Feature detectors and difference operators in BP networks

Summary form only given, as follows. Hidden units are commonly thought to function as feature detectors. In the present work it was shown, however, that hidden units in networks trained using the backpropagation (BP) algorithm actually function as difference, or novelty, detectors most of the time. This was illustrated in four ways: (1) the form of the hidden unit weight vectors, (2) correlations between these weights and the training patterns, (3) the vectors obtained when these weights are multiplied by various input patterns, and (4) the weight vectors obtained from a piecewise linearized model of the network