A modification of Kohonen's Learning Vector Quantization is proposed to handle hard cases of supervised learning with a rugged decision surface or asymmetries in the input data structure. Cell reference points (neurons) are forced to move close to the decision surface by successively omitting input data that do not find a neuron of the opposite class within a circle of shrinking radius. This simulates habituation to frequent but unimportant stimuli and admits problem solving with fewer neurons. Simple estimates for the optimal shrinking schedule and results of illustrative runs are presented.