Post

First problem: the same electrical stimulation does not always produce the same neural response. Responses were much more variable across days than within a day, motivating a model that accounts for the brain’s ongoing state rather than treating stimulation as a fixed input-output mapping.

So we trained a deep forward model to predict the population response to multielectrode stimulation. On held-out days, it outperformed 1-to-1, linear, nonlinear, and dictionary-based models across multiple measures of prediction accuracy.

Prediction is useful. Control is the harder problem. Given a desired neural response, we used either gradient optimization or an inverse neural network to find a stimulation pattern predicted to produce it, then tested those patterns in vivo. Both beat conventional approaches.