Skip to content
PodSubs
ShowsPeopleGuestsJobsNewsEvents
More
BlogPodCred
Menu
ShowsPeopleGuestsJobsNewsEventsBlogPodCredNetworksSign in
Join nowSign in
  • Shows
  • People
  • Jobs
  • News
  • More
    • Guests
    • Events
    • Blog
    • PodCred
    • Networks
    Join nowSign in

Mind & Matter

Mind & Matter artwork

Mice Learn Just by Exploring, and It Speeds Up Later Training | Marius Pachitariu | 235

Jun 26, 20251 hr 50 minAudio
ShareXLinkedInFacebookWhatsAppEmail

Send us Fan Mail How brains compute and learn, blending neuroscience with AI insights. Episode Summary: Dr. Marius Pachitariu discusses how the brain computes information across scales, from single neurons to complex networks, using mice to study visual learning. He explains the differences between supervised and unsupervised learning, the brain’s high-dimensional processing, and how it compares to artificial neural networks like large language models. The conversation also covers experimental techniques, such as calcium imaging, and the role of reward prediction errors in learning. About the guest: Marius Pachitariu, PhD is a group leader at the Janelia Research Campus, leading a lab focused on neuroscience with a blend of experimental and computational approaches. Discussion Points: The brain operates at multiple scales, with single neurons acting as computational units and networks creating complex, high-dimensional computations. Pachitariu’s lab uses advanced tools like calcium imaging to record from tens of thousands of neurons simultaneously in mice. Unsupervised learning allows mice to form visual memories of environments without rewards, speeding up task learning later. Brain activity during sleep or anesthesia is highly correlated, unlike the high-dimensional, less predictable patterns during wakefulness. The brain expands sensory input dimensionality (e.g., from retina to visual cortex) to simplify complex computations, a principle also seen in artificial neural networks. Reward prediction errors, driven by dopamine, signal when expectations are violated, aiding learning by updating internal models. Large language models rely on self-supervised learning, predicting next words, but lack the forward-modeling reasoning humans excel at. Related episode: M&M 44: Consciousness, Perception, Hallucinations, Selfhood, Neuroscience, Psychedelics & "Being You" | Anil Seth *Not medical advice. Support the show Support my work: Good Chemistry: Personalized, science-based health consulting. Work with Dr. Nick Jikomes directly. Affiliate Partners: Visit this link to see my affiliate partners and get discounts codes for products & services to support health. For all the ways you can support my efforts.

Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.

Transcript

From the show. Tap a time to play from there.

Nick JikomesWhether food, drugs or ideas, what you consume, influences, who you become. On the mind and matter podcast, we learn together from the best scientists and thinkers alive today about how your mind body reacts to what you feed it. Before starting mind and matter, I spent 10 years in academia doing scientific research. I got a PhD in neuroscience, where I focused on neuro endocrinology and the neurobiology of behavior, and before that, I specialized in molecular developmental and evolutionary genetics. I use my scientific background to help parse and translate the information that guests share on the podcast. In addition to the podcast, I write long form written content inspired by the show, where I integrate what I've learned across episodes. I also have a free weekly newsletter where I provide you with upcoming guests, share links and provide commentary on scientific studies and research that I'm reading and more, visit mind and matter.substack.com. To find all of my content. You hope I got that pronunciation. Okay, thank you for joining me. And can you tell everyone a little bit, a little bit about yourself and what you study?

Marius PachitariuYeah, sure. Thanks for having me. I am a group leader at the Janila research campus. What that means is that I have a lab here. I've been here for eight years now. I'm not at a university. I'm at a research institute. We are funded through a charity. That means that I don't have some of the responsibilities that come with being a professor at university, so I can spend most of my time doing research, and that's what we've been doing. And I think we're here to talk about some of this research.

Nick JikomesYeah, yeah. And so what is your field of study? Would you say that you're a computational neuroscientist? What is your field that you're in now? But also, what's your background in terms of how you were trained?

Marius PachitariuYeah, I would say right now I am a neuroscientist. Period, I did certainly arrive here via computational neuroscience. It's not unusual for people in this field of neuroscience to need to have multiple kind of backgrounds, and they get here from all sorts of disciplines. In my case, I started with kind of a pure math undergraduate degree, and kind of transitioning slowly into more applied math. Then the Applied Math became computational neuroscience. Then the computational neuroscience mixed a little bit with machine learning in my PhD that I did at the Gatsby unit in London, and then eventually my post doc, I realized I can make more progress if I also do some of my own experiments. So that's when I actually picked up some experimental skills as well. Started doing experiments, and that just kept going in my lab, where I now have kind of a full complement of kind of experimental side to the lab and computational side to the lab, and everything else that's needed to run such such a lab.

Nick JikomesAnd so what is your lab study at a high level? What are the kinds of neuroscience questions and the kinds of systems that you're working with,

Marius PachitariuI would say, at the very highest level, maybe the lab studies just how network computations happen in the brain through kind of coordinated activity between all these neurons that we have in our heads, right? A lot of it, a lot of the complexity emerges from just the vastness of the neural networks that each of us has in their heads, but also that the animals that we study have in their heads something as, you know, small as a mouse, which is the organism that we Study, has, you know, more than a million neurons. And if you put that in perspective, that's actually more than some of the biggest neural networks that we have today, the artificial neural networks that we have today. And so a lot of the complexity arises from the kind of high dimensional nature of this system. And I would say that where our computational kind of skills come together with the experiments that we can do, is to really develop new methods to study such high dimensional systems,

Nick Jikomesbefore we get into like, some of your work, concretely, some of the recent research. Some of the tools and the methods and all that. Let's just talk at a high level about computational neuroscience, and neuroscience more generally, you said something earlier about network activity. Can you define for people who aren't familiar with neuroscience, or not intimately familiar, how do we think about things at different scales when we start to think about how neurons compute things and how information is encoded, going from individual neurons to circuits to networks. What do these things mean?

Marius PachitariuYeah, yeah, good. We can certainly start at a single neuron level, but one could start even below that, right? So like you said, there's multiple scales at which computation takes place. They all happen simultaneously, kind of all at once, and when we think of any one particular computation that eventually leads to something measurable, like I can reach my hand out to pick a cup, there was the effect of all these different scales and all these systems, kind of computing, you know, kind of propagating neural activity until it led to me reaching out my my hand. And so it starts at the level of single neurons, which are themselves computational devices. Single neurons are very small, right? Something like 1010s, of microns across, like a single cell. However, they have these processes that extend pretty far away relative to their sizes, their dendrites, their axons, those integrate all sorts of inputs from other neurons. This happens at very, very small, microscopic level. Synapses get triggered. Voltages inside of this neurons kind of accumulate this activity. Ion channels open and close. So a lot of stuff goes on and on and on until, kind of the output of this neuron, classically thought, is an action potential that travels down of its axon and then goes to kind of contact other neurons that are going to be receiving this message. And it's this property that makes it a network of neurons. Every neuron is connected to something like from 1000 to 10,000 other neurons, and of course, they're all kind of recurrently connected to each other. Not every neuron is connected to every other neuron, so the exact properties of which neurons connect to the to each other leads to this kind of dynamic of exchanging messages back and forth with action potentials, this kind of dynamic through the particular structure of those connections leads to a computation and leads to this final outcome that is, you know, me reaching out my hands to pick up a cup,

Nick Jikomesand yeah, I want to, I want to talk about the question of computation at multiple levels. We're just starting very basically like when we say computation, what exactly do we mean in biological systems, whether we're talking about the single neuron level or the circuit level or the network level, most people have an intuition for what a computation is. You can have a handheld calculator in your hand or your smartphone and all of these computer devices that we've made. But when we think about biological systems, are they do we mean that they are literally computing things, or is this just kind of like a metaphor?

Marius PachitariuThat's a good question. I think we can avoid. There's a whole right kind of debate about, you know, is the brain a computer? Is it like a calculator? I don't think that's necessarily the right kind of debate. I think one way to think about it is actually not unique to the brain. It's the same for a lot of physical systems that we make models of these systems. The models of these systems describe their behavior, you know, to as good of a level as we understand it. And then we can make models of neurons as computational systems, such a model would, for example, you know, some of the simplest models are just a neuron as a perceptron. It's a great simplification, but we can imagine a single neuron gets 1000 different inputs that arrive at different synapses with different weights, those get kind of summed up and the neuron outputs the sum, the weighted sum of those inputs. The model is clearly a computation. It's something I can write down and simulate in the computer. Whether the neuron behaves closely to this model is something that we now have to test, right experimentally. We test incrementally, such tests people have run before. In some cases, this can be a good approximation. Can be a good model. In other cases, it turns out you have to model some more, because, you know, like, individual synapses might actually, for example, be on similar branches of the neuron, and they might have, like a sub unit of computation happening over there, and it's only the results of the sub computation that gets kind of transmitted forward, and the neuron as a whole integrates all of that. So I would say the kinds of models that we make of neurons and of networks of neurons are very much kind of compute. Things. They have inputs and outputs. They transform a set of inputs to a set of outputs. Whether or not the neurons in the networks actually do that is on the experimental neuroscience to to prove or to disprove.

