Searching for Meaning in RNNs using Deep Neural Inspection

Abstract

Recent variants of Recurrent Neural Networks (RNNs)---in particular, Long Short-Term Memory (LSTM) networks---have established RNNs as a deep learning staple in modeling sequential data in a variety of machine learning tasks. However, RNNs are still often used as a black box with limited understanding of the hidden representation that they learn. Existing approaches such as visualization are limited by the manual effort to examine the visualizations and require considerable expertise, while neural attention models change, rather than interpret, the model. We propose a technique to search for neurons based on existing interpretable models, features, or programs.