Accelerating Deep Convolutional Neural Networks via Filter Pruning

He, Yang

Accelerating Deep Convolutional Neural Networks via Filter Pruning

He, Yang

Permalink

Publication Type:: Thesis
Issue Date:: 2021

Open Access

Copyright Clearance Process

Recently Added
In Progress
Open Access

This item is open access.

Adobe PDF

Download contents and abstractAdobe PDF (587.45 kB)

Adobe PDF

Download thesisAdobe PDF (2.84 MB)

View statistics

Full metadata record

Field	Value	Language
dc.contributor.author	He, Yang
dc.date.accessioned	2022-08-22T00:02:07Z
dc.date.available	2022-08-22T00:02:07Z
dc.date.issued	2021
dc.identifier.uri	http://hdl.handle.net/10453/160643
dc.description	University of Technology Sydney. Faculty of Engineering and Information Technology.	en_US.UTF-8
dc.description.abstract	The superior performance of deep Convolutional Neural Networks (CNNs) usually comes from the deeper and wider architectures, which cause the prohibitively expensive computation cost. To reduce the computational cost, works on model compression and acceleration have recently emerged. Among all the directions for this goal, filter pruning has attracted attention in recent studies due to its efficacy. For a better understanding of filter pruning, this thesis explores different aspects of filter pruning, including pruning mechanism, pruning ratio, pruning criteria, and automatic pruning. First, we improve the pruning mechanism with soft filter pruning so that the mistaken pruned filters can have a chance to be recovered. Second, we consider the asymptotic pruning rate to reduce the sudden information loss in the pruning process. Then we explore the pruning criteria to better measure the importance of filters. Finally, we propose the automatic pruning method to save human labor. Our methods lead to superior convolutional neural network acceleration results.	en_US.UTF-8
dc.format	Thesis (PhD)
dc.language.iso	en_US	en_US.UTF-8
dc.relation	https://opus.lib.uts.edu.au/bitstream/10453/160643/2/02whole.pdf
dc.rights	The author owns the copyright in this thesis including all reproduction and reuse rights for the work. The work may not be altered without the permission of the copyright owner. Attribution is essential when quoting or paraphrasing from this thesis.
dc.rights	au.edu.uts.lib/ppc
dc.rights	info:eu-repo/semantics/openAccess
dc.title	Accelerating Deep Convolutional Neural Networks via Filter Pruning	en_US.UTF-8
dc.type	Thesis
utslib.copyright.status	open_access	*

Abstract:

The superior performance of deep Convolutional Neural Networks (CNNs) usually comes from the deeper and wider architectures, which cause the prohibitively expensive computation cost. To reduce the computational cost, works on model compression and acceleration have recently emerged. Among all the directions for this goal, filter pruning has attracted attention in recent studies due to its efficacy. For a better understanding of filter pruning, this thesis explores different aspects of filter pruning, including pruning mechanism, pruning ratio, pruning criteria, and automatic pruning. First, we improve the pruning mechanism with soft filter pruning so that the mistaken pruned filters can have a chance to be recovered. Second, we consider the asymptotic pruning rate to reduce the sudden information loss in the pruning process. Then we explore the pruning criteria to better measure the importance of filters. Finally, we propose the automatic pruning method to save human labor. Our methods lead to superior convolutional neural network acceleration results.

Please use this identifier to cite or link to this item:

http://hdl.handle.net/10453/160643