MLP classifiers: overtraining and solutions
Zheru Chi · 2002
Training a multi-layer perceptron (MLP) classifier is difficult to control and as a result its performance on unseen patterns is unpredicted. Overtraining is one of many problems in training an MLP classifier. In this paper, the author first discusses the overtraining problem based on an artificial two-input two-category classification problem. The author then suggests five solutions to the overtraining problem, which are supported by experimental results.