Download
Description
An important question in neuroscience is what determines whether neurons to learn “selective” representations (where they are important to one and only one task aspect, for example visual cortex) over “mixed-selective” ones (where they are important to multiple tasks or tasks aspects, for example PFC).
Is it architecture, learning algorithm, or task structure which matters most?
In this project I approach this question computationally by training artificial neural networks to learn multiple tasks an analysing neuronal selectivity. I find all three factors play a roll in determining whether neurons eventually become selective or mixed-selective.
Why
This work was completed at Harvard in collaboration with Cengiz Pehlevan and Sam Gershman as part of Sam’s course on computational cognitive neuroscience. It was never published…I should pick it up again at some point.
