Cascaded vector quantization by non-linear PCA network layers

The different mechanisms of principal component analysis (PCA) and vector quantization are combined in an architecture of one functional layer which implements vector quantization without using winner-take-all nets. After introducing cascaded vector quantization, the paper introduces a new network (the binary cascade network) which is composed of lateral inhibited neurons for PCA. They have bell-shaped activation functions which provide binary cascaded quantization stages. It is shown that this architecture is nearly optimal in terms of resource distribution