Invariant feature representation by sparse vectors using adaptive subspace self-organizing map

Extraction of invariant features is a crucial process in pattern recognition. In this paper, a universal framework for encoding invariant features with sparse vectors is described. Using Kohonen´s adaptive subspace self-organizing map (ASSOM) [T. Kohonen, 2001], external inputs are encoded in a sparse vector form. These sparse vectors can be used to differentiate the input patterns as well as to represent the invariant features of a pattern category. Experiments using human speech data as an example are described. More importantly, the algorithm can be immediately applied to other forms of input data.