,
4 tweets,
3 min read
Read on Twitter
New work out on arXiv! Reverse engineering recurrent networks for sentiment classification reveals line attractor dynamics (arxiv.org/abs/1906.10720), with fantastic co-authors @ItsNeuronal, @MattGolub_Neuro, @SuryaGanguli and @SussilloDavid. #tweetprint summary below! 👇🏾 (1/4)
We analyze recurrent networks trained to perform sentiment classification, a standard natural language processing (NLP) task. We find that RNNs trained on this task as surprisingly interpretable using tools from dynamical systems. (2/4) 
The network dynamics are organized around a roughly 1-D approximate line attractor, which we identify by studying the eigendecomposition of the recurrent Jacobian of the dynamics at approximate fixed points. (3/4)
Finally, we can understand how the network processes individual words (tokens) by looking at projections of the embedding vectors onto the principal eigenvectors of the system. Overall, we think these tools will help us demystify and understand how recurrent networks work! (4/4)
