Competitive Learning with Fast Neuron-Insertion

Abstract

Adaptive Vector Quantization (AVQ) is to find a small set of weight vectors that well approximates a larger set of input vectors. This paper presents a fast AVQ method Competitive Learning with Approximate Neuron-Insertion (CLANI). Though neuron-insertion techniques can much enhance the accuracy in AVQ, a naive implementation requires a large computational cost proportional to the number of input vectors. Approximate neuron-insertion has an advantage that its computational cost is independent of the number of input vectors. We theoretically estimate the computational costs of CLANI and the other conventional methods. The effectiveness of CLANI is demonstrated in vector quantization simulations and an image compression application.