A unified DNN weight compression framework using reweighted optimization methods

To address the large model sizes and intensive computation requirements of deep neural networks (DNNs), weight pruning techniques have been proposed and generally fall into two categories: static regularization-based pruning and dynamic regularization-based pruning. However, the static method often...

Full description

Saved in:
Bibliographic Details
Main Authors: Mengchen Fan, Tianyun Zhang, Xiaolong Ma, Jiacheng Guo, Zheng Zhan, Shanglin Zhou, Minghai Qin, Caiwen Ding, Baocheng Geng, Makan Fardad, Yanzhi Wang
Format: Article
Language:English
Published: Elsevier 2025-09-01
Series:Intelligent Systems with Applications
Subjects:
Online Access:http://www.sciencedirect.com/science/article/pii/S2667305325000821
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849318336022708224
author Mengchen Fan
Tianyun Zhang
Xiaolong Ma
Jiacheng Guo
Zheng Zhan
Shanglin Zhou
Minghai Qin
Caiwen Ding
Baocheng Geng
Makan Fardad
Yanzhi Wang
author_facet Mengchen Fan
Tianyun Zhang
Xiaolong Ma
Jiacheng Guo
Zheng Zhan
Shanglin Zhou
Minghai Qin
Caiwen Ding
Baocheng Geng
Makan Fardad
Yanzhi Wang
author_sort Mengchen Fan
collection DOAJ
description To address the large model sizes and intensive computation requirements of deep neural networks (DNNs), weight pruning techniques have been proposed and generally fall into two categories: static regularization-based pruning and dynamic regularization-based pruning. However, the static method often leads to either complex operations or reduced accuracy, while the dynamic method requires extensive time to adjust parameters to maintain accuracy while achieving effective pruning. In this paper, we propose a unified robustness-aware framework for DNN weight pruning that dynamically updates regularization terms bounded by the designated constraint. This framework can generate both non-structured sparsity and different kinds of structured sparsity, and it incorporates adversarial training to enhance the robustness of the sparse model. We further extend our approach into an integrated framework capable of handling multiple DNN compression tasks. Experimental results show that our proposed method increases the compression rate – up to 630× for LeNet-5, 45× for AlexNet, 7.2× for MobileNet, 3.2× for ResNet-50 – while also reducing training time and simplifying hyperparameter tuning to a single penalty parameter. Additionally, our method improves model robustness by 5.07% for ResNet-18 and 3.34% for VGG-16 under a 16×pruning rate, outperforming the state-of-the-art ADMM-based hard constraint method.
format Article
id doaj-art-5018871cd3fd472fa2eedf1f87e22e70
institution Kabale University
issn 2667-3053
language English
publishDate 2025-09-01
publisher Elsevier
record_format Article
series Intelligent Systems with Applications
spelling doaj-art-5018871cd3fd472fa2eedf1f87e22e702025-08-20T03:50:53ZengElsevierIntelligent Systems with Applications2667-30532025-09-012720055610.1016/j.iswa.2025.200556A unified DNN weight compression framework using reweighted optimization methodsMengchen Fan0Tianyun Zhang1Xiaolong Ma2Jiacheng Guo3Zheng Zhan4Shanglin Zhou5Minghai Qin6Caiwen Ding7Baocheng Geng8Makan Fardad9Yanzhi Wang10The University of Alabama at Birmingham, USACleveland State University, USA; Corresponding author.Clemson University, USACleveland State University, USANortheastern University, USAUniversity of Connecticut, USAWestern Digital Research, USAUniversity of Connecticut, USAThe University of Alabama at Birmingham, USASyracuse University, USANortheastern University, USATo address the large model sizes and intensive computation requirements of deep neural networks (DNNs), weight pruning techniques have been proposed and generally fall into two categories: static regularization-based pruning and dynamic regularization-based pruning. However, the static method often leads to either complex operations or reduced accuracy, while the dynamic method requires extensive time to adjust parameters to maintain accuracy while achieving effective pruning. In this paper, we propose a unified robustness-aware framework for DNN weight pruning that dynamically updates regularization terms bounded by the designated constraint. This framework can generate both non-structured sparsity and different kinds of structured sparsity, and it incorporates adversarial training to enhance the robustness of the sparse model. We further extend our approach into an integrated framework capable of handling multiple DNN compression tasks. Experimental results show that our proposed method increases the compression rate – up to 630× for LeNet-5, 45× for AlexNet, 7.2× for MobileNet, 3.2× for ResNet-50 – while also reducing training time and simplifying hyperparameter tuning to a single penalty parameter. Additionally, our method improves model robustness by 5.07% for ResNet-18 and 3.34% for VGG-16 under a 16×pruning rate, outperforming the state-of-the-art ADMM-based hard constraint method.http://www.sciencedirect.com/science/article/pii/S2667305325000821Deep neural networksWeight pruningModel compressionAdversarial training
spellingShingle Mengchen Fan
Tianyun Zhang
Xiaolong Ma
Jiacheng Guo
Zheng Zhan
Shanglin Zhou
Minghai Qin
Caiwen Ding
Baocheng Geng
Makan Fardad
Yanzhi Wang
A unified DNN weight compression framework using reweighted optimization methods
Intelligent Systems with Applications
Deep neural networks
Weight pruning
Model compression
Adversarial training
title A unified DNN weight compression framework using reweighted optimization methods
title_full A unified DNN weight compression framework using reweighted optimization methods
title_fullStr A unified DNN weight compression framework using reweighted optimization methods
title_full_unstemmed A unified DNN weight compression framework using reweighted optimization methods
title_short A unified DNN weight compression framework using reweighted optimization methods
title_sort unified dnn weight compression framework using reweighted optimization methods
topic Deep neural networks
Weight pruning
Model compression
Adversarial training
url http://www.sciencedirect.com/science/article/pii/S2667305325000821
work_keys_str_mv AT mengchenfan aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT tianyunzhang aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT xiaolongma aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT jiachengguo aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT zhengzhan aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT shanglinzhou aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT minghaiqin aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT caiwending aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT baochenggeng aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT makanfardad aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT yanzhiwang aunifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT mengchenfan unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT tianyunzhang unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT xiaolongma unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT jiachengguo unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT zhengzhan unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT shanglinzhou unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT minghaiqin unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT caiwending unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT baochenggeng unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT makanfardad unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods
AT yanzhiwang unifieddnnweightcompressionframeworkusingreweightedoptimizationmethods