Computational Cost Reduction by Selective Attention for Fast Speaker Adaptation in Multilayer Perceptron

Abstract

Selective attention learning is proposed to improve the speed of the error backpropagation algorithm of a multilayer Perceptron. Class-selective relevance for evaluating the importance of a hidden node in an off-line stage and a node attention technique for measuring the local errors appearing at the output and hidden nodes in an on-line learning process are employed to selectively update the weights of the network. The acceleration of learning time is then achieved by lowering the computational cost required for learning. By combining this method with other types of improved learning algorithms, further improvement in learning speed is also achieved. The effectiveness of the proposed method is demonstrated by the speaker adaptation task of an isolated word recognition system. The experimental results show that the proposed selective attention technique can reduce the adaptation time more than 65% in an average sense.