Single layer look-up perceptrons

Concerns a single layer perceptron (SLP) which incorporates n-tuple pattern recognition techniques in an SLP architecture to produce a single layer look-up perceptron (SLLUP) which can learn the same types of nonlinear mappings as a multilayer perceptron but with a fraction of the training and computation. An additional very desirable property of the SLLUP is that it produces a quadratic error surface and so convergence to optimal performance is assured. It is argued that the SLLUP is basically an interpolation system which is able to generate an estimate of a continuous mapping function from a sparse set of training examples and is well suited to dealing with simple nonlinear mappings such as parity detection. The ability of the SLLUP to work on the very complex mapping problems of speech recognition and text to speech synthesis is also examined and compared with the performance obtainable using the multilayer perceptron. It is seen that the SLLUP can very nearly equal the performance of the MLP in these problems, suggesting that the MLP also does little more than a straightforward sample interpolation