用于矢量量化的神经网络竞争学习算法

Neural Network Competitive Learning Algorithms for Vector Quantization

对典型的竞争学习算法进行了研究和分析，提出了一种基于神经元获胜概率的概率敏感竞争学习算法（ＰＳＣＬ）。与传统竞争学习算法只有一个神经元获胜而得到学习不同，ＰＳＣＬ算法按照各神经元的获胜概率并通过对失真距离的调整使每个神经元均得到不同程度的学习，可以有效地克服神经元欠利用问题。

Neural network competitive learning algorithms are widely used for vector quantization. In this paper, some typical competitive learning algorithms have been specially investigated and analyzed . A new competitive learning algorithm based on the neuron win probability is presented for vector quantization. Unlike the traditional competitive learning algorithms where only one neuron will win and learn at each competition, every neuron in the proposed probabilitysensitive competitive learning algorithm (PSCL) will win to some extent, depending on its win probability and adjustment of distortion distance to the input vector. The new algorithm is shown to be efficient to overcome the problem of neuron underutilization.