A new approach to the design of neural network classifiers and its application to the automatic recognition of handwritten digits

Describes a procedure for simultaneously building and training a neural network. Its salient features are the following: (1) the resulting network uses neurons with binary outputs, which makes hardware implementations straightforward; (2) the network has one single layer of trainable connections, therefore, training is fast; (3) the additional layers perform explicit Boolean functions, therefore these layers require no training and they can be implemented in hardware with standard logic gates; and (4) the procedure gives insight into the complexity of the problem. The application of this procedure to the recognition of handwritten digits is presented. The structure of an application-specific integrated circuit, which is in the design phase, is briefly described.<<ETX>>