Project in Prairie
We use Recurrent Neural Networks (RNNs) as functional and fully observable models of cortical networks. To this end, we employ algorithms from AI to train RNNs to perform various tasks developed in the field of cognitive and systems neuroscience. We then seek to understand the relationship between the structure of the trained networks, the dynamics of their activity, and the computations they implement.
The methods that we develop for understanding and interpreting trained networks are inspired by statistical physics and directly related to the field of mechanistic interpretability in AI.