A Survey of Quantization Methods for Efficient Neural Network Inference
Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer
Abstract
This chapter provides approaches to the problem of quantizing the numerical values in deep Neural Network computations, covering the advantages/disadvantages of current methods. Over the past decade, people have observed significant improvements in the accuracy of Neural Networks (NNs) for a wide range of problems, often achieved by highly over-parameterized models. Achieving efficient, real-time NNs with optimal accuracy requires rethinking the design, training, and deployment of NN models. Model distillation involves training a large model and then using it as a teacher to train a more compact model. Loosely related to NN quantization is work in neuroscience that suggests that the human brain stores information in a discrete/quantized form, rather than in a continuous form. Gray and Neuhoff have written a very nice survey of the history of quantization up to 1998.