DON’T JUDGE A BOOK BY ITS COVER - ON THE DYNAMICS OF RECURRENT NEURAL NETWORKS

Doron Haviv, Alexander Rivkind, Omri Barak

Abstract: To be effective in sequential data processing, Recurrent Neural Networks (RNNs) are required to keep track of past events by creating memories. Consequently RNNs are harder to train than their feedforward counterparts, prompting the developments of both dedicated units such as LSTM and GRU and of a handful of training tricks. In this paper, we investigate the effect of different training protocols on the representation of memories in RNN. While reaching similar performance for different protocols, RNNs are shown to exhibit substantial differences in their ability to generalize for unforeseen tasks or conditions. We analyze the dynamics of the network’s hidden state, and uncover the reasons for this difference. Each memory is found to be associated with a nearly steady state of the dynamics whose speed predicts performance on unforeseen tasks and which we refer to as a ’slow point’. By tracing the formation of the slow points we are able to understand the origin of differences between training protocols. Our results show that multiple solutions to the same task exist but may rely on different dynamical mechanisms, and that training protocols can bias the choice of such solutions in an interpretable way.