Nick JikomesSo, yeah. So, so maybe we can just a very simple level. We can we can model neurons and networks, we can say, Okay, we've created this toy model. It's a computer inspired model. Some calculation is happening, like, say, something is being summed together. So if I think of like a handheld calculator, it's built electronically such that if a certain pattern is inputted, another output, a certain output is reliably given. So if it's some sequence of digits and then the plus signal and then some sequence of digits, it will reliably add those things together and have a defined output. And at least some of the time, neurons tend to behave in that that way, just generally speaking, you can model them as something like a computational device.

Marius PachitariuYeah. Yeah. And, like you said, right, a lot of the things that can get modeled that way, you know, especially in deterministic way, like a calculator would be, tend to be, tend to be true on the smallest scale, like, if I give I raise the voltage of a neuron, I can predictably, 100% make sure this neuron is going to fire an action potential send that message down the axis on but then as it goes to kind of higher scales. Now we have these networks of neurons. They're kind of coupled together strongly. Sometimes it doesn't matter which neurons fires the action potential, as long as the message gets relayed. And so you get into these more, less deterministic cases. Of course, physically, everything is still deterministic. But you get into cases where the behavior of a single neuron might not be, you know, might be sometimes fail on individual cases, for example, me reaching out a hand, a neuron in my brain might activate 50% of the time that I do that another neuron might activate the other 50% of the time. So that's where, kind of the the network effects start to come into play, where it's not necessary for, you know, every neuron to behave like a calculator, just for the system as a whole, is kind of a larger scale, to behave relatively, you know, deterministically. And then it comes to, of course, things that make us more human are the fact that we cannot always be reliably predicted. Right? We do things that no one expects sometimes, and exactly how that kind of decision propagates from kind of the lowest levels of single neurons firing through to networks that, you know, there may go one way or another in their kind of activation patterns. And me ending up saying one thing or another, you know, that's, that's the kind of, you know, kind of fascinating thing that we hope that one day we'll get closer to understanding, and it's honestly one of the things that makes neuroscience pretty exciting to work in, because it's a way of understanding ourselves. And I think a lot of people get into it because of that.

Nick JikomesWhat are some examples you might give of famous or well characterized computations that happen in the brain, mostly at the single neuron level. Just to get people started thinking about what's out there,

Marius Pachitariusome of the most famous models at the single neuron levels are the ones that led to Nobel prizes, so things like the Hudson Huxley model for in their case, it was for the giant squid axon, which is one of the biggest neurons that you can find anywhere on the planet. That's good, because it makes it easy to work with. It's kind of macroscopic. It's easy to record from. And what their model did, it was a model for the exact way in which an action potential might be triggered and propagate. Action potentials, you know, they are the message. They're the central kind of unit of computation. So very important to understand exactly how these action potentials develop and propagate, and what made that research exciting, from a computational point of view that even back then, right in the 50s and 60s, with pen and paper, they were able to make measurements, kind of extrapolate from those measurements a certain shape of a relation Between ion channels opening as a function of time and the voltage it generated in the neuron. And then from there, they were able to abstract away a computational model with a few subunits of computations in terms of different ion channels opening for different ions coming in and out, like calcium and sodium and so on. Them, and then make predictions with that model that, you know, kind of fit the data pretty well. And that was, you know, frankly, one of the first kind of computational neuroscience applications. And it, in some ways, it was the the way that the field originated beyond that in terms of of single neuron computations. Let's see what came after. Hubel and Wiesel are another famous example, famous example, they were recording from individual neurons in the brains of awake monkeys while these monkeys were kind of looking ahead at a computer screen, and on that computer screen various visual stimuli were displayed, and this was the first time when they were able to figure out what neurons in the kind of central Brain, kind of higher up in cortex, what those neurons were responding to. So what were the messages about the action potentials that these neurons were generating were recorded with an electrode, and they were able, the researchers were able to associate those messages to the exact stimulus being shown to the animal. And again, that was also done at a single cell level, kind of recording neurons, one at a time with an electrode.

Nick JikomesAnd earlier, you said something implicit in what you said was that this is a mistake that a lot of people make, especially early in their education, is that computations don't start at the neural level. They actually start lower than that. This is not necessarily gonna be the focus for our discussion, but can you give people just a basic sense for computations that happen within cells?

Marius PachitariuYeah, let me. Let me think of a good example of that. So a lot of neurons, especially the ones in layered structures such as cortex or hippocampus, that means it's got multiple layers. There will be neurons that have a Soma, kind of, in the deep layers, and they'll have an apical dendrite that extends all throughout these layers, and that makes the neuron kind of, you know, the soma is the place where we usually look at for computations, because that's where the action potential originates. But then there's the whole kind of shape of this neuron with all its dendrites, and the exact way in which these extend out that is a computational device as well. You can get connections onto this neuron that are formed, let's say in layer one at the very top, these tend to be kind of more intracortical connections from other neurons that are kind of traveling parallel to these layers. And you can have connections much closer to the cell body, where the cell body is in layers, let's say four and five, where those connections might be from neurons that, say in the thalamus, that bring in sensory information to this neuron. And then you might have kind of connections somewhere in the middle that are from inhibitory neurons in this network, which, for example, can modulate, or, like, shunt the activity that's propagating either from the top to the bottom or from the bottom back to the top. And the fact that the activity, for example, the inputs at the top, they're so kind of far away, it means that for them to actually have an effect on the soma, they kind of need to accumulate, or like, to potentially create something, some inputs that happen more locally. And that kind of computation requires a non linearity. And the fact that it requires a non linearity makes this model of a neuron as the kind of a simple linear addition insufficient. It means you now have to think of kind of separately, as the inputs to this neuron as being its dendrites. And then you can model the dendrites as different compartments. You can go into a lot of detail with this, and people do. You can model it with hundreds of compartments, where every compartment might, in fact, be the unit of computation. But again, that's not entirely known how many compartments you'd need to model a neuron very well. Neurons have been shown to do single neurons have been shown to do nonlinear computation, such as the XOR problem. That's a famous nonlinear one that the perceptron cannot do. Turns out that neuron can do that. So

Nick Jikomesbasically, every neuron, one neuron, is one cell with one nucleus, and at any given instantaneous moment, it can send itself out only one action potential at a time. But they tend to have very elaborate physical structure. Elaborate morphologies. So one neuron might send branches to different parts of the cortex. They might send them to different parts of the brain entirely that are very, very far apart compared to the size of the nucleus of the cell. And all of these different branches can basically take in separate inputs, have different sensitivities, different levels of information weight, they contribute to the whole cell itself. And so there's all these, like little sub computations happening in different physical parts of a cell that one way or another need to get integrated before the computation at the level of that neuron kind of happens.

Marius PachitariuAbsolutely yes. And one a good way to think about is that the output tends to be fairly homogeneous. Goes down the axon. It's the action potential. It regenerates down the axon. Even when the axon branches out, it goes to kind of all places. Then it's pretty similar. The input is where different parts of the cell can behave differently. On the output, it tends to behave similarly.

Nick JikomesSo even at the level of one neuron, thinking about one neuron at a time in the brain, it's unfortunately not as simple as, you know, a simple calculator doing simple additions. There's a lot of complexity happening within that one cell before we even sort of scale this up to thinking about large populations of neurons.

Marius PachitariuYes, that will be correct. The exact level to which that complexity comes into play is something we don't know yet, but we know that it's not as simple as a single unit of computation,

Nick Jikomesand so as we scale up our thinking to the level of the whole brain, ultimately in the organism, but just populations, larger numbers of neurons in the brain, and some chunk of the brain or something, what what starts to change or become different in terms of what we have to think about and how systems level effects and the dynamics are just gonna behave, start to behave differently than the single neuron level.

Marius PachitariuYeah, so there's a few different levels before we get, you know, we get there, right? So we can, we can kind of build up. So if we start from a single neuron, it might be embedded in a local network. This local network could be made, let's say it's a neuron in cortex. This local network will comprise kind of a small area around this neuron going up and down. What would we call a cortical column? A cortical column is basically goes down through the layers of cortex, layer one through layer six. Maybe, you know, you can, we can define it spatially, and it's to the extent we can define it as maybe one by one millimeters or a little bit below that it tends to be a region of space where neurons are kind of highly interconnected to each other, and it tends to this pattern of connectivity tends to be repeated throughout, let's say, the entire cortex, with a few modifications between different brain areas. It said this is basic as soon as, as soon as we put a neuron into this local context, which is still local, we're still not talking about brain wide computation. It's still very local. The Network dynamics, complexity starts to come into effect, and that is given by the fact that, let's see, if I have two neurons, I can only connect them in a few different ways, where you can have one neuron excite the other one and the other one could excite that or not. If I have three neurons, I might have a few more patterns of you know which ones excite each other. As we start thinking of the actual number of neurons just in these local networks, which is on the order of 100,000 or something like that, probably even more, and

Nick Jikomesyou're speaking about the rodent brain

Marius Pachitariuat this point. As far as we can tell, cortical networks have this property across, at least, you know, mammalian organisms, right? Which have which have cortex the the kind of very high interconnectivity at the local level is the case across mammalian species

Nick JikomesI see. So even even though, even though the size of the brain varies considerably when you get down to certain scales, the brains of different species can start to look at least approximately like one another in terms of their density and their overall connectivity patterns, even though at a different scale, things start to look very

Marius Pachitariudifferent Absolutely, and you'd have a difficult time distinguishing them if you've got two pieces of different brains, the main way to distinguish them would be by thickness. If you've got something that's very thick, like several millimeters. Then you'd know that it's, well, several centimeters. Then you'd know that it's a primate brain, or maybe a human brain, in a mouse brain, this thickness goes up to about one millimeter. As far as the density of connections, though, that is quite similar in the sense that in local circuits, there is very high connectivity, this leads to this combinatorial complexity problem that I was trying to kind of illustrate by going at more and more neurons. Right? The more neurons you have, the more combinatorially many patterns of connectivity you can have between these neurons. And if we don't know these patterns of connectivity, which we almost never do. It becomes very difficult to even predict how a network of these neurons will respond respond to its inputs. Because while we may be able to control the inputs, let's say, if there's like sensory inputs, just show an image, show a movie to an animal, we cannot really control what happens. As soon as one neuron starts to excite another, and that excites another, it excites another, and you start having all this reverberation of activity in the circuit, it can lead to outcomes that appear non deterministic, because we're getting close to something that people call chaotic behavior. We're getting close to this chaotic regime where just slight differences in initial conditions can quickly diverge and lead to kind of different outcomes in the computation.

