This is one of a series of entries outlining some of the ideas in my dissertation work.

Last entry I talked about how one theory proposes that the cortical minicolumn, a cylindrical structure of about 100 neurons, is the primary functional unit of the neocortex, and how its main functions are to learn patterns through association, pass inferences up the cortical hierarchy, and push predictions down the hierarchy.

Today I'll be talking about this last part, prediction. Basically, if each minicolumn has learned a pattern, it can signal to higher parts of the hierarchy that it is currently detecting that pattern. But it can also make an educated guess about what's coming next, and prepare the minicolumns lower in the hierarchy for incoming data.

What would be the use in doing that? Well, Jeff Hawkins thinks prediction is the key feature of intelligent systems, and he gives many examples. I'll use some of his, and some others he doesn't.

At the Cognitive Science Society meeting this summer, Jeff Elman's keynote address talked a lot about prediction in language. People are faster reading sentences with predictable outcomes, such as "The lifeguard saved the girl" rather than "The lifeguard saved some money". It turns out you are constantly trying to predict what comes next, in language, in vision, in hearing, in touch. When the unexpected unicorns you have a moment of minor disorientation as the bottom-up input and the top-down prediction are at odds with one another.

Basically, the predictions streaming downward in your neocortex help speed up things like language comprehension, and help you resolve missing or noisy input, like when you're talking on your cell phone with a crappy connection. In visual processing, prediction helps you determine the likely path of moving objects, as in driving or playing tennis. It also helps you determine wholes from only partial input. If you see the back end of a horse from behind a barn, you're likely to predict the front end of a horse that you can't see yet.

Illusions help make this idea a little more solid. A Kanizsa triangle is an illusion in which the most important features of the shape, the corners, are accentuated (by the "pac-man" shapes). Your cortex directly senses the three corners, and "fills in" the sides of the triangle, even though you aren't directly sensing them, so that your subjective experience is that you are seeing a triangle. In a recent study with macaques, researchers measured the activation in areas V1 (lower in the hierarchy, processing primitives like lines and contours) and V2 (higher in the hierarchy, processing simple shapes) in response to Kanizsa figures and complete objects.

They measured the monkeys' response in V1 to normal objects. Then they showed them the Kanizsa figures and measured the response in V1. It turns out that activity in V1 in response to the illusory lines is almost the same as to seeing actual lines, with one exception...there is a delay. So what's likely happening when you see an illusion like this is that you get the input for the corners, that flows up the hierarchy to V2 and activates the concept of "triangle", which in turns drives predictions down the hierarchy that if you're seeing a triangle, you should be seeing lines where the sides are. This predictive activation gives the subjective impression of lines, though not as strongly as if you sensed them directly.

Prediction also works across ways of sensing. Hawkins gives the example of sitting at his computer, with his back to the door, and hearing his cat meow behind him. The auditory input travels up the hierarchy and activates the concept for his cat, which pushes predictions about what he should see down the hierarchy. So his visual processing is already primed to see a cat when he turns around, which will help process the information more quickly. If he turns around and there's an elephant standing there, a conflict between his prediction and experience will result, and he'll either have to resolve it as a new pattern or part of the existing pattern, or reject it as an anomaly.

So that's a fairly broad and simple view of the neocortex. I haven't mentioned its interactions with other areas of the brain, or activity that is more horizontal than vertical. But this gives the gist of a theory of cortical function, in which small neural modules acquire robust representations of patterns in the world, use them to signal to higher parts of the cortex that their pattern is present, and then signal to lower parts of the cortex about what they predict will happen next.

That's the theory in a nutshell. If this stuff makes sense, let me know. If not, let me know. I may discuss how I'm actually implementing these ideas in a computer simulation, but that's going to be even more technical.