Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Adnan, Mohammed, Jain, Rohan, Jacobs, Tom, Sharma, Ekansh, Krishnan, Rahul G., Burkholz, Rebekka, Ioannou, Yani
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2605.27541
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917537969602560
author Adnan, Mohammed
Jain, Rohan
Jacobs, Tom
Sharma, Ekansh
Krishnan, Rahul G.
Burkholz, Rebekka
Ioannou, Yani
author_facet Adnan, Mohammed
Jain, Rohan
Jacobs, Tom
Sharma, Ekansh
Krishnan, Rahul G.
Burkholz, Rebekka
Ioannou, Yani
contents Dynamic Sparse Training (DST) methods train neural networks by maintaining sparsity while dynamically adapting the network topology. Despite the promise of reduced computation, DST methods converge significantly slower than dense training, often requiring comparable training time to achieve similar accuracy. We demonstrate both analytically and empirically that Batch Normalization (BN) adversely affects sparse training, and propose SparseOpt, a sparsity-aware optimizer, to address this. Experiments on ResNet models across CIFAR-100 and ImageNet demonstrate consistently faster convergence and improved generalization with our proposed method. Our work highlights the limitations of current normalization layers in sparse training and provides the first systematic study of the interaction between Batch Normalization, sparse layers, and DST, taking a significant step toward making DST practically competitive with dense training.
format Preprint
id arxiv_https___arxiv_org_abs_2605_27541
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SparseOpt: Addressing Normalization-induced Gradient Skew in Sparse Training
Adnan, Mohammed
Jain, Rohan
Jacobs, Tom
Sharma, Ekansh
Krishnan, Rahul G.
Burkholz, Rebekka
Ioannou, Yani
Machine Learning
Dynamic Sparse Training (DