Rapid Vector Quantization and Classi cation with Neural Networks

The majority of today's Neural Networks are either Multi-Layer-Perceptron networks (MLP) or Feature-Maps like Kohonen's Self-Organizing Map (SOM). Usually they are simulated on ordinary single-processor von-Neumann hardware to be used for some kind of vector quantization, classiication or coding. However, these simulated Neural Networks are expensive in respect to the computational costs they demand. This report will review brieey several methods for speeding up the quant-ization with Feature-Maps and then present a novel approach for a rapid quantization with MLP networks. It will show how the properties of the Euclidean Distance can not only be used to speed up the search for the winning output neuron in Feature-Maps but in MLP networks as well.