it? [01:33:16] Speaker A: Yeah, yeah, I think so. And in doing so, I think we will learn a lot about the brain. [01:33:24] Speaker B: That's a fantastic place to end it, Dan. So is there anything that we didn't cover that you want to make sure that we discuss? [01:33:34] Speaker A: No, I don't think so. This covered a lot of ground. [01:33:38] Speaker B: Well, congrats again on the marriage. I hope it's going swimmingly. And when is the first? [01:33:44] Speaker A: It's like not being married the better. [01:33:49] Speaker B: Was this. Was. Is your partner someone that moved with you to Yale or. [01:33:54] Speaker A: Yeah, so actually my. She moved with me to Montreal. Oh. [01:33:58] Speaker B: Oh, okay. [01:33:59] Speaker A: Yeah, yeah. So we were long distance during COVID She was in Colorado, had a weird flanking of the name of that state, Colorado. And I was in New York and then we moved to Montreal together. [01:34:14] Speaker B: Wow. Okay. [01:34:15] Speaker A: And some. [01:34:15] Speaker B: Well, that's. It's nice to meet you. Yeah. Now that your life is. You're basically retired and you're not busy at all, it's nice you can enjoy the marriage. [01:34:25] Speaker A: Right. [01:34:26] Speaker B: Don't have kids yet. Don't have kids yet. Wait on the kids. Yeah. Okay. Anyway, thanks so much, Dan. I appreciate it and I hope we connect again soon. [01:34:35] Speaker A: Yeah, absolutely. Are you going to go to ccn? [01:34:38] Speaker B: I don't know. When is CCN this year? [01:34:40] Speaker A: It's next week. [01:34:42] Speaker B: Oh, then no, no, I won't be going. Are you doing. Are you participating in that GAC or you're not going? [01:34:48] Speaker A: No, no, I'm. I mean, I'm going because it's in New York and it's a train ride away and I can charge my own lab. [01:34:55] Speaker B: Yeah, there you go. Yeah, but you're not presenting something. [01:34:58] Speaker A: But I'm not presenting. [01:34:59] Speaker B: Yeah. Yeah. Okay. Well, it should be a good time. Okay. Yeah. So. So thanks for doing this and yeah, absolutely. Much luck. [01:35:07] Speaker A: This is a lot of fun. [01:35:15] Speaker B: Brain Inspired is powered by the Transmitter, an online publication that aims to deliver useful information, insights and tools to build bridges across neuroscience and advanced research. Visit thetransmitter.org to explore the latest neuroscience news and perspectives written by journalists and scientists. If you value Brain Inspired, support it through Patreon to access full length episodes, join our Discord community and even influence who I invite to the podcast. Go to BrainInspired Co to learn more. The music you hear is a little slow jazzy blues performed by my friend Kyle Donovan. Thank you for your support. See you next time, [01:36:00] Speaker A: Sam.