…because if we were relying on directly predicting the world (and doing what we do in science) it's too slow, and it demands too much processing power and too many people.

And chimpanzees can't do science…so how did we get here?

Now we can do the first insight:
The brain doesn't actually directly predict the world…it tries to anticipate itself as it's causally interacting with the world.

The brain is actually predicting itself as you are causally interacting with the world.

There's a level of you in which you don't have to do anything:
If your eyes are open, your retinas are firing.
If you're moving your muscles, there are nerve impulses that are firing.

So you have a level of your processing (your sense organs and your muscles for example) that automatically generate sensory motor patterns.

Even if you're sitting absolutely still, you're getting sensory motor patterns.

Now let's think about what those are:
There's a baby. The baby is moving its head, its eyes are open, there's different colors and intensity of light…and what happens is its retina fire in different patterns according to that.

The retina is firing.
There's a pattern of neural firing.

And the baby is also moving its head, so there's a pattern of neural firing in the muscles.

So there's retinal firing (I'm simplifying by the way) and there's neural firing in my muscles in my neck. There are complex sensory motor patterns like that.

Let's call that the base level.

That level that happens just automatically from moving in the world.
Even if you're moving randomly like a baby that's just starting to learn.

Now there's a level “above that”….I don't mean literally above, I'm just talking theoretically.
There's a higher level brain area (it's probably thalamus and V1, but we don't need to do the anatomical stuff) and what it's trying to do is it's trying to detect patterns in the retinal and muscular firing.

It's trying to detect those patterns and complete them ahead of time.
That's what it's trying to do.

Now it's really bad at it when it starts.

But here's the difference:
That brain area can “talk” to the retina (the nerves in the retina) and the nerves in your neck…and those are trying to cooperate with that higher brain area. They’re trying to communicate with it, because they all belong to the same living communication and cooperation system.

So the retina and your nerve muscles can provide very clear feedback to that brain area of how it screwed up, of the mismatch between the pattern they tried and the pattern that actually unfolded.

That's prediction error.

Using that information, the higher brain area learns, it adjusts, it changes its sensitivity to which patterns it's detecting and completing.

And there's no magic there. Neural networks have been doing this since the mid 1980s. We know how this happens.

Now the brain area “tries again” (and this cycling is happening very fast) to detect patterns in the retina and the muscles…and because it's done some correction, its predictions are a little bit more precise…and still makes a lot of mistakes, but because its predictions were a little bit more clear and precise, the feedback it gets is a little bit more clear and precise.

And now the whole thing starts to cycle.
It's doing this in a self-organizing fashion.
There's no little man, no little homunculus in there doing it.

Now that cycling still takes a long time.

And part of the insight is:
There’s the base level, then brain area one (trying to anticipate the base level) and then I can have a brain area two that's trying to anticipate the neural firings in brain area one and detect them and complete them ahead of time…and then it gets feedback.

So this medium level is getting improved by error signals from the base level and correction signals top down.

That's what's going on.

Now, your brain has a high stack of these…very complex.

So you're getting all of these top down, bottom up signals, going at multiple levels…which means you get very good anticipation of your sensorimotor patterns.

Which results in:
You actually start to anticipate the causal patterns in the world that are causing those sensory motor patterns.

That's the crucial move.

As you get really good at detecting and completing ahead of time the sensory motor patterns…through that you get very good at anticipating the causal patterns in the world that are actually making the sensory motor patterns.

That's what it means to model the world.
It’s not forming a picture of the world.

You get all these constraints (top down, bottom up) so that you start anticipating the world well.

Top down…it's often called your priors. What's up there.
And it's often called your beliefs…for reasons I'll talk about later, that's a bad term.
We just use it because we originally talked about it that way and it's a mistake.

But you've got these top down constraints and you have bottom up error, top down constraints.
Priors and prediction error.

That's predictive processing.