Visualisation and 'Diagnostic Classifiers' Reveal how Recurrent and Recursive Neural Networks Process Hierarchical Structure (Extended Abstract)

In this paper, we investigate how recurrent neural networks can learn and process languages with hierarchical, compositional semantics. To this end, we define the artificial task of processing nested arithmetic expressions, and study whether different types of neural networks can learn to compute their meaning. We find that simple recurrent networks cannot find a generalising solution to this task, but gated recurrent neural networks perform surprisingly well: networks learn to predict the outcome of the arithmetic expressions with high accuracy, although performance deteriorates somewhat with increasing length. We test multiple hypotheses on the information that is encoded and processed by the networks using a method called diagnostic classification. In this method, simple neural classifiers are used to test sequences of predictions about features of the hidden state representations at each time step. Our results indicate that the networks follow a strategy similar to our hypothesised ‘cumulative strategy’, which explains the high accuracy of the network on novel expressions, the generalisation to longer expressions than seen in training, and the mild deterioration with increasing length. This, in turn, shows that diagnostic classifiers can be a useful technique for opening up the black box of neural networks.

Elman Backpropagation as Reinforcement for Simple Recurrent Networks

Simple recurrent networks (SRNs) in symbolic time-series prediction (e.g., language processing models) are frequently trained with gradient descent--based learning algorithms, notably with variants of backpropagation (BP). A major drawback for the cognitive plausibility of BP is that it is a supervised scheme in which a teacher has to provide a fully specified target answer. Yet agents in natural environments often receive summary feedback about the degree of success or failure only, a view adopted in reinforcement learning schemes. In this work, we show that for SRNs in prediction tasks for which there is a probability interpretation of the network's output vector, Elman BP can be reimplemented as a reinforcement learning scheme for which the expected weight updates agree with the ones from traditional Elman BP. Network simulations on formal languages corroborate this result and show that the learning behaviors of Elman backpropagation and its reinforcement variant are very similar also in online learning tasks.