On the induction of temporal structure by recurrent neural networks
Language acquisition is one of the core problems in artificial intelligence (AI) and it is generally accepted that any successful AI account of the mind will stand or fall depending on its ability to model human language. Simple Recurrent Networks (SRNs) are a class of so-called artificial neural networks that have a long history in language modelling via learning to predict the next word in a sentence. However, SRNs have also been shown to suffer from catastrophic forgetting, lack of syntactic systematicity and an inability to represent more than three levels of centre-embedding, due to the so-called 'vanishing gradients' problem. This problem is caused by the decay of past input information encoded within the error-gradients which vanish exponentially as additional input information is encountered and passed through the recurrent connections. That said, a number of architectural variations have been applied which may compensate for this issue, such as the Nonlinear Autoregressive Network with exogenous inputs (NARX) network and the multi-recurrent network (MRN). In addition to this, Echo State Networks (ESNs) are a relatively new class of recurrent neural network that do not suffer from the vanishing gradients problem and have been shown to exhibit state-of-the-art performance in tasks such as motor control, dynamic time series prediction, and more recently language processing. This research re-explores the class of SRNs and evaluates them against the state-of-the-art ESN to identify which model class is best able to induce the underlying finite-state automaton of the target grammar implicitly through the next word prediction task. In order to meet its aim, the research analyses the internal representations formed by each of the different models and explores the conditions under which they are able to carry information about long term sequential dependencies beyond what is found in the training data. The findings of the research are significant. It reveals that the traditional class of SRNs, trained with backpropagation through time, are superior to ESNs for the grammar prediction task. More specifically, the MRN, with its state-based memory of varying rigidity, is more able to learn the underlying grammar than any other model. An analysis of the MRN’s internal state reveals that this is due to its ability to maintain a constant variance within its state-based representation of the embedded aspects (or finite state machines) of the target grammar. The investigations show that in order to successfully induce complex context free grammars directly from sentence examples, then not only are a hidden layer and output layer recurrency required, but so is self-recurrency on the context layer to enable varying degrees of current and past state information, that are integrated over time.