Stochastically competitive learning algorithm for vector quantizer design

The problem of vector quantizer (VQ) design for practical applications can be divided into two phases: 1) to search a globally optimal codebook using a given set of training data; and 2) to make it adaptive to the new signals outside the set. The most widely used technique for VQ design is the generalized Lloyd algorithm (GLA), while the Kohonen learning algorithm (KLA) is a very promising alternative due to its inherently adaptive capability. However, both the GLA and KLA tend to get trapped into poor local optima due to their “greedy” nature in the search process. By incorporating the principle of stochastic relaxation into the KLA, we propose a stochastically competitive learning algorithm (SCLA), which will approach the global optimum regardless of the initial configuration due to its capability of “pulling” itself from local optima. Based on the SCLA, a coding scheme is then outlined in detail to design a codebook both globally optimal for a given set of training data and adaptive to new data outside the set