METHOD AND ALGORITHM OF RECURSIVE DEEP LEARNING QUANTIZATION FOR WEIGHT BIT REDUCTION

摘要

A system for reducing weight storage bits for a deep-learning network and a method thereof include a quantization module and a cluster-number reduction module. The quantization module quantizes neural weights of each quantization layer in a deep-learning network. The cluster-number reduction module reduces a preset number of clusters for a layer having a clustering error which is the minimum value of clustering errors of a plurality of quantization layers. The quantization module performs re-quantization based on a preset number of clusters reduced with respect to the layer. The cluster-number reduction module further determines another layer having a clustering error, which is the minimum value of the clustering errors of the quantized layers. In addition, the cluster-number reduction module reduces the preset number of clusters with respect to another layer until the recognition performance of the deep-learning network is reduced by a preset threshold value.