Intensity encoding in unsupervised neural nets.

Abstract

The requirement of input vector normalisation in unsupervised neural nets results in a loss of information about the intensity of the signal contained in the input datastream. We show through a simple algebraic analysis that the introduction of an additional input channel encoding the root-mean-square intensity in the signals cannot restore this information if the input vectors have to be, nevertheless, all of the same length. We suggest an alternative method of encoding the input vectors where each of the input channels is split into two components in such a way that the resultant input vector is then of fixed length and retains information of the intensity in the signals. We further demonstrate, by using synthetic data, that a Kohonen Net is capable of forming topological maps of signals of different intensity, where an adjacency relationship is maintained both among the signals of the same frequency composition at different intensities and between signals of different frequency compositions at the same intensity. A second experiment reported here shows the same behaviour for less artificial inputs (based on a cochlear model) and additionally demonstrates that the trained network can respond appropriately to signals not previously encountered.