Diagnostic classification and symbolic guidance to understand and improve recurrent neural networks

This paper describes a search through a variety of methods to inspect and understand the internal dynamics of gated recurrent neural networks, using a task focusing on a key feature of language: hierarchical compositionality of meaning. We study both gated recurrent units and long short term memory networks with a technique called diagnostic classification, in which a simple neural meta-model is trained to qualitatively evaluate hypotheses about the information that is encoded in the hidden state of a trained network. Using this method, we analyse the hidden layer activations, but also the gates of trained networks. We explore the potential limits of diagnostic classification using not only the accuracy of the resulting diagnostic classifiers, but also studying their weights to understand where and how information is encoded. As a result, we develop a detailed understanding of the computations implemented by the networks to execute their task, which we then utilise to improve their performance with a new training regime we call symbolic guidance. This regime uses symbolically generated targets that are specifically designed to leverage our knowledge about the inherent bias of the networks and leads to a clear improvement against the best previous model.