A Statistical Complexity Framework for Topology Preserving Adaptive Vector Quantization
Sonmez, Kemal M
Date: March 20 - March 22, 1996
We propose a statistical model complexity framework for topology preserving adaptive vector quantization. In this setting, the adaptation of the neighborhood function during training of the codebooks, which is essential for producing global organization, may be regarded as increasing the statistical model complexity as more data become available. Therefore, the training is equivalent to on the fly optimization of the bias/variance trade-off.