ABSTRACT
It is shown in this paper that a particular adaptive self-organizing network is able to create sets of wavelet- and Gabor-type filters when randomly displaced or moving input patterns are used as training data. No analytical functional form of these filters is thereby postulated. It is plausible that the same kind of adaptive system could create many other kinds of invariant-feature filters, if there exist corresponding transformations in the training data. Such a system, called ASSOM (Adaptive-Subspace SOM), can act as a learning feature-extraction stage for pattern recognizers, being able to adapt to an arbitrary sensory environment.
