Vectorized backpropagation and automatic pruning for MLP network optimization

An analysis of the backpropagation algorithm is presented, and the significance of vectorized backpropagation and automatic pruning for better learning performance and multilayer perceptron (MLP) network optimization is revealed. During the learning phase, the network which uses vectorized backpropagation converges within 20%-50% of the iterations required for the standard MLP to converge without affecting the test set performance. The network pruning algorithm reduces the number of hidden nodes and connection weights. The pruned network, with only 40% connection weights of the unpruned network, gives the same learning and recognition performance as the parent unpruned fully connected network