A filter neural network

This paper proposes to add a filter layer to a dot product matching neural network. The purpose of the filter layer is to discard those unfavourable choices by checking the lower and upper bounds of each exemplar with the test pattern. The product is a supervised, fast learning filter neural network. It has a better generalisation capability than an ordinary dot product matching neural network. The new neural network is tested for speaker-independent spoken number (in English) recognition. An accuracy of 96.5% is reported for the test data. Without the filter layer, the recognition rate falls to 94.0%.