Automatic growing of a Hopfield style neural network for classification of patterns

Hopfield networks, a type of recurrent neural network, with a modification of the perceptron learning rule or Hebbian learning as the learning paradigm, may be used as a tool for classification. The usual method of using the Hopfield network as an associative memory has certain shortcomings however. Brouwer (1993, 1994) proposed a method which overcomes these shortcomings. This method avoids the problem of overloading the network that is due to attempting to store all of the elements of the training sets. The training algorithm described there is based upon a further extension of the Widrow Hoff learning rule. The results are quite good, however the connection matrix in case of Hopfield style networks is huge. This paper describes how the original network may be replaced by a much smaller network which is allowed to grow automatically, during training, from a one neuron network to the size required. This paper concludes with the results of applying the method to classification of diabetes patients and classification of cervical cells.