Design of Noise Shaping Quantizers for Data Compaction of Neural Networks

Abstract

This paper focuses on the quantization problem of connection weights of neural networks. Our previous work proposed a class of quantizers, called noise-shaping quantizer, for the quantization of neural networks. The performance of the proposed quantizer depends on the error diffusion filter. This paper proposes a systematic design method of error diffusion filters based on features of learning data, which is used for the learning of neural networks. In the proposed design method, satisfactory error diffusion filters are given by solving a kind of traveling salesman problems.