Lecture 2: Receiver-operating characteristics

Neurons act as if they were performing reinforcement learning.

Neural Encoding

How does the brain encode the sensory stimuli?

from Stimulus: $s$ to ⟶ Response: $r$

⟹ We try to guess $P(r \mid s)$

⟶ but: it’s a hard problem

so instead, we compute backward probability:

from $r$ to ⟶ $p(s \mid r)$

Psychophysical threshold curves: experiences with light flashes ⟹ how the stimulus relates to the behavior

Based on the output of the cell, what would an ideal observer decide?

from histogram distributions

There are two types of wrong answers: false positives and false negatives (miss)

Then: we identify a discrimination threshold to separate the distributions

Sometimes:

you want a tight threshold (to absolutely avoid misses (to don’t want to miss the presence of a tiger!))
you want a loose threshold (to avoid false positives (to don’t want to begin the race in advance not to get disqualified!))

Then you can plot the evolution of the threshold: receiver-operating characteristics

Have a monkey try to guess where a bunch of moving points is heading: to the right or to the left?

Neurons in area V5 have a space receptive field

Weird paradox: if we can tell what the monkey will do, what do the other neurons do?

Even more surprising: a single neuron can perform better than the monkey itself!

How neural activities encode the link between stimulus and behavioral response?

Can we understand what happens between neural responses and behavior?

$r$ ⟶ $p(\underbrace{a}_{\text{action}} \mid r)$