Artificial neural networks work with numbers. To apply them to non-numeric data one has to represent this data by numbers. Finding an appropriate numberic representation is a difficult task which usually involves some general heuristics, common sense and some intuitions about the behaviour of neural nets and the problem itself. In this paper we present an approach to automatic discovery of optimal representations of data for feedforward multilayer networks. The resulting algorithm extends in a natural way the standard backpropagation scheme.
展开▼
