A fast learning technique for the multilayer perceptron

The authors propose a learning technique for the multilayer perceptron based on an error function which is linear in the output layer weights, allowing for optimal adaptation of these weights. The weight adaptation sequence was modified to take full advantage of the proposed learning method. The application of the technique to parity problems indicates a drastic improvement in the convergence time compared to the conventional backpropagation technique. The number of iterations needed for convergence by the technique was about 10% of that needed by conventional backpropagation. This result makes the technique a more attractive alternative for real-time nonlinear adaptive filtering tasks. A perceptron using this method can be a very strong candidate for common nonlinear filters, with the advantage that the nonlinearity in the perceptron structure is not restricted as in such filters