Nick JikomesAnd when we start thinking at larger scales about systems, and popular, you know, multiple circuits, populations of neurons. How do you guys start to approach this in the lab, experimentally? How do you where? Where are we at today, in terms of technology, tools and measurements for how many neurons that can be recorded, and what the population size actually is that can be studied in something like a mouse,

Marius Pachitariuyeah. So that's where my lab and others we've been kind of putting our effort and energies into trying to record as many units of these systems. And it is at this kind of medium scale where everything is connected to everything, where we can start to record, at this point, from 10s of 1000s, almost 100,000 neurons at the same time, there are trade offs that we make in order to be able to record activity at these kinds of scales. I mean, firstly, you know, this is a lot. Up until not too long ago, not 1015, 20 years ago, it was still common to lower single electrodes into the brain record neurons one at a time. There was very kind of an arduous task. It was like a whole PhD of seven, eight years just to put the data set together of a few 100 neurons, which then was like exploited and analyzed, you know, very in depth for many, many more years after that, nowadays, we have several techniques that allow us to record neurons more efficiently. Maybe I can start with the one that is a more natural successor to like the single electrode recordings. These are high channel density probes. They're essentially still electrodes. They're kind of these long probes, except that you don't just get one channel of recording on them. You can get up to a few 1000 channels. In fact, one of the most well known examples that it's not the only one, but it's called the neuropixels probe. It's been developed over the past 10 years. We're now up to kind of better versions. There's been your pixels two, there's been neuropixels Ultra. Just last week, the nhp non human primate version of this probe was published. And these probes have really allowed people to record not one neuron at a time, but now something like, let's say, 500 neurons at a time. And then there's a different set of techniques that arose from kind of a separate, more, I guess, physics field, which are based around imaging. So instead of lowering electrodes into the brain, instead we put a big microscope above an animal. We replace a piece of the skull of the animal with a piece of glass, and we put certain molecular proteins inside of the neurons that will fluoresce when the neuron activates, when the neuron fires an action potential. There's a lot of calcium coming into the cell. It binds to these proteins that most of the time we use what are known as G camps. These are genetically encoded. Calcium modulated proteins. When calcium comes in, because of the action potential, the cell is nicely fluorescent, and if it's fluorescent, that means you can see it with a microscope. And that's what we do in the lab. We have been kind of building better microscope and better kind of ways of recording from large number of neurons with microscopy. And so those are techniques that get us to those 50,000 100,000 neurons. Yeah.

Nick JikomesSo when we talk about the types of animal research that you do that we'll come to when we talk about recording from large populations of neurons. There's essentially two types of ways of doing that. One is to stick a whole set of tiny, tiny, thin little electrodes into the brain. These are like tiny, little, thin wires. Each wire can have multiple channels on it, and each of those channels can pick up a signal from one neuron that you can isolate. But because you have many of these electrodes on the little implant that you put in an animal's brain, you can record from hundreds or even 1000s of neurons, potentially at a time. So you've got action potentials that you record from a bunch of different neurons in some little chunk of the brain. And then there's this other technique, which is a microscope based technique that combines, it's basically like the things from physics coming together with things from molecular genetics, where the the optical capabilities to point a microscope at a brain and record the image were developed together with techniques that allow you guys to engineer cells, neurons to literally light up, and then you can see them, literally watch them with a microscope. And so these two techniques are now at a place, and they both have, you know, pros and cons in different ways, but we've got two different potential ways to record from 1000s, or even 10s of 1000s of neurons at a time, meaning that we're talking about individual neurons. We're not talking about like an fMRI signal that smears everything together, correct,

Marius Pachitariucorrect, correct. Yeah, and let me tell you, you know, that's making projects much faster. It's making our ability to reason about this process is, you know, the computations and the behaviors much faster. We don't need to wait for, you know, many years to collect the data set. We can go into the lab, collect the data set already with 1000s of neurons, you know, kind of analyze it, and then come back the next day, and based on what we analyzed, just kind of set up a new experiment and collect a new data set. It's really been empowering, you know, people, a lot more people, to do these kinds of experiments, too, and one of those is myself, because I have a computational background. I when I was, you know, doing my PhD, I never thought that I would work in a wet lab at all until, you know, it was made sufficiently easy and approachable that even you know someone with a competition background like myself imagine themselves doing it. And then it turns out, you know, it's not as difficult as we thought.

Nick JikomesOne thing I want to touch on before we get into some other stuff, is when you start to think about it. You know, if you're recording from 1000 neurons, or 10,000 neurons, or whatever, the brain is, ultimately an organ with millions of neurons, many, many millions. If it's a primate or a human brain, but even a mouse brain has millions of neurons. As you said, this is more, I think you said it's more than the artificial neural networks that are doing all of the impressive things that we see in the AI world, even a simple animal like a mouse, that's nowhere nearly as sophisticated as a human primate, is already this very high dimensional creature with many, many neurons, when you're recording from that many neurons, and then we remember that neurons are expensive from a metabolic perspective, even just one action potential. Think about all of the protein channels that have to open and close, all the parts that have to move, like all of this stuff is takes energy. When we, you know, I'm trying to anticipate some of the more computational questions we'll get to. But when you're recording from this many neurons and starting to think about how the brain encodes information, how do you have to anchor that in thinking about metabolism and efficiency from, like an at literally, an ATP perspective?

Marius PachitariuYeah, no, that's definitely a good approach to come at this problem. And it turns out it's one of what I would call kind of several lines of computational thinking that has moved us over the years and and this goes back to maybe as far back as the 90s or something like that, when a lot of kind of, what we think of as the biggest computational theories arose what you're describing, or the way that we would, the way that people thought to to put that into a into a model, is to think along the lines of, how can I achieve a given computation? So I can write down a computation I want to achieve this with the smallest number of spikes, so by minimizing the total number of spikes across the network. Things that I'm going to generate in order to produce this computation. Then turns out, there is a constraint that we put on a problem, and those kinds of constraints can kind of limit the space of possible solutions to the problem. And it's one line of modeling to create models of this sort that have these kinds of constraints, efficiency constraints, and then see how well they match things that we see in the brain. This has taken a lot of different forms. You know, the the efficient coding hypothesis is kind of its its main kind of umbrella it can apply pretty well, and there's been some good predictions made at the level of, for example, sensors in our retina. Turns out, you know, those operate in a pretty efficient way, and they tend to align with the predictions of efficient coding theory. They can also align. The model can work also pretty well in species that are pretty small, something like a fly, or, you know, you know, in that region, an insect, right? And an animal that still has to do pretty complicated things like, you know, fly around the world without bumping into things and avoiding predators and all sorts of things like that. But it has, like a very limit now. It has a very limited kind of computational budget, because it has small, relatively small number. It's still like 100,000 small number of neurons, a very limited budget, energy wise, there, the models can still apply pretty well. Now you come to something like the human brain, and all of a sudden those efficiency constraints start to kind of not apply as much anymore,

Nick Jikomesmeaning the predictions of those types of models don't do as well.

Marius PachitariuThey don't do as well, and they can only go so far, because it turns out that you can't this is a constraint. Efficiency is a constraint. You cannot let the constraint limit the kinds of things you might do. What What I mean by that is bigger brains, especially in humans, they have to solve these kinds of complicated problems that we are not even sure how you would solve them at the level of of a neural network, and so kind of driving that by a constraint that you put on the brain isn't a good way of finding out what the problems are, or like finding out what does the this network compute? Yes, maybe you know it's not going to go way out of its budget of ATP in order to do that. But at the same time, if we're looking at all the action potentials that are happening in the brain, we're looking at a lot of the kind of action potentials that happen even, for example, in something like a very sensory area when our eyes are closed. For example, we do this in mice. We record neural activity in complete darkness in their visual cortex. Turns out there's still a lot of activity. It's structured in certain ways. We understand some of it, but it's certainly not completely quiet. It would be the as would be the prediction of, you know, efficiency as a fundamental principle. Yeah. And like my perspective, and it's kind of we could explore this in different ways, but I would say my personal perspective on it is that brains and networks that are this big, they find certain solutions to these problems that are kind of, you know, non kind of idiosyncratic. There are ways of connecting neurons that you wouldn't have thought about, that maybe like happen by chance at some point during evolution, the places that are kind of hard to escape from and really kind of optimize efficiency, which is why the brain is not very efficient, right? It consumes 10 times more energy per cell than any other part of the body.

