139
Learning/prediction/prognosis/Frith: thesis: we learn through prediction. The brain uses mistakes in the prediction to make it better next time.
>
Learning, >
Trial-an-Error.
Solution: the brain makes two different predictions about my movements:
1. inverse model: which command sequence will generate the motion? The brain must "conclude back" from the output of the motion device to its input.
2. forward model: what movements will occur when a sequence of commands are sent?
Test: can only be done by real movement.
I 140
N.B.: it does not need a real movement to test whether both predictions are compatible.
Brain: the brain can adjust my models (forward and backward) without making a movement.
Def Helmholtz Machine/Frith: the Helmholtz Maschine is a machine that can learn in this way: to compare the backward and the forward model without performing the movement already.
Forward model: what movement is there when I send this command?
Backward model: what commands must have been sent for this movement to occur?
N.B.: that happens in sleep.
>
Dream.
Wake-Sleep Algorithm: uses two types of prediction:
1. Recognition: predicts which object will trigger these sensory sensations (backward).
2. Generate: predicts which sensations this object will produce (forward).
I 166
Prediction/error/Bayesian Theorem/Frith: what happens if a mistake is noticed in the prediction?
>
Bayesianism.
Backwards: the activity pattern that should be perceived by the eyes, ears, etc.: P (X I A).
Mistake: helps us to come to a better understanding of the world.
P (A I X): is this new conception (update).
Correction: the brain now runs through a loop, in which each time a different activity pattern is predicted. The error is reduced every time.
Knowledge: if the error is small enough, the brain "knows" what is outside. However, the brain continues to go through this process incessantly.
>
Brain, >
Brain states, >
Brain/Frith, >
Knowledge.