How Your Brain Organizes Information

Perhaps one of the most fascinating abilities of the mammalian brain is that it can generate flexible behavior that is generalized across contexts. For example, let's imagine you spent a few weeks learning how to cook just the perfect lasagna. You learn that skill in your own kitchen, where you are familiar with the location of all the utensils, ingredients and appliances as well as the layout of the kitchen itself.

Finally, you get invited to your friend's house to cook your famous lasagna. Even though you have never used their kitchen before, you are still able to navigate and cook efficiently. But how is it possible?

Because if you think about it, you have never been in that particular situation, so this must be a totally new problem, right? Well, the answer lies in the ability of your brain to generalize. You're able to kind of strip away the particular sensory context of your kitchen and extract the abstract notion of lasagna cooking procedure.

At the same time, you know the general principles of how kitchens work, so one quick glance at the location of the oven and the utensils is enough to relate that particular layout to your inner model of kitchens. This generalization requires information about the world. to be organized into a multi-purpose coherent framework known as a <u>cognitive map</u>.

This video is the first in a two-part series. Today, we will lay the foundation for the problem of cognitive mapping introduce some general ideas and look at the experimental data from modern neuroscience. And in the second part, we'll go through a particular computational solution to the problem of cognitive mapping, known as the <b><u>Tolman-Eichenbaum</b></u> machine which might be the stepping stone on the path towards the first artificial<i> hippocampus</i>.

If you're interested, stay tuned. Let's suppose you put a rat in a maze that looks like this and train it to find its way to the end of the maze to a tasty reward. After a few of these training trials in a familiar maze <i>A</i> you put a rat in a new maze <i>B</i> which has several radial arms like so.

Normally, the rat would run to the familiar top path, but what would happen if you blocked it? Now the animal is faced with a choice. "<i>Which alternative path to pursue instead?

</i>" Let's think about possible scenarios. One possibility is that rats learn by associations which means that they would take a path that was most similar to the one that originally led to food, either path <i>NINE</i> or <i>TEN</i>. If, however, animals had some sort of internal map of the spatial layout, they would choose path <i>SIX</i> which pointed in the direction of food but which had never been experienced before, so it's not directly associated with reward.

And indeed, this is what's usually observed in the real life. This is one example of an experiment conducted in 1930's by an American psychologist, <i>Edward Tolman</i> which led him to coin the term <i><u>"Cognitive map"</i></u> relating to this idea that animals have something like a mental map of surrounding space. However, it wasn't until 40 years later that neuroscientists were able to peek inside the brain to see how such a map is manifested in neural activity.

Although cognitive maps serve as general purpose representations of the world as we'll see further, historically, they were usually studied in the context of spatial behavior. The predominant view is that the workhorse of cognitive mapping in the mammalian brain is the <i>Hippocampal formation</i> which includes the <u><i>hippocampus</u></i> a seahorse shaped structure and <u><i>Entorhinal cortex</u></i> which serves as a kind of gateway through which information flows in and out of the hippocampus. What's fascinating about these structures is the existence of spatially selective neurons which paved the way to understanding how knowledge is organized into a map like structure at the level of single cells.

Let's briefly review the major types of such map building neurons. The hallmark of hippocampus are <u><i>place cells</u></i>, neurons that become active when an animal is in a specific location or place in an environment. They are thought to code for the current location in a context dependent way since the response of a single cell is totally different in different surroundings.

Upstream to the place cells are <i><u>grid cells</i></u> that are located in the entorhinal cortex and fire in regular periodic patterns. arranged on a hexagonal grid as the animal moves in the environment. Notably, this regular pattern is quite stable and is mostly invariant to a particular surrounding.

Entorhinal cortex also has what are called "<i><u>Object-vector cells</i></u>" that activate whenever the animal is at a certain distance and certain direction away from any object in the environment while hippocampus has neurons that are similar to object-vector cells but which respond selectively to a specific object and not others known as <u><i>landmark cells</u></i>. Both regions also have <i><u>boundary cells</i></u> coding for the presence of boundaries head direction cells that code for the direction the animal is facing, and a few other cell types. As you can see, there is a whole zoo of cellular responses in the hippocampal formation and it might seem unclear how do we even make sense of all this mess.

