Inverse-step competitive learning

Reviews several variants of the competitive learning rule: simple competitive learning, the Kohonen self-organization map, and frequency-sensitive competitive learning. They then propose a novel learning rule based on competitive learning, called inverse-step competitive learning (ISCL). The isolated points play a more important role than the normal points in simple competitive learning, because the learning step is proportional to the distance between the input pattern and the weights value. The basic idea of this learning rule is to take a learning step which is a descending function of this distance. The authors give the first results of the ISCL rule and compare it to simple competitive learning