Error estimation and learning data arrangement for neural networks
Kenji Fukumizu, Sumio Watanabe · 2002
The expected squared error of a trained neural network is analyzed from the statistical point of view. First we derive an estimation formula of the expected squared error at each input point. It tells us in which area the network response has high accuracy, and it works as a confidence index. We can utilize the confidence index for a criterion of rejection in applications. Second, a novel method of collecting learning data is proposed based on the error analysis. The proposed method can provide more suitable learning data than those taken from the true input distribution.>