An overall pattern can be distinguished, however. From all these examples you can see that the entorhinal cortex provides a kind of a general coordinate system allowing the brain to perform vector computations and estimate distances. Hippocampus, on the other hand, forms a more specific code, providing the brain with information about particular location and landmarks in this coordinate system.

Although such neurons historically have been discovered during spatial behavior, what's crucial is that this selectivity is not restricted to physical space. For example, if you train a rat to press a lever adjusting the frequency of the sound, you'll see that certain neurons in the hippocampus become selective to a particular frequency range like a conventional place cell but in the one-dimensional space of sound frequencies. Neurons in the entorhinal cortex also develop a frequency dependent pattern of activity but which is periodic resembling grid cells that were squished to the one dimension of the frequency space.

In another study, human subjects were trained to navigate in a highly abstract two-dimensional space of bird silhouettes characterized by leg and neck lengths. Participants could independently vary the length of legs and neck with the controller and their task was to morph a bird into a particular configuration, while activity of their brains was monitored through <i>f</i><b>MRI</b>. Remarkably, activity of their entorhinal cortex showed signs of a hexagonal symmetry as people mentally moved in this conceptual space of birds which is incredibly consistent with the grid cell code.

All of this suggests that the hippocampus and entorhinal cortex in tandem might indeed construct a multi-dimensional and multi-modal representation of the world. But how can it be that the same machinery is used to solve computational tasks in both spatial and non-spatial domains? Using what seems like similar algorithms, there must be a way we can conceptually unify them into a single type of problem.

Luckily, there is a simple and elegant mathematical formalism that connects physical and abstract spaces known as <i><u>graph theor</i></u>y. Notice that elements in all the tasks we have seen so far have a notion of being connected by a certain relation. For instance, neighboring locations in a room are physically connected to each other and you can move along them in horizontal and vertical directions.

Similarly, members of a family fall onto a tree-like structure with corresponding relations such as parent, sibling, etc. . These type of problems are naturally described by a graph.

Essentially, a graph is a mathematical structure that consists of a set of points called vertices or <u>nodes</u> and a set of lines called edges that connect pairs of vertices. Now the vertices can represent any kind of object or entity and the edges can correspond to any kind of connection or relation between the vertices. For instance, we can construct a graph of a two-dimensional space space by connecting each location node to its four neighbors in a square grid-like manner.

To effectively perform tasks, however, it is essential that you know where you are located on this graph at every point in time. Otherwise, there is no point in having organized knowledge in the first place. To keep track of where you are, you can use what's called <i>path integration</i>.

In the physical space, it refers to using self-motion cues such as movement speed and direction to accumulate movement vectors and update your position. For example, you know that taking a sequence of steps in the direction of north, then west, then south, then east will take you back to where you started. Many animal species such as insects, birds, rodents, and humans all can path integrate.

In fact, as we'll see in more detail in the next video, networks of grid cells can perform path which can be modeled in a neural network with what's called <i>attractor dynamics</i>. For arbitrary graphs, however, you'll need a more abstract but similar notion of path integration that is a finite set of rules of how to add different types of relationships. For example, that taking the root sibling parent is equivalent to parent, and so forth.

And the very same graph can be used for generalization. Notice that the task of navigating in a two-dimensional space of birds Is essentially the same as walking around the room. The underlying structure of connections is fundamentally the same.

What's different is the type of incoming sensory cues. Similarly, if you have an internal model of a family tree, not only you can use it for different families, but you can also use it to apply to things like taxonomy since it also has a tree-like structure. Just relabel the relationships and you're good to go.

So in conclusion, if the hippocampal entorhinal system can construct these relational graphs and carry out path integration and root finding on them, then it makes perfect sense how this system could be reused by the brain across different modalities. But how do we know which graphs to build in the first place? How can the brain come up with such a structured representation?

Well, to understand this, we need to first address another super important concept and that is the idea of a <i>latent space</i>. As it follows from the name, latent space is something that is not directly observable from external cues. For example, let's view the world through the eyes of a mouse, which is performing an alternation task.

This simply means that it is running in a T-shaped maze, and each time it reaches this split, a choice whether to turn left or right needs to be made. Now, the task is such that the reward is always alternating between the sides. So the animal has to learn that if on previous trial, it turned left and received a tasty treat that means that now it should go to the right and vice versa.

