Invariant feature maps for analysis of orientations in image data

We present a method that uses competitive learning and a neighbourhood function in a similar way to the self-organising map (SOM) [3]. 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 rotation of these subspaces is weighted by the neighbourhood function. Non-linear activation functions are introduced so that each node performs nonlinear PCA on the training data captured in its Voronoi region. We show that this network may be used for position invariant detection of bars at varying orientations.