Summary
International Symposium on Nonlinear Theory and its Applications
2017
Session Number:C0L-B
Session:
Number:C0L-B-2
Consideration on Quantization Functions in Quantized Neural Networks
Takumi Kadokura, Hidehiro Nakano, Arata Miyauchi,
pp.568-571
Publication Date:2017/12/4
Online ISSN:2188-5079
Summary:
Quantized Neural Network (QNN) is a kind of neural networks in which its weights and activations are quantized. Since QNN can reduce computational quantity and energy consumption by quantization, it is expected to be used on embedded devices. This paper investigates quantization functions used for quantizing gradients in QNN. By performing the numerical experiments, the performances of some quantization functions are compared. We then show that there exists a quantization function which can keep high performance with low quantization bit rate.