A new learning paradigm for neural networks

Introduces a new way of inferring the structure of a temporal neural network from a set of training data. The approach is to learn a grammar which describes and generalises the input patterns, and then to map this grammar onto a connectionist architecture so allowing the network topology to be specialised to the training data. The resulting network has as many levels as are necessary, and arbitrary connections between levels. The resulting grammars are called strictly hierarchical and map straightforwardly onto a connectionist architecture using a relatively small number of neurons. The authors have performed some experiments on the recognition of hand-written isolated digits, using the simplest possible (supervised) grammatical inference algorithm to generalise a nonstochastic context-free grammar from the training data