A RAM-based neural net with inhibitory weights and its application to recognising handwritten digits

A method for introducing inhibitory weights into RAM based nets has been developed. The inhibitory weights leads to a more robust net and much lower error rates can be obtained. In the paper we describe how the inhibition factors can be learned with a one shot learning scheme. The main strategy is to choose the inhibitory values so that they minimise the error-rate obtained in a crossvalidating test performed on the training set. The inhibition technique has been tested on the task of recognising handwritten digits. The results obtained match the best error rates reported in the literature