A Training Scheme for Pattern Classification Using Multi-layer Feed-forward Neural Networks.

摘要

This study highlights on the subject of weightudinitialization in multi-layer feed-forward networks.udTraining data is analyzed and the notion of criti-udcal point is introduced for determining the initialudweights for the input to hidden layer synaptic con-udnections. The proposed method has been applied toudartificial data. The experimental results show thatudthe proposed method takes almost 1/2 of the train-uding time required for standard back propagation.