Neighborhood distortion functions and self-organizing feature maps

Summary form only given, as follows. Neighborhood distortion functions have been introduced as an aid in training self-organizing feature maps (SOFMs). These distortion functions are obtained by generalizing vector quantization (VQ) to account for the SOFM neighborhood mechanism; this formulation is called multivector quantization (MVQ). Unlike VQ, where the distortion is introduced by mapping an input to the best matching codevector, in MVQ the distortion is introduced by mapping an input to a set of codevectors. For SOFMs, this set consists of the weight vectors of the best matching unit and its neighbors. Learning rules for training SOFMs to preserve topology by minimizing MVQ distortion measures were also considered.