An adaptive and fully sparse training approach for multilayer perceptrons
Fang Wang, Q.J. Zhang · 2002
An adaptive and fully sparse backpropagation training approach is proposed in this paper. The technique speeds up training by combining a sparse optimization concept with neural network training. The sparse phenomenon due to neuron activation, which is inherent in neural networks, is exploited in both feedforward and backpropagation phases. A new computational algorithm with sparse pattern reuse and refreshment has been developed together with the adaptation procedure of a new set of parameters which regulate the sparse training process. The proposed training approach has been applied to speech recognition and circuit extraction problems and achieved significant speed-up of training.