Using Constructive Learning in Speaker Verification. A Comparison with Multilayer Perceptron

摘要

This work deals with the use of Artificial Neural Networks for the speaker recognition problem. Specifically, we have been using backpropagation Multilayer Perceptron (MLP), with encouraging performance. However, MLPs have several problems in actual applications. One of them is the size of the network: it is impossible to determine a priori the optimal number of hidden layers an the size of each of them. If the number of weights is insufficient, the network is not able to learn the desired function, and if the number of them is too high the network tends to overfit. There are some theoretical and empirical results on this matter, but in practice only "trial and error" is really useful, and each "trial" involves the training of a new network. To deal with the previous problem, we are testing a different approach: a constructive neural network, using a Projection Pursuit Learning algorithm. The advantage of a constructive teaming is that we can add neurons incrementally in the training stage to reach the desired learning error, without having to train a new network, like in MLP. Here we present some preliminary results of the first experiments performed, that are compared with those of MLP.