Nick JikomesSo, I mean, it's really easy to get really into the weeds here. We have to keep in mind some things simultaneously for probably the rest of the discussion, which is nobody knows exactly how this stuff works. There's many different things that we can talk about and many different perspectives we can share. So we'll try and do a good job of explaining there's this model or this theory, and then there's this other way of thinking about things. But one thing I want people listening to keep in mind is two separate but related. Questions are, how does the brain compute the stuff that it computes or that it's doing? But then there's also the question of, what is it actually doing? Like before you know, and you kind of need to think about both separately and and together in parallel to to start to think about some of this stuff. So I want to ask you some a big question here, and there's no answer. I'm just making that clear for everyone. You can give your perspective and talk about maybe some of the other perspectives that are out there, but I just want to ask you the general question, what do you think, and what do neuroscientists generally think the brain is actually ultimately built to learn and encode? So for example, is it built to learn and encode sequences of actions, and to figure out how to do that from experience as efficiently and effectively as possible. Is it built to construct representations of stuff outside the organism, or to abstract regularities that you know? Are these more abstract things? Is it all the above? Can you start to talk about maybe your perspective and some of the general perspectives on this.

Marius PachitariuYeah, I'm sure it's all of the above, but perhaps I could share what I what I think might be like, one of the more crucial ones, and I think one of the more crucial ones for humans. And you know, we'll see when we get to study this in animals. There are certainly people going that way. Is that our brains are very good at making models of the world, kind of like you said, like representations of things outside, making models of the world, but also to they're very good at running those models forward in time, right in our heads, we can imagine what would happen if I did this or if I did that, yeah, mental time travel, that's right, but into the future, And that is an extremely important skill to have, because if you can predict how things how the world is going to react to your actions, you can avoid some of the most costly actions that you can take. Yeah, and you could really kind of find or narrow in on what is are the good actions that you can take. But I think a lot of that depends on, not only on having a model inside of our head of how the world works, but also our ability to kind of Gen. It's generative AI, right, but in our heads for kind of extrapolating into the future how the world would change had if we do this or that, or if we don't do anything at all, which is also very much something we do all the time. Why do I think that's the one of the most important things? It's, it's fairly, you know, it's we think it's more unique to us humans. We build on top of that ability very much. I think, a lot of the knowledge that we transmit, we transmit abstractly between people, and it's knowledge about how things work. It's knowledge about models of the world. We make models of the world, and we transmit it to our our you know, children, we transmit them to other. You know, members of our tribe. We transmit them through society to you know, our whole thing is to make models and to transmit them forward.

Nick JikomesAnd you know all of these things. You know, you might emphasize them or de emphasize them based on the person you're talking to. Might have different opinions on which ones to emphasize. But also these things might exist to different degree, different degrees in different types of organisms, like, obviously, a centipede is presumably not making the types of abstract models that a human being has, even though both animals need to, for example, chain together sequences of of, you know, muscular, motor, motor action patterns in an efficient manner. So, so when we think about mice, how do you sort of think about this in mice and and what the type of brain a mouse has needs to do are mice capable of abstraction a little bit or like, how do you start to think about what the mouse brain is built for? Right?

Marius PachitariuWe think so. Let's see where, what's a good place to start my lab more directly works right now on how animals, like mice, make models of visual environments, so not necessarily, how they, you know, extrapolate models in their head. That's seen something a little more uniquely human, although we have seen it in other people's work. For example. People in areas like the hippocampus, where we see replayed patterns of activity that correspond to the intention that an animal, like a rat or a mouse has of where to go forward. And that's that's pretty exciting research, because it's very similar to the kind of generative properties that that we think are useful for simulating the environment. We work on the part of like, how do we assimilate properties of the environment? A lot of that has to do with forming memories of one kind or another. In our case, because we're more of a visual, visual systems lab, we look at how these visual memories are being formed, and it turns out there are certain parts of the visual system that develop representations of the world develop these visual memories as an animal experiences it and becomes very familiar with it over time. We know they're new because they're not there, you know, kind of on the first day of training, we don't have these very selective neurons that respond based on the visual environment that the mouse is in. These get developed over days, and we find that animals can use these representations to respond well, to drive their behavior. First of all, so being visually familiar with with certain environments allows them to, afterwards, quickly learn which one is going to provide the reward for them. Reward training, kind of reinforcement learning, is something that a lot of animal labs do in order to get animals to produce an outcome of a computation. But then we also find that these representations kind of develop by themselves as animals just kind of explore these worlds that we create them, rather than, you know them, needing the animals to kind of forage for rewards. The foraging part isn't necessary in order to build an internal model of the world. And yes, we find neurons that seem to be behave like that in the mouse brain. And so we're kind of, you know, we just had a paper published about that, and we're continuing that work going forward,

Nick Jikomeswhen we start to think about the population activity of neurons, how large number of neurons behave, and what these patterns do, and how they encode information and stuff. What do we so one, one thing to talk about, maybe, is the general rhythm or pattern that the population has. How does, how does one neuron in the population behave with respect to other neurons in the population? This is where we start to talk about things like the correlations between neurons and when you get into the weeds here, this has different implications for how information can be encoded by neurons, but trying to tie this to something that people might have some familiarity with, people might have an intuition that, for example, when you go under anesthesia or you go into non REM sleep, you're unconscious, so there's sort of less going on in terms of conscious experience. The neurons in that state tend to have a pretty rhythmic and correlated pattern of activity. And you see something very different when an animal is awake, and especially when it's awake, and it's like focused and thinking about stuff with respect to populations and the correlations between neurons, how do we start to think about the correlations they exhibit and how that relates to things like the robustness or efficiency of encoding and attention, and just how, how different types of information encoding might be allowed or disallowed with these different types of correlation patterns?

Marius PachitariuYeah, let me, let me start actually, with the example that you gave, because I think it nicely brings in together two ideas we discussed, the sleeping or anesthesia example. Turns out, yes, it's oscillatory activity during sleep a majority of the time, except for REM sleep. But that's actually a lot of action potentials. It's it's not a fundamentally reduced state where, you know, the brain is trying to conserve energy or minimize action potentials. It's more or less the same energy budget when we're sleeping and like, technically not doing anything with our brains. So it's just a patterning difference. It is definitely a patterning difference. The exact role of sleep, as you know, is not well understood, whether it has something to do with kind of cleaning out your brain and clearing bad stuff or things of that sort. But we do know something very interesting during sleep, and that is that we dream, and dreams are kind of the you. Are fundamental generative models in our brains being played forward, if they very much depend on the experiences we're internalizing right over the course of the day, sometimes over many, many years, we can dream of things that happened a very long time ago. So dreams are a great example of something that happens during sleep with patterns of activity that sometimes can look like the ones during wakefulness. Sometimes they do not. Now, you pointed it out as an example of kind of network dynamics operating differently, and that's perhaps one of the, I would say, most clear differences that we can study and think about in terms of network dynamics, because it's so very clear that under anesthesia, especially, and under certain forms of like very Deep Sleep, neurons are very much coordinated. They might fire a lot of action potentials in a burst, then quiet, then burst, then quiet. These are what people call up and down states. There are models of up and down states. We built a model when I was just kind of starting out in my neuroscience career. We built them all of that. There's other models of how up and down states might be produced. It can be a shift, for example, in the amount of inhibition that a network has to play with. We think that more inhibition, kind of stabilized regime prevent some of this kind of runaway excitation that leads to these bursty population events. But ultimately, we think that that's just kind of a way of the core of, well, correlating a network to the point where it's not really able to do the things it's normally supposed to be doing. It's just a way of, kind of impairing the normal neural computations, and perhaps that impairment is what leads to the loss of computation. So

Nick Jikomesat a very high level, the extent to which neurons are firing in a coordinated fashion will determine what they're computationally capable of. Clearly, the neurons, you know, as they're firing noise, they do during anesthesia or deep sleep, they're clearly not capable of doing the things that they do when they're awake.

Marius PachitariuYes, except for dreams, except

Nick Jikomesfor dreams,

Marius Pachitariuyeah, and I was just saying that that's kind of in some ways, with something we would call a one dimensional pattern of activity, if all the nerves go up and down together. This is the whole population having one mode, you know, of coordinated activity. We can project the neural activity onto a single vector direction in neural space and get most of what's going on, most of the activity going on over there.

Nick JikomesBut when there's generative systems are in play during dreaming or during waking, that's not the case.

