Can Artiicial Neural Networks Discover Useful Regularities?

It is argued that the success of artiicial neural networks (ANNs) to date has depended almost exclusively upon the judicious choice of input representations which eeectively recode high-order input parameters as low-order parameters. It is demonstrated that, where such a recoding is required but not provided, BP ANNs fail to generalize. Of course, BP ANNs can utilize the spurious low order statistics associated with almost any`natural' problem. However, these provide partial information about the underlying mapping, and therefore do not, in general, permit generalization. In order to obtain generalization it is necessary to hand-craft inputs so that parameters which were coded as relations between input components are coded by a single input component. Thus the responsibility for enabling an ANN to learn a given task lies ultimately with the designer of the ANN. It is argued that this is an undesirable state of aaairs, in a eld where it is widely accepted that the process of learning supposedly obviates the need to hand-craft computational models.