Let's try to think about the type of a mental model, a cognitive map that must be built in the brain of this mouse. In other words, what are the relevant behavioral variables and their structure required to perform this particular task? Well, first of all, there is the spatial component since you need to know where you are in this "T-maze" physically and update your position.

So we can reasonably expect to see conventional place cells when looking inside the hippocampus. But notice that information about physical location alone is not enough to solve the problem of obtaining the maximum amount of reward. Since the animal supposedly knows that it needs to alternate turns at every trial, it has to remember the direction of the previous turn.

In other words, there is an additional abstract variable encoding where you are going and where you have been. To completely capture all the relevant information about this task, you need a configuration of a cognitive map which keeps track of both the location in the physical 2D space of the maze and a binary location in this abstract space of left and right trials. Here is another way to think about it.

At first, when the animal hasn't learned the nature of this task and is just exploring the maze it's cognitive map only has a spatial component encoding the 2D location. But over the course of several trials, the mouse learns that, "<i>Aha! I need to alternate the directions</i> <i>because the reward seems to be always located on the arm</i> <i>opposite of the previous trial.

</i>" When this happens, the cognitive map is expanded with a new dimension and the mental representation of the T-maze kind of splits into two cloned versions of it one for turn left trials and another one for right turns, and now all of a sudden you need to update your position in this expanded space which now has an additional axis. Remarkably, we find cells whose firing is modulated by both the physical location and the direction of the future turn. Such neurons were termed "<u><i>Splitter cells</u></i>".

They uniquely encode the location in the full expanded version of the cognitive map, which is remarkably consistent with this idea that the hippocampus keeps track of all task relevant variables no matter how abstract they are. This split dimension into left versus right trials is an example of a latent space since it is not directly observable from the sensory cues. There is no light switch that would signal you where to turn.

Instead, you need to infer your location in the latent space based on previous observations. Another example includes training the animals to run in a virtual reality and perform what's known as tower accumulation task. Essentially, as they run on the track, they are presented with visual cues on both sides and at the end of the track, they need to choose whether to turn left or right.

The direction of the reward is indicated by which side had the higher number of towers. So for example, if you encounter in total nine towers to your left and only seven to your right, it indicates that you need to turn left to get a reward, since it has a higher number of queues. This number of towers, or better yet the difference between the two sides forms a <i>latent evidence space</i>, and it turns out that there are hippocampal neurons which form place fields in this latent space.

As you can see, location in latent spaces and even their pure existence in the cognitive map is extremely important. But the problem is, it is not directly observable from individual sensory cues. Instead, latent spaces are built from sequences of sensory observations.

For example, remembering your previous choice can affect the future one. But why bother with building such relational graphs in the first place? In the beginning of this video, we saw an example of how knowledge acquired in one setting can be successfully generalized to other contexts.

This is because once the structured representation is built, it can be abstracted away from particular sensory observations. You know that '<i>north-west-south-east</i>' will close a loop in any environment, be it your room or a local park and such generalization effectively requires an existence of a <i><u>factorized representation</i></u>. Here is what factorization means intuitively.

Let's imagine you are buying a car and would like to predict how much you will enjoy driving it. For simplicity, the space of all possible cars is formed by the particular model and its color, and your goal is to maximize the value of your subjective happiness, <i><b>Gamma</i></b> which is a function of two variables car model and color. There is some hidden distribution of gamma but you don't know it since it would require driving every possible combination of models and colors to find that distribution.

However, it makes sense to separately learn the distribution of happiness along the color axis for example, by driving a Tesla of different colors and the distribution of happiness along the car axis by driving each model in red configuration. If these two distributions are independent, for example, if you prefer blue over yellow, and that is the case irrespectively of the car model, then the underlying two-dimensional distribution is said to be factorizable. Mathematically, we can describe the joint distribution to be a product of the two corresponding one-dimensional, also called <i>marginal distributions</i>.

It is enough for you to know the two factors separately to predict the joint distribution and that means that such factorization allows you to predict the value of gamma for combinations you have never encountered before. Also notice that since you only need to know a pair of one-dimensional distributions, it is much less information compared to storing the two-dimensional distribution of all possible combinations and for the brain that problem of information storage is even more relevant since cognitive maps have a much larger number of dimensions. Similarly, in the brain, there is factorization of every experience into its <u>Structural component</u> which is the position on this relational graph, and <u>Sensory component</u>, which defines the particular setting of the outside world.

In fact, evidence for this factorization can be seen in the responses of individual neurons in the hippocampal formation. Recall that grid cells in the entorhinal cortex and in particular, the medial portion of it, remain stable across environments. Forming this coordinate system, information about this structural basis is then fed into the hippocampus.

A second stream of information containing purely sensory cues is provided to the hippocampus by a different part of entorhinal cortex lateral entorhinal area which doesn't have such prominent grid patterns of activity. The hippocampus then forms a conjunctive representation. unifying the two streams of information and embedding the particular sensory information into the structural backbone.

This difference between factorized and conjunctive representations can be demonstrated by a famous phenomenon called <u><i>hippocampal remapping</u></i>. Essentially, it refers to the tendency of place cells to change their firing patterns in different sensory contexts. For example, one neuron might activate when the animal is in the middle of a room.

But after you change the order by spraying with something minty that very place cell might shift its preferred location or stop firing in that environment altogether. Notice that this is completely different from grid cells, which are largely invariant to such contextual perturbations and only care about space. All right, let's try to tie everything together.

In this video, we saw evidence both on the behavioral and cellular level that the brain must have an internal model of the world known as a cognitive map. Despite the word '<i>map</i>' in the name it is not restricted to representing physical space. Rather, cognitive maps are a systematic way to organize knowledge in some kind of structure.

The main purpose of such representations is to effectively utilize the inherent regularities and repetitions of the outside world order to minimize computational effort and generalize your knowledge. For example, transitive rules such as that "if A is bigger than B and B is bigger than C" then "A must be bigger than C" apply in a variety of different scenarios from distances and sizes of predators to sound frequencies and paychecks. So there is no need to learn that transitive structure every time.

You can just reuse it. We also talked about how such structural backbones can be viewed as organizing knowledge as a relational graph which needs to take into account latent spaces that are not directly observable but rather need to be inferred from sequences of sensory observations. Incorporation of latent spaces allows the hippocampus to keep track of abstract variables, such as the amount of sensory evidence in the tower accumulation task and the position in the space of left versus right trials in the automation task.

And finally, we discussed why factorizing knowledge into structural and sensory components is useful to the brain since it allows it to generalize and make the problems more computationally feasible. Evidence for such factorized representations can be found in the entorhinal cortex whose medial and lateral parts provide the hippocampus with the two separate streams of information while the hippocampus then generates a unified, conjoined representation, embedding the structure into sensory context in order to solve particular behavioral tasks from running in mazes and pressing sound frequency levers to cooking a lasagna in your friend's kitchen. Now I know this was a lot to take in and there are a lot of missing pieces.

The thing is, what is described in this video is a territory of hard debates and ongoing research so we are still lacking a unified interpretation. Hopefully, things will become more clear in the second part of the video series where we will take the concepts and experimental observations we have seen here and try to build a computational system that would be able to generalize and learn latent spaces. In the meantime, if you're interested in diving deeper into the topics related to computational neuroscience you should definitely check out our today's sponsor, <b><u>Brilliant.

org</b></u> Brilliant is the perfect platform to expand your knowledge in STEM fields and gain new skills. Unlike traditional textbook-based learning, Brilliant's courses are designed to be highly engaging and interactive making learning an enjoyable and rewarding experience. Hands-on exercises and stunning visualizations allow you to develop a deep understanding of the underlying concepts in the context of real-world applications embedding problem-solving skills into your cognitive map.

Brilliant offers a wide variety of courses for every level and background from fundamentals of math and computer science to neural networks and differential equations with new lessons being added every month. For example, if you're interested in computational models of the brain, check out their crash course on reinforcement learning, covering the foundations of this framework where an artificial agent is trained by trial and error. Don't hesitate to dive into Brilliant and begin expanding your knowledge in bite-sized chunks.

Go to <u>"brilliant. org/ArtemKirsanov/"</u> to get a 30-day free trial of everything Brilliant has to offer and the first two hundred people to use this link will get a twenty percent discount on the premium membership. If you enjoyed this video, press the like button, share it with your friends & colleagues, subscribe to the channel if you haven't already and consider supporting me on Patreon to suggest video topics and enjoy the bonus content.

Stay tuned for more interesting topics coming up. Goodbye, and thank you for the interest in the brain.