Luke Rosedahl

PhD Student in Dynamical Neuroscience

It was A!

Why? Because the stripes were a certain thickness and angle.

What could we learn from that? Read below!

The disc you just made a category judgment for is called a Gabor Disk and is used to study human learning. Now, you might be thinking: but I don't see discs like that in real life, and you would be right. In normal life, we don't often have to categorize gabor disks. So why use them? For that very reason: because we don't see them often, we have no pre-existing notions of categories for them. So any behavior we see in the lab is likely due to the experiment; similar to how you randomly guessed which category the disk on this page belonged to, subjects in the lab start off randomly guessing.

So why were you correct if you said the disk was category A? What made it an A disk? The answer to that depends on how the experimenter set up the categories, but the short answer is that the thickness of the black and white stripes and the angle of the stripes determines what category it belongs to.

Keep reading to learn more about the different kinds of category structures or click here to return to my home page!

Different Category Types

One kind of category structure, which is relatively easy to learn, is created by setting a simple rule like "all disks with bars thicker than X are in A." This kind of category is learned through Rule-Based learning, because it has a simple rule that humans can understand. If you were shown a lot of disks separated like that, you would quickly realize that the thick bars are A and would categorize them correctly.

There is a second, more complex kind of category structure. This structure involves rules that involve both thickness and angle and do not have simple category bounds. An example of this would be saying "if the thickness is greater than the angle, it is category A." Because thickness and angle are separate things with different units and the bound is not simple, humans do not understand this type of rule.

Well, if we can't understand that type of rule, what is the point of the category? Good question! It turns out that even though humans cannot consciously understand the rule, they can still learn which disks belong in category A. They achieve this through a neural system that keeps track of aspects of each disk that determine which category it belongs in. This kind of learning we call 'Information-Integration' learning.

Information Integration Learning

Information Integration categories are learned through reinforcement learning. Basically every disk you see triggers a certain subset of visual neurons which connect (synapse on to) to decision neurons. If the decision neuron for the A category has greater activity than the B category neuron, you select the A category. If that was correct, the synapses between the most active visual neurons (those that responded to that disk) and neuron A are strengthened. So next time you see that disk, the A decision neuron gets greater input from the visual neurons and responds stronger. Over time, this strengthening of synapses results in the decision neuron knowing which visual neurons are more active for its category and weighing those more heavily.

The neural systems involved are pretty complicated, so let's look at a simplified version and use a relatable metaphor: baseball. Imagine you are playing a game of baseball with your friends. You have played enough baseball to know whether to swing or not for regular pitches, but have not encountered any fancy pitches (knuckleballs, etc.). For each pitch, you make a decision to stay or swing, and the brain uses the feedback you receive to make better decisions in the future.

Over the course of the game, you receive 4 pitches. The first two, a pitch out of the strike box and a pitch over the plate, you know what to do with because you have played before. The third pitch however, you have never seen before. Look at the visuals for the different pitches and read the descriptions below to see what happens in your brain as the game progresses! The right column shows an example image that might be used in a laboratory setting to study this learning, so you can see how category learning studies relate to real life!

The pitcher throws a pitch that is not headed over the plate. Visual neuron 3 (Blue, on the right) responds to that particular trajectory of the ball and excites the 'Stay' decision neuron, so you don't swing. That was the correct decision, so the umpire calls the ball out (positive reward) and there is dopamine reinforced feedback for the Blue-Stay neuron pathway, making it stronger.

The pitcher throws a pitch that is headed over the plate. Visual neuron 1 (Red, on the left) responds to that particular trajectory of the ball and excites the 'Swing' decision neuron, so you swing and hit the ball! That was the correct decision, so there is dopamine reinforced feedback for the Red-Swing neuron pathway, making it stronger.

The pitcher throws a knuckleball (the first you have ever seen). Visual neuron 2 (Green, middle) responds, but the system doesn't know what that means and it excites both Swing and Stay neurons. The decision is randomly selected as Stay, which is incorrect and results in a strike (negative reward). As the reward is negative, there is negative dopamine feedback for the Green-Stay pathway, making it weaker.

The pitcher watches you make the incorrect decision and decides to throw another knuckleball. This time, because the feedback weakened the Green-Stay pathway, the Green neuron excites the Swing neuron, so you swing and hit the ball (positive reward). As the reward was positive, there is positive dopamine feedback for the Green-Swing pathway, making it stronger.

The game is over and you won! Over the course of the game, the Red-Swing, Blue-Stay, and Green-Swing pathways were strengthened, so you learned when to swing, when to stay, and what to do for a knuckleball!

This is a simplified version of how Information-Integration categories are learned. Synaptic strengthening of the connections between visual neurons and decision neurons result in the decision neurons weighing the visual neurons which respond to its category more strongly, so over time the brain can learn to categorize the disks even though it doesn't understand the actual rule.

Now you know the two types of categories (Rule-Based and Information-Integration), how they are different (Rule-Based has a rule humans can understand), and how humans can still learn Information-Integration tasks (similar to how we learn when to swing at a baseball). So how do we use this information?

We use this information to study the effects of a variety of conditions (such as the delay before being given feedback, distractions with other tasks, the accuracy of the feedback, where in the visual field the disk is presented, etc.) on learning. The results of such experiments teach us more about how humans learn and what factors can positively (or negatively!) impact learning.

Interested in more specifics? Check out my Current Research page! Otherwise, return to my home page to read more about what I do as a PhD student!

Live Free!