Marius PachitariuThat's not the case, then you know, this one dimensional mode of like the whole population activating or not, becomes very uninteresting, has very little activity in it. It really becomes this very high dimensional pattern where there's all sorts of groups of neurons that might be activating together sometimes. And really, there's no good way, at least, for this kind of cortical recordings that we do. There's no good way of reducing that to say, oh, you know, not. It's not one dimension, but maybe it's five. Maybe it's like 10 different axes, and that's all the brain is doing. No, we don't see that in any of the scenarios we've looked at. There is not a way to reduce activity of in our case, you know, 10,000 100,000 neurons that we record, we have no way of reducing that to, let's say, just five dimensions. So

Nick Jikomespeople, people have tried and failed to come up with models of the brain and to analyze these high dimensional data sets where you've got many neurons recorded. And people have tried to say, Okay, what if we can explain everything you know? If we can explain the sleeping brain in one dimension, maybe we can explain the waking brain in three or five or 10 and they can't, we can't do it,

Marius PachitariuYep, absolutely. And we can also, we can prove that we can't do it right. As as we record more and more neurons, and we get more and more samples of this kind of neural data, we can develop the kinds of analyzes of the, you know, if you want, of the mathematical properties of this kind of cloud of points that the neural activity inhabits, we can quantify those mathematical properties. And those quantifications can tell us, for example, that you know, there is no linear, for example, set of dimensions that you can project into that span the whole. Base, but also that there's no kind of nonlinear set of dimensions. So nonlinear manifold, which is where it gets more complicated and more mathy, and then you have to think of kind of the local properties of this mathematical object that we call a manifold, because it resembles, you know, the mathematical concept of a manifold. It's kind of a curved space, but linearly. Locally, it's still linear. Locally, it still looks like a low dimensional space. We don't see that in the brain as soon as we record enough neurons, if we record very few neurons, or if we put the brain into a very kind of low dimensional state. For example, I don't know, ask an animal to do the exact same thing 1000 times, yeah, which is what a lot of experiments are actually like, right? Right? Then, yes, it looks like there's very little going on. But as soon as we ask the the animals to do lots of different things, or we show lots of different stimuli with different properties, we see that the brain, kind of the neural activity just jumps around in the space and just bounces into all sorts of different patterns.

Nick JikomesSo it's not just that the brain, brain activity of a waking animal doing interesting things is many dimensional you can actually prove that it's sort of this hyper dimensional thing that can't even in principle, be projected down to a linear combination of, you know, even dozens of dimensions.

Marius PachitariuThat is correct? Yes. And

Nick Jikomestrying to what do you think that means, in very, very simple terms, is that connected to what you said before? Perhaps that you know, this has to do with the idiosyncrasies of evolution and biology, and that, you know, things aren't, things aren't built from a blueprint by an engineer. They're kind of cobbled together by this random walk through evolutionary space over time. Yeah, no,

Marius PachitariuI think that's that's probably a part of it. It's good to think about it in terms of information coming in. Let's say to our eyes, we're studying a visual system, and what's going out. What's going out is actually the output of our system is fairly low dimensional. There's we can really describe what we do on a moment by moment basis with, you know, just a few bits of information, right? However, the amount of information coming in through our very high dimensional sensors, our retinas are some of the best cameras on the planet. Basically, this is extremely high dimensional information. As we walk around, we're bombarded with very high dimensional information, lots of photons and all sorts of configurations that our brains are somehow very well optimized to extract information from and distill it down to these very few bits of information that go out, which constitute our our decisions, right? We we see a familiar face. We say, Hi. That's the output the computation itself has involved, like a very high dimensional process. Now, what's really kind of surprising is even as the data coming into our retinas is so high dimensional, it appears that the way the brain solve this problem is to make it even more high dimensional inside of the brain. So we go from something like a million retinal ganglion cells. I might have my numbers wrong. I know the factor of expansion. I might not know the absolute numbers. There's a lot in the heat. There's a lot. But then the factor of expansion of how many inputs come in through the retina and through the ganglion cells and through a nerve, optic, optic nerve, it expands by a factor of 50 or so inside of our visual cortex. So for every little single pixel, let's say, in our field of view, there's another 50 neurons, kind of somehow encoding that pixel in different ways, let's say, but they're not putting a single pixel, they're encoding patterns of

Nick Jikomespixels. There's more ways to encode the pixels than there are incoming pixels.

Marius PachitariuYeah, and so that's, that's right, another way to say, like, emotionality gets expanded before ultimately it has to collapse, because ultimately, are relatively low damage.

Nick JikomesIn principle, in theory, what would be the possible advantages of doing that, dimensionality, expansion, before going in the opposite direction? Yeah, I mean in principle.

Marius PachitariuSo I mean there is an empirical answer to this, which is that when we try to design artificial. Official neural systems, right? Official neural networks, this kind of expansionality is beneficial, and these systems work better the more neurons they have to so

Nick Jikomesthat's just, we know that's true from that perspective. So we know it's true, even if we can't really explain why.

Marius PachitariuWe have a few you know, actually, it turns out, the theory for explaining why was developed before the neural networks were developed. And the neural networks were not developed with, necessarily with this theory in mind. But this theory is goes back to kind of statistical learning, I believe is the name. No, I may be wrong about this. This is not some of the initial ways of thinking about it is, if you have a kind of a non linear computation, you have to do, what you could do is take your inputs, project them into a high dimensional space, and then do a linear computational in this space. That's much easier to do. Now, there are actual mathematical proofs that this kind of thing works almost always. Some of those proofs are related to neural networks that, say, a neural network with an intermediate width that can go very large, as you know, infinite, can do any computation, as opposed to a perceptron that cannot by itself. But when you put many of them, infinitely many kind of in a row, it can do any comp. So those are the kinds of theories that talk about dimensionality, expansion. In practice, with neural networks and artificial systems, we don't just make them high dimensional. We also make them kind of hierarchical and we make them sequential, yeah. So the input keeps getting processed and processed and processed and processed. So

Nick Jikomesbasically, the idea here, the simple idea here, is that if you do this dimensionality expansion, it can make computations easier to perform, and then you can project them down into a lower dimensionality, and then, you know, execute, ultimately, a motor action,

Marius Pachitariucorrect? Yes, that is the idea that has mathematical support. I would say neural artificial neural networks both validate that, in the sense that they do expand the input into high dimensional space, but in some ways, they also invalidate that, because they also require kind of many repetitions of this computation to happen. Don't see that as much in the human brain. If you look at the human brain, or any mammalian brain, some of the early visual cortical areas are very large, so we've got what's known as v1 the primary cortex, and v2 these alone take up, depends how you define it, but by some definitions, you could say they in The human brain, they take up something like a third of all the visual processing areas, and something like the mouse brain, they take up more than half easily. So then there's a lot of these higher order computations actually happening in smaller and smaller and smaller brain areas. Right, right, right. Make also correspond to the kind of the space contracting that way.

Nick JikomesSo let's start, well, actually, let's go to one more place before we talk about some of your recent experimental results to set this up. So there's a notion you see this, both in the machine learning and AI world. So people may have heard this in some context, even if they're not researchers in this area themselves, but it's the notion of supervised versus unsupervised learning. Let's just start with a basic definition of those

Marius Pachitariutwo things. Yes, these are coming entirely from the machine learning field, where supervised learning is any computation that you set up where you have labels that you're trying to predict from inputs. So it's kind of it's another way of saying a regression problem. People might be very familiar with linear regression. That's the most fundamental supervised learning problem. You've got inputs and outputs. You model them as a linear operation, but you can also predict the outputs from a deep neural network with as many parameters as you want, and that's also supervised learning, because you're predicting labels such as, you know, this image, there's a cat, or there's a dog, or, you know, one of the many labels you could put on an image. And you have a lot, a big data set of such pairs of inputs and outputs, inputs and labels that you train your classifier on, and that's typically described as supervised learning, and it's been the foundation for almost everything in. Machine learning these days. On the other hand, a supervised

Nick Jikomeslearning algorithm would be one where the algorithm is given the things to look for or to do ahead of time. It's given some information and kind of told what to

Marius Pachitariudo, correct Yeah, it's instructed on a kind of a case by case, an example by example, what to do. And if you have a big enough data set, then you will be able to, hopefully know what to do in a new situation, a new input that you have never seen before, still able to produce the right output. On the other hand, unsupervised learning is something that has also been kind of a big idea, and it's only becoming more practical these days, but it's been a big idea for a while, because you realize pretty quickly, when it comes to human brains and animal brains, that the amount of pure supervision that we have is very low. There's very few times that, you know, a parent points to something says, that's a dog, that's a dog, you know, it's a zebra. You know, we get some of this, but it's very low compared to what any of the artificial systems these days get in their training. And instead, it's been thought for quite a while that the brain must also be able to do some kind of unsupervised learning, where it can extract patterns from the environment without necessarily being told what those patterns are. And the reason why it made sense to think about it that way is because our environments are are, in fact, very highly structured. The photons that arrive on our retina are not just random combinations of things. They arrive in, like very specific patterns and being able to abstract away and kind of make certain, you know, understand that certain combinations of photons, certain patterns are much more likely than others. That's a form of unsupervised learning, typically and especially from a pragmatic point of view, let's say you have an artificial system you want to do this in. It is very common these days to do something like unsupervised learning to prepare a system for the eventual supervised learning that you're going to do with a potentially a much smaller set of labeled examples. And when you do it that way, we call the unsupervised learning, and as unsupervised pre training, we're pre training a system to abstract away the kinds of patterns it sees in the data, so that when labels come later, it's able to more easily associate those labels with abstract concepts in the data, rather than with the kind of pixel level properties, which might not be a Very good association.

Nick JikomesOkay, so there's a notion that unsupervised learning can help an organism learn the general, more abstract structure of stuff, and then when it has to do more concrete learning, like, you know, matching pictures together or generating an action when it sees something like in some of the experiments we might describe that what you might call supervised learning, or task based learning, that can happen faster if there's already kind of some scaffold there that emerged in this unsupervised manner.

