Distributed arithmetic perceptron

The shift of the nonlinearity from the neuron to the input allows the realisation of any mapping by a single perceptron. The resulting perceptron is unimodal and consequently there are no problems of local minima and excessive time-consuming training procedures. In the paper a method is proposed for carrying out this preprocessing in a more general way. Moreover, it is shown that the weights of the connections can be explicitly determined from the training set.