Weight Quantization for Multi-layer Perceptrons Using Soft Weight Sharing

Abstract

We propose a novel approach for quantizing the weights of a multi-layer perceptron (MLP) for efficient VLSI implementation. Our approach uses soft weight sharing, previously proposed for improved generalization and considers the weights not as constant numbers but as random variables drawn from a Gaussian mixture distribution; which includes as its special cases k-means clustering and uniform quantization. This approach couples the training of weights for reduced error with their quantization. Simulations on synthetic and real regression and classification data sets compare various quantization schemes and demonstrate the advantage of the coupled training of distribution parameters.