A survey on neural network quantization for vision tasks

PhD Qualifying Examination


Title: "A survey on neural network quantization for vision tasks"

by

Mr. Rongzhao ZHANG


Abstract:

Recent years have witnessed a great prosperity of deep neural networks 
(DNNs) in both academia and industry. However, since a DNN model usually 
consists of multiple cascaded layers and a huge number of parameters, it 
is computationally expensive to deploy, prohibiting the usage of DNN in 
resource-limited scenarios like mobile applications and embedding 
devices. Neural network quantization is a powerful technique that can 
effectively reduce the bitwidth (e.g., from 32-bit to binary) of 
parameters and operations in a neural network with only moderate or even 
no performance drop, so that the computation burden of DNNs can be 
largely reduced. In this survey, we review a range of neural network 
quantization approaches in detail. We first introduce the background of 
neural network quantization researches. Then I divide existing 
quantization methods into several categories: STE-based methods, 
optimization-oriented ones and architectural approaches as well as 
knowledge distillation; the representative approaches in each categories 
are elaborated. I also summarize their applications to different vision 
tasks and the corresponding performance. Moreover, I discuss the future 
directions for deep model quantization and its potential applications to 
medical image analysis challenges.


Date:			Thursday, 26 November 2020

Time:                  	2:00pm - 4:00pm

Zoom meeting:           https://hkust.zoom.us/j/4355150756

Committee Members:	Prof. Albert Chung (Supervisor)
 			Prof. Chi-Keung Tang (Chairperson)
 			Prof. Tim Cheng
 			Prof. Weichuan Yu (ECE)

**** ALL are Welcome ****