Bias Estimation for Neural Network Predictions

Abstract

This paper looks at the problem of performance assessment in the use of neural networks for classification tasks. It is well-known that the prediction obtained from a trained neural network is subject to errors, both in terms of bias and variance in the estimated error rate.

In order to estimate these measures, it is customary to reserve some data as a test set. This is reasonable if data are plentiful, but when the data set is small in size, this is likely to reduce the accuracy of the network estimates, simply because there are not enough data left for adequate training. An alternative approach will allows the use of all the data, is to employ the bootstrap method.

Here we give a brief introduction to the bootstrap, and then report on some computational experiments on artificial data sets in order to investigate the potential of this approach for the estimation of error bias.