Learning rare categories in backpropagation

Abstract

Hierarchical systems of neural networks based on the backpropagation algorithm were used to test the hypothesis that rare categories could be learned more accurately and in shorter training times than in nonhierarchical neural networks also based on the backpropagation algorithm. In two artificial data sets, the problem of learning rare categories was quantified and an existing solution was shown to be inadequate. HNNs were compared to nonhierarchical neural networks. In both artificial examples, HNNs performed better than nonhierarchical neural networks in terms of sensitivity and time to train. Specificities were not significantly different. In two real-world examples, these results were confirmed.