Future project ideas

Fitting an RNN to reproduce all the inactivation data on the poisson clicks data (Alex).

Idea would be to initialize a recurrent neural network, and arbitrarily designate some fraction of neurons to represent each of the brain regions (FOF, PPC, striatum, SC...). Set up the cost function of the network to match the performance of the rat (meta rat, or maybe just synthetic data with the same general trends) on the control data, and the inactivation data for each brain region.
Qs/goals: (1) demonstrate a network can match those patterns of behavior. (2) Do neurons in each brain region match ephys data from real neurons? (3) What do the neurons outside of the designated areas do? (4) assuming the answers to 1-3 are reasonable...what can we infer about how the network is working? (5) What predictions can we make about simultaneously recording in one brain region while we inactivate another?
Challenges: (1) How to set up the inputs, and read out the decision? (2) training the network. Can standard RNN methods be used with inactivations? (3) Do we get degenerate solutions where all the computation is being done outside of the designated brain areas, and the network just uses the area inactivations as another "input"? (4) how sensitive are the results to training method/initialization, etc

Fitting an RNN with a variable integration timescale (Alex)

Ahmed and I have demonstrated that we can train rats to switch, on a session-by-session basis to integrate on a long or short timescale. So, if we train an RNN to be able to switch integration timescales (being a perfect integrator vs. a leaky integrator), how does the network perform that switch?
Qs: Probably much easier to train
Challenges: harder interpretation questions. Probably much more sensitive to training method, network setup, etc. 

Published work

Contextual integrator of Sussillo, Mante et. al.