Noise-Shaping Quantization Harmonized by Training Data for Compaction of Neural Networks

Abstract

This paper addresses the quantization of weight coefficients in trained neural networks to achieve more compact representations. To mitigate the degradation of the input-output relationship in neural networks due to quantization, we introduce the noise-shaping quantization method. This method quantizes coefficients and distributes the resulting quantization error over coefficients not yet quantized. In this paper, we proposed an approach to adjust the magnitude of the quantization error using training data. The effectiveness of the proposed method was then validated through an image classification problem. Finally, we related the quantization problem of neural network weight coefficients to the quantization problem of time signals in dynamic systems and discussed its distinctions from previous quantization techniques.