Fast formation of invariant feature maps

We present a neural network that uses competitive learning and a neighbourhood function in a similar way to the self-organising map (SOM) fi]. The network consists of a number of modules that are positioned in an array (normally in one or two dimensions) where each module performs a subspace projection and the learning rate used within each module is weighted by the neighbourhood function. By using a subspace method that always gives convergence of orthonormal basis vectors within the modules, we demonstrate that this method can allow fast formation of filters of basic invariant features. The adaptive subspace self-organising map (ASSOM) [1,2] is in some ways similar and it has been proposed as a modular neural network for invariant filtering, providing invariances in scale, phase, rotation, translation etc. In this paper we show that our method can also provide these invariances with much improved performance in terms of the processing time required during training.