Marius PachitariuYeah, absolutely. That would be the idea, and it's been a very popular idea with theoreticians for quite a while, there's quite a few algorithms that or models that say that have made predictions of what things would look like in the brain if unsupervised learning was going on, and a lot of those were promising. And so a lot of people in my generation, for example, kind of seeing those kinds of models, we've been very, kind of inspired to, like go and find more directly something of that sort happening in the real thing, right in the real brain and with real neurons. And we think that's what we found. You

Nick Jikomeswant to set up some of these experiments a little bit. What is the general experimental paradigm you're using? What does it look like? What do the mice have to do?

Marius PachitariuYeah. So in this particular paradigm, mice are head fixed on an air floating ball or kind of, kind of a treadmill. Turns out mice are very good runners, just for the same reason that hamsters have, you know, kind of naturally run on hamster wheels. Just about all rodents, they just kind of like to run on wheels. So we have fixed them on wheels, we have them run, and then we pick up that signal that. You know, the spinning of the wheel, and we couple it to a virtual reality setup, which consists of monitors arranged around the mouse. And as the mouse is kind of running and moving forward the ver it moves forward in the virtual reality, and gives it an impression that it's traversing some environment. Now that might seem abstract, and you know, matrix like, the reason we do this is so we can use our fancy, sophisticated, very large microscopes, right that are kind of fixed in place, that requires us to fix the animals in place as well. And then we can come close to the micro with the microscope and record from these very large neural populations, while the animal is kind of having an experience that's, you know, not exactly natural, but it has certain of the same features.

Nick JikomesSo the mouse is literally running on a ball, and it's looking at computer screens that have, what people might think of is kind of like a video game, like environment, usually a maze of some kind, and the screen, and the way it moves and presents pictures to the animal is tied to the animals running its motion on this ball. And so it's very much as if the mouse is actually running through a physical space. And for all intents and purposes, it certainly looks like the mice think they are.

Marius PachitariuYeah, we don't know what the mice think. And we don't think that we are. You know, we give them a little more credit than that. We think they do know however, we think they're processing things in kind of the same way, or like reusing some of the same computational computations in their brains to process this kind of data. I don't think they're imagining that they're not fixed in place anymore. For our purposes, this has a few beneficial properties. It's kind of, it's a departure from some of the more classical experiments. In classical experiments, stimuli are kind of shown statically, or an abstract stimuli shown on a screen in front of an animal. In our case, these stimuli are presented in a way that the animal might experience them if they were running, yeah, it's

Nick Jikomesa dynamic environment. The animal has to use its navigation abilities. Things are, you know, it's the stimuli are tied to the animals motion through this virtual space. So it's, obviously, this is an artificial setup, but it's quasi naturalistic in that sense.

Marius PachitariuYes, and we think it's part of the reason why it can drive the kinds of learning processes that that would happen under naturalistic scenario. And when we started the this type of experiments, what we did is we set up kind of this, these different corridors, different regions of the world. We put different naturalistic pictures, really, they're photographs that I took with my phone, of like, different visual textures around the world, really, not just here. And then we put these on the walls, and we wanted the animal, the mouse, to learn that some visual textures are rewarded, and some are not. Some environments visually are good environments. Some environments are kind of boring. They don't have anything in them. We train the animals to respond to those good environments, the rewarding environments to recognize them and respond. These animals are water restricted, so they're motivated to work for water rewards, and that is how we get them to report to us. You know, whenever they lick, it's a report that they think they're going to be getting water in this environment. And then after training for, you know, anywhere from two weeks to three, four weeks or so, the animals are reporting on, you know, almost perfectly, which environments are good and which environments are bad. And then we record neurons in their

Nick Jikomesbrains. How do they do that? How do they tell you?

Marius PachitariuSo they tell us that, because we don't just deliver water in the good environment, we ask them to lick the lick port for us in order for water to be delivered. So

Nick Jikomesif in the one environment they're reliably licking and in the other environment they're not, this means the animal understands they're in environment a versus environment

Marius PachitariuB. Yeah, exactly. There is no punishment for licking in the wrong environment. So in some ways, the animals could, could just lick everywhere and always get the reward. But it turns out, they're, they're conservative with their licks, and they, you know, when they know they know, yeah, right, yeah. And then they lick, and if they know they're not going to get water, they're not even going to bother, you know, yeah.

Nick JikomesOkay, and so, so what comes next in the story? How do you how do you start to suss out, you know, supervised versus unsupervised types of learning brain

Marius Pachitariuso far this, you know, resembles, you know, other studies that have been done before, including the neural recording part, which we recorded neurons before this process, and we recorded neurons after this process, and we wanted to compare the activities before versus after,

Nick Jikomesand you're basically looking at what we would call the visual parts of cortex, correct

Marius Pachitariuand multiple different visual areas in cortex, in more of a big majority of the brain areas that the mouse has dedicated to visual processing. We're recording from all of them at the same time, and that's where the technology is really, you know, enabling us to be able to do this, rather than going and recording from one area at a time over many days, which would be, you know, kind of difficult, we can record literally all of them at the same time. And so have something like a survey of what different brain areas are computing.

Nick JikomesAnd so what would, what were you guys expect, like, you know, under different hypotheses, in terms of how the visual parts of the brain would be learning about this environment, what is sort of the space of possibilities you expected to see here.

Marius PachitariuYeah. So I think if we, if we had not been paying attention to what other people were doing, you might think, okay, higher order brain areas should be more plastic, you know, and change their activity as the animals learn a visual task. Because why would you change no very low level spatial computations, right? The very low level units in the brain is for the retina, and certainly the case, you kind of expect that they kind of deterministically report the same thing regardless of what's happening. They're low units. And so we might have predicted that only higher order brain areas would change their activity. However, we already knew from other groups that neurons, just about everywhere, change their pattern of activity. Not necessarily did they become more active, but they might become more selective between the different environments. So in the beginning, you might have lots of neurons that respond. For the sake of argument, let's say there's a leafy place in the world and there's a place with circles, because those are the stimuli we used, kind of leaky and circley. And there's at the beginning, there might be lots of neurons that respond to both after learning, a lot of the neurons will be very selective. They'll respond to either one or the other. So this increase in selectivity, sometimes associated with increases in firing rate, is something that had been reported all across the kind of the visual areas, including at the lowest level of the hierarchy. And so we need to expect that it looked like that in our data too. As far as increases of discriminate, the neurons are able to discriminate the stimuli better. But we also found one brain area where this increase was highest, and that's kind of where we we focused our analysis a little bit more. That was indeed a higher order visual area. Now I should say that what I described so far has been kind of how we've been, how we did actually do the project. We didn't, you know, plan out exactly where we would go from here. We just kind of got to this place. We had all the data. We were trying to think through the data, this kind of massive data set, what could it possibly mean? Until we actually realized there's another crucial experiment we could do that would test whether or not these neurons are changing their properties because the mouse is learning a task, or because the mouse is just by itself, internalizing a model of the world. And that's what we call unsupervised learning, in this case, where we went ahead with a new cohort of animals that were not water restricted, we took advantage of the fact that, like I said, mice like to naturally run. They'll just run forward if you if you leave them there. So they ran. They explored these worlds for also something like two weeks uninterrupted, interrupted, you know, an hour a day or two, but

Nick Jikomesthere's nothing for them to really learn, in the sense of, there's no task to do, there's no reward to get. They're behaving. They're moving through the same environment, just like the other mice that wanted to sort of get to the end to get the water. But they had no they had no such thing to learn.

Marius PachitariuYeah, they had no motivation. Other they're they're into. Internal motivation of kind of running forward and exploring and perhaps, you know, make paying attention to what's on the walls. And we think they were paying attention to what's on the walls for kind of two different reasons. The first reason is that when we look at their neurons, their neurons went through exactly the same modifications that the neurons in the task animals went through. So the supervised animals and the unsupervised animals, the neural activity changed in almost the same way. The few exceptions, there's one exception to that, but as far as like the system as a whole, goes the tuning properties of the neurons changed almost exactly the same way. And there's a second reason why we think they're internalizing the model of the world, is that if you take these animals that are just experiencing the world by themselves, and then you ask them to do the same task that the supervised animals were doing. Now these animals can do that task much faster. In fact, they can. Many of them can do that task on the first day, which, when it comes to animal training, is, is pretty fast. How

Nick Jikomesdoes that compare to the animals that had had to learn right away?

Marius PachitariuMost of the animals that that had to learn right away took about a week or so. Okay, the task asked, and I think people might have an experience of this, because, for example, you know, when I try to train my cats to do things that can take a very long time and a lot of repetitions of the same thing, or training a dog to do certain things, it takes a while, right? So these mice we're actually learning in one day, which for animal training, I'm saying, is really fast.

Nick JikomesOkay? And so the basic idea here is, when an animal has an experience with an environment, when it's already familiar with the environment, it's easier for it to learn stuff, new things in that environment,

Marius Pachitariuabsolutely. And we think it's because of these neurons that were forming new abstract representations of the world around it.

Nick JikomesAnd to what extent were those neurons localized in one place or a few places versus distributed? What can you say about that?

Marius PachitariuYeah, so I mean, that's interesting because there was, there was one set of brain areas that we call the more medial regions. They overlap to some extent, with an area called the retrosplenial cortex. It's a part of the navigation system in at least in rodents, it's connected to the hippocampus and other navigation systems that that's the area where we saw entirely new neurons with Visual Properties develop. So there were very few neurons that even responded to images before, but after this unsupervised learning, we saw like a large population emerge that actually had visual response properties. Now this was restricted to this particular set of medial regions, but it wasn't the only change that happened in the system. There were also changes that happen in the primary visual cortex in other higher order visual areas. It's just that these these areas already had neurons tuned to stimuli before that were selective to stimulate before, especially primary visual cortex responds strongly to visual stimuli. It's just that the patterns of responses the tuning changed, and the tuning changed in kind of a gradual way towards towards a place where one and when an animal was experiencing the same kind of repetitive corridor for many, many days, it would start to essentially develop an entirely new dimension of neural space just for that corridor. So the same way we're talking before that. You know, neural computations are extremely high dimensional. It turns out that over time, things that repeat a lot, things that we might form into visual memories, kind of start to stick out from the cloud and develop their own dimension that's orthogonal to all the other dimensions that existed there before. And so this process is kind of it's more one of changing the angle of a population vector rather than introducing a whole new set of neurons, like what we saw in the medial areas.

Nick JikomesYeah. What might the behavioral advantages be of that type of operation? It's just, it's intuitively, you might think, okay, so the brain is sort of making it easier to. Pick out this thing that you're very familiar with and therefore probably have to interact with more, or something like that.

Marius PachitariuYeah, that, I think the gradual nature of the process, where it actually took kind of a couple of weeks for this dimension to kind of really stick out and be completely separated. It's the gradual nature of the process that gives us the insight here, because, in the beginning, let's say animals run through this kind of leafy corridors. There's leaves on the on the walls, if we put them into another corridor that also has leaves, but kind of in different places, different you know, doesn't it looks the same kind of abstractly, but it's really, it's a different location, let's say in this forest, then they still respond in that environment, as if they think that there should be reward in there, like they respond in a similar way. And that we think is arising from the similarity of visual representations, right? So as you're forming a memory of a place, visually, you might be forming, you know, some abstract you know, leafy places are all good. They all give me rewards. So whenever I see a leafy place, I should lick, I'm going to get reward. But then with time, as you start seeing the same place so many times, it gets kind of seared into your brain, right? It's this, and it's, I think it's an experience would be a lot of us have when we recognize something that we've seen a lot, just kind of see a pattern, and we know we've seen this before, right? That is called visual recognition memory. Yes, as visual recognition memory develops, it develops for very specific exemplars. It develops for specific instances of the world, not for like the general concept, not just for leafy places in general that are good. This particular leafy place where, you know, there's a tree to the left and a tree to the right, and some you know, Bucha this place, this is what's good, and that requires a very specific memory for that particular pattern. And we think this kind of, this dimension, kind of sticking out of the cloud, is an instantiation of that memory being formed and being very specific to that place.

Nick JikomesYeah, it's a way to make a place stick out because it's important, because, you know, you get water there, or something good, or something like

Marius Pachitariuthat, yeah. And, and you've seen it so many times that you you can recognize, like, very small details, and know if the small detail has changed,

Nick Jikomesyeah, yeah, yeah. And presumably, the the memory of that location is much higher resolution than than other locations.

Marius PachitariuExactly. It's a high resolution, yeah, version of it inside of your internal model.

Nick JikomesAnd you know, when we talked about supervised versus unsupervised learning, I was starting to think like, okay, obviously we can discriminate between supervised and unsupervised. When you guys create a whole situation where you introduce the labels or the conditions for the reward, like it is supervised, that sense. But actually, when you step back and think about it, especially when you think about it for simpler and simpler organisms, in what sense do they ever get, like something supervised, like almost everything in their life is going to be a form of unsupervised learning? Is that reasonable to think about

Marius Pachitariuit? Gets almost everything in their life will be a form,

Nick Jikomesyeah, in what sense is there supervised learning for like a fruit fly in the real world?

Marius PachitariuSo I think that that's where we bring in something, you know, called reinforcement learning, right? Which is it's often being distinguished from these two. Reinforcement learning is more about, you know, getting reward, getting good outcomes for actions that you take. It's kind of more inherently of a kind of a low dimensional computation, because you only get so few rewards in the lifetime. But then, if you have a good algorithm, you can kind of figure out what led to that good reward. What sequence of actions led to that good reward? How does that relate to the sensory environment?

Nick JikomesAnd I suppose, I suppose in the absence, in the absence of having an external agent, like an experimenter, or, you know, like, like your mother or something, if I think about an animal who's taught by observing its mother, in the absence of having those kinds of labels, I guess, in a sense, you know, if certain sort of reward, rewards are baked into the brain just intrinsically by evolution, those can be thought of as just preordained labels, and so they're supervising the learning in that sense,

Marius Pachitariuabsolutely. And every organism that we've looked at has had a big part of its brain dedicated to a monitoring for those rewards and to kind of associating good things that they that animals do to those rewards. Yeah, a lot of it is the kind of the dopamine system that people talk about a lot. It's a very kind of widespread uh. Signal that gets distributed throughout the brain. It's, if you want, it's the, it's the, it's the learning signal, essentially, for a lot of these, even for the higher dimensional computations, yeah,

Nick Jikomesand, I mean, this is where you would talk about, so, you know, dopamine, it's associated with motivation, and the, you know, obtaining rewards. But this is where, oftentimes, this is, like, the first sort of computation you learn about is the reward prediction error, which is, I mean, if you want to describe it for people, it's actually a powerful but remarkably simple phenomenon. Can you talk about that a little

Marius Pachitariubit? Yes. So it also has to do with the brain making predictions about the world. So let's say you have this complicated machinery we've talked about, we've been talking about so far, that makes a model of the world, and it's able to predict, based on the model of the world, that certain things are good and certain things are bad. And then you you get into a situation, you predict, okay, something good is about to happen. Because I don't know, I walked into the elevator, I pressed the bottom for my floor, and I expect it will go to my floor. I don't know something like that, but then all of a sudden, something unexpected happens, and the elevator just doesn't move. Let's say, then we would say, if we were to record from certain neurons in your brain that are dopamine neurons, we would possibly see a spike, a reward prediction error, a spike of neural activity associated to that moment when we realized that something that was supposed to happen didn't happen. Now, the way we study these in a lot of animals is related to rewards, because that leads to some of the biggest spikes, if an animal expects a reward will happen, like Pablos dog, when the bell rings and the reward doesn't happen, that's when we see some of the biggest such spikes, that are reward prediction errors. And it is thought that this signal is especially good for learning because if your model was so sure that something good was going to happen, but it didn't. It means your model was wrong. Your internal model was wrong. You need to adjust it. To adjust it, you need a signal. The signal is the reward prediction error that then gets distributed, kind of across brain areas and through magic we don't know, is able to change that internal model to say, you know, sometimes there might not be a reward, no, get to get used to it. That's fine. It's normal. It's the new normal. There's

Nick Jikomessome kind of comparison going on between, you know, you've got a model because you're like, I have an expectation of what things should look like in the world, and then they don't look that like. They don't look that way. So the expectation is violated. It it's some kind of like subtraction, like operation,

Marius Pachitariuyeah, and that happens in advance, right? The Model predicting reward is something that is a recognition of a situation where, if you just do this, or if you just do that, you might be able to get a candy bar or the machine, right, and then it's when it that has already been distilled into the expectation, something good is going to happen. If it doesn't, candy bar does not come out. We have a prediction here. We have to go back in time figure out what parts of the prediction were wrong. But that, you know, this is not actually maybe a machine dispensing candy bars, or maybe I didn't put in the right amount of money, or maybe it's out of candy bars. And for this particular thing, I usually get so we have to still figure out which of those parts of the model was wrong. And with you know, then correct accordingly, that's where it gets complicated, right? This, this initial part of a reward prediction error distributed to two different brain areas is by no means a simple thing, but then actually associating it with the internal model is where it can get really hairy and

Nick Jikomesa little bit of time that we have remaining. I just want to ask you some general questions to get your perspective as a neuroscientist and as a computational guy on things like the large language models and the AI systems we build in the digital artifacts that humans have invented, high level, just your general take on this stuff. When you look at the large language models that are out there, you know, people are using chatgpt Every day, and all of the other llms that are out there, they can obviously do very impressive things. But to what extent are they, how would I even say it? To what extent are they actually reasoning, or do you think that they're implementing the types of algorithms that biological systems like brains are doing? Can you just kind of compare and contrast, or maybe tell us if you think something, is there something sort of fundamental that we think brains are doing, that the llms are not, or vice versa?

Marius PachitariuYeah. I mean. That, that there's all sorts of things to talk about there. I can give you more of my my perspective on it, rather than than reflecting a more general perspective. I mean, the more general perspective in neuroscience, right? Is that, for better or for worse, right? There's something more special about how brain computations work that we haven't really captured the magic, the magic of reasoning, into the artificial machines yet, right? And then the critical part of that becomes the yet, because some people think you can scale the current systems to a point where they do become looking like more like reasoning, even if, currently they don't do that. Now my perspective on it is that humans and the kind of reasoning we have has been arrived at through evolution to some particular idiosyncratic version of reasoning that we have where our brains are it, I don't think our brains are like doing it in a in the simple, like correct way. I think they're doing it in this very convoluted way where it just kind of happened through evolution to get us to this point, I think it'll be very difficult to replicate that in a machine, because it's just it's reverse engineering our brains, which is what we do in neuroscience, but very converted, very like strongly connected network, it's very hard to reverse engineer. But that doesn't necessarily mean that whatever version of reasoning an artificial system gets to isn't reasoning. It just might be a conversion of reasoning. It might not be what we do, but you know, as far as like outcomes go, we can still argue whether the outcomes are, you know, the the responses that llms Give us are good or bad, and that's really most of what we argue about, right? The response is good or bad. Now I think one part of that though, that I think ties more closely to my research, that I think is interesting to mention, is that the way llms are trained is technically something people might call self supervised learning, in the sense that llms are not well, they're trained in multiple stages. The main stage, the more like self supervised stage, is the one that takes the longest, and it's technically not done with supervised data. Llms do not have, well, they they are fine tuned on relatively big data of questions and answers. But that's not all they've been trained on. That's just a fine tuning step. It's a relatively small part of how these llms are trained, because they're initially just trained to predict the next word in every sentence, right? And that is what we would call a self supervised training, because it we can just give it books and websites and emails and everything, and we don't need to label it specifically to say, you know, this is a particular form of communication, or it can just learn by predicting the next token, the next word in the sentence, yeah. And the reason why that's neither supervised nor unsupervised is we don't give it specific labels, but it makes its own labels. In a sense, to it's still using the same kind of supervised learning algorithms, the deep neural networks, that take an input and produce an output, but it's making its own labels of what is it going to predict. It's making its own task of predicting this next word. And that's where I think a lot of the kind of frustration with llms comes in, because, in principle, predicting the next word, if you can do it really, really well, can be a mechanism to generate entirely long sentences. Instead of predicting the word, you produce the word and you produce the next word and produce next word. That's how LMS work. But then, as far as like researchers, especially those of us thinking about the brain, go, that is certainly not how. You know, humans are trained to reason by predicting the next word, right? We reason across big, much bigger future timescales. Also we have pretty long timescales into the past, but llms have that as well. What llms Do not have easily right now is the ability to kind of do a lot of kind of forward modeling, and like a pick, the pick, the best version out of those, and the best specialized models that we hear about that do some of the most impressive tasks, like solving very complicated. Math problems, like international math problems and things of that sort. Those are the ones that do have some of this forward reasoning. And yeah, I mean, that's, I think that's kind of my perspective on it's just, it's another computational system. It doesn't have to be the same as what humans do. It can be studied for kind of what it is, and it will be very interesting to see just how you know, far these kinds of systems can get to for us neuroscientists, I think we mostly think of them as like another interesting system to study,

Nick Jikomesyeah, yeah, it's another system. Just like you're studying multiple systems in biology, you've got different species, different strains. Obviously, in the realm of biology, you've got, you know, evolution by common descent, so they're all related to each other. You know, this is almost just like an alien branch of of we might think of as a phylogeny and and so therefore it might just have more, more distinctive properties. It might work in different ways, because it has like this independent origin,

Marius Pachitariuyeah, exactly. And that that kind of contrast can often lead to kind of realizations about how the systems work, yes, yeah, for certain properties and what's not

Nick Jikomesgiven how much of the brain is about the body, in a sense, right? You've got to move this body through space to survive and and propagate. Much of the brain is about, it's about, you know, learning things about the environment insofar as they can inform patterns of muscular contraction that the neurons are generating the llms don't have that. They don't have bodies and muscles like that. So how does the sort of embodiment or disembodiment play into this? Do you think that's an important thing in terms of what these systems will and will not be able to do? I

Marius Pachitariuthink my perspective on this is that the embodiment problem is certainly a difficult one, and one where humans are in animals, that that's where we still are much better than anything else, right? We like so far, we've been able to even a few years ago, right? We would say that it's our cognitive abilities that, like, set us apart. But now we have llms that can mimic some of those things that we didn't even think was possible before. Our motor abilities are still, you know, far from being reached by robots or any kind of, yeah, and artificial systems. But I think a lot of that has to do also with our biology, with how efficiently our cells are working. Some of those systems, you know, principles you started with, right the efficiency of them, they tend to apply quite a bit more to other parts of our body that have been kind of optimize through even much longer evolution before that. And so while I do think that this is a hard problem, that's not, you know, it's it needs a lot of work to be solved on the robot side, making robots with, you know, llms that are sufficiently embodied, I personally do not think that is what sets us as humans apart. I think it is those internal models and our ability to reason. We can do very sophisticated things without moving a finger, just by thinking people can, you know, write entire novels about continents they haven't been to, just from their, you know, Paris, apartment or something like that. And I think, you know, a lot, a lot of those kinds of properties are things where we have built our societies around and we have built, you know, it's why the internet is such a drain on our attention. Is because it connects us to a lot of knowledge and a lot of new things from all over the world. It's, I don't think it is our, you know, muscular systems that are going to be where, you know, llms are suddenly going to become conscious because they are embodied. No, I don't think so

Nick Jikomesinteresting. Well, I think it's a great place to leave it. Your work is fascinating. Marius, I want to thank you for your time. I encourage people to go check out your work and and work sort of in the realm that you play in, if they're interested in, in learning and biological systems, but also how some of these computational principles apply, and some of the crosstalk between the ML world and and the biology world. Is there any anything you want to leave people with, any final thoughts or things you want to reiterate from our discussion? In today.

Marius PachitariuWell, I want to thank you for inviting me again. It's, it's great to be on your on your podcast. I think, you know, the only thing I would iterate is that not the specifics of what we talked about, but rather, I think the the fascination around studying this system is real, you can it can be very exciting, especially if you're a student, if you're someone who's like me, you know what, 1520, years ago, looking for, you know, some kind of interesting field, right? Because I had these mathematical skills that I felt like it wasn't that useful. I, you know, solving math problems, the kinds of pure theory math problems that math, you know, professors in mathematics, solve, isn't directly impactful into the world right now, but you can take a lot of those same kinds of skills and you can apply them to these fascinating systems and study them, whether you're trying to understand how brains work or how machines work, it's, it's, it's, it's these complex systems that kind of have fascinating us for so long that I think are really worth getting into, because they can keep you motivated for very long periods of time, for many years and probably for entire lifetimes and careers.

Nick Jikomes: If you like to read like I do, then you should check out the read wise app. It's totally free to download, and if you go through the link in the episode description or on my support page, you can get a free 60 day trial for zero cost. Readwise is a really cool app that allows you to import digital highlights from over 30 different platforms, such as Kindle, iBooks, PDFs and more. It makes it really easy for you to review and find and share excerpts and quotes from whatever it is that you're reading, whether it's PDFs of scientific papers or books that you're reading on your Kindle. And you can even save and use highlights from physical books. I read a lot of physical books before bed, for example, or when I'm traveling. And the readwise app actually integrates with the camera on your phone so that you can take a photo of a page that you're reading, highlight a specific chunk of text from that page and export it to the app so that you can reference it or find it or share it later. They even allow you to export really cool looking graphics of some of the highlights that you have in the books or the papers you're reading, so that you can share them in an engaging way. So check out the readwise app today. If you are a reader and you like to reference and organize what you read, go through the link in the episode description or on my support page or in my free weekly newsletter@mindandmatter.substack.com and try the readwise app, 60 days free today. Transcribed by https//otter.ai

Open the transcript file

Hosts and Guests

Nobody is credited on this episode yet.

Credit the people on this episode

Be the first to rate this episode

More episodes

All episodes
  1. Seed Oils, Omega-3s & Lessons From 93 Years of Life | Artemis Simopoulos | 313

    Nov 17, 20261 hr 35 minAudio

    Send us Fan Mail At 93, Artemis Simopoulos explains why oils once used in paint now dominate the food supply, how a high omega-6 intake slows the making of EPA and DHA, and why cholesterol-lowering corn oil did not...

    Credit the people on this episode

  2. Tendons, Collagen, mTOR & Exercise Adaptation | Keith Baar | Episode 312

    Nov 10, 20262 hrAudio

    Send us Fan Mail Keith Baar explains why tendons stop adapting after 10 minutes of loading, why long low-load holds beat fast reps for connective tissue, and how vitamin C, collagen, caffeine and mTOR shape training....

    Credit the people on this episode

  3. GDF15, Morning Sickness & Mitochondrial Stress | Stephen O'Rahilly | Episode 311

    Nov 3, 20261 hr 29 minAudio

    Send us Fan Mail Dr. Stephen O'Rahilly explains GDF15, a stress hormone that makes us feel sick and eat less, and why morning sickness, and its severe form hyperemesis gravidarum (HG), comes from GDF15 made by the...

    Credit the people on this episode

  4. Psilocybin vs. SSRIs for Depression, Emotional Blunting & How Psychedelics Work | Robin Carhart-Harris | 310

    Oct 26, 20261 hr 56 minAudio

    Send us Fan Mail Robin Carhart-Harris explains why his trial of psilocybin therapy vs. the SSRI escitalopram for depression missed one primary measure but favored psilocybin on almost everything else, and why SSRIs can...

    Credit the people on this episode

  5. This Animal Survives 18 Minutes Without Oxygen | Gary Lewin | 309

    Oct 12, 20261 hr 40 minAudio

    Send us Fan Mail Nick talks with Prof. Gary Lewin (Max DelbrĂĽck Center, Berlin) about naked mole-rats, which in lab tests survived 18 minutes in pure nitrogen by shutting down brain activity, keeping the heart beating,...

    Credit the people on this episode