A sparse matrix approach to neural network training

摘要

A new training technique, based on sparse matrix concept is developed for the training of multilayer perceptron. The proposed approach exploits the patterns of neuron activations in neural networks and substantially reduces the amount of computations in backpropagation. The proposed training algorithm is applied to word recognition with TI20 real speech data. Compared to techniques without using the sparse concept, same or better recognition accuracy is achieved and training speed is substantially improved.