Design and initialization of two-layer perceptrons using standard pattern recognition techniques

A neural network weight initialization method for two-layer perceptrons comprising direct input-output connections and squashing and concentric hidden units is presented. The method can be used to monitor the classification performance as a function of the network complexity without any error backpropagation training. The user can then select the best initialized network and train it further using the EBP algorithm. As the selected network is already properly initialized, the training will only take a few steps and the danger of converging toward a local minimum should be very small. The authors' method is illustrated with a broad phonetic speech classification problem.<>