Pattern grouping strategy makes BP algorithm less sensitive to learning rate

Yingyang Lu, Shenchu Xu, Boxi Wu, Zhenxiang Chen

Abstract: In the normal backpropagation learning process, the whole set of target patterns is learned again and again. We think that, the pattern grouping strategy (PGS), in which the patterns to be learned are divided into subgroups and are learned subgroup by subgroup, may be helpful for the BP training process. The investigation of the affect of the learning rate on the success rate of the BP learning process with PGS may serve as a proof.

External IDs:dblp:conf/ijcnn/LuXWC99