Principal components extraction by autoassociative feed-forward networks

A methodology for testing the ability of autoassociative, feedforward, backpropagation networks to extract the principal components from a training set is described. Some simple examples, treated analytically and validated by numerical trials, suggest that such behavior may be typical. It is hypothesized that N-H-N autoassociative networks tend naturally to discover principal components in pattern sets subject to mild constraints. It seems probable that the time required by the network to extract good approximations of the principal components is short compared to the time that might be spent trying to overfit the patterns and that the neural net method might be competitive with the classical procedure in real-time adaptive processing applications