next up previous
Next: 3. Concurrent Processing Up: 2.4 Backpropagation Neural Networks Previous: 2.4.5 Local Minimum Problem

2.4.6 Generalization

A trained backpropagation network is able to detect and classify an input pattern that has not been seen during learning. This feature is called generalization, borrowed from the psychology terms. Neural networks are known to be good at classifying noisy input patterns, but not at classifying a pattern that is intermediate between two solid patterns from the training samples. In other words, neural networks are good at interpolation but not extrapolation [BJ91]. Also, there may exist overfitted input data, the unseen input pattern such that it can be classified into one of the trained output response undesirably [Hay99]. Suggested solution includes modification of network architecture and more adequate training samples.

Kiyoshi Kawaguchi