How to train RNNs on chaotic neural data?
Date:
I presented my work on how to train RNNs on chaotic neural data. This includes theoretical findings on the relationship of RNNs long-term behaviour and its loss gradients in training, as well as a new variant of teacher-forcing, sparsely forced BPTT, which allows to train on chaotic neural data.
Here are a short video describing our reseach and the corresponding poster. An abstract for this presentation can be found in the abstract proceedings of the Bernstein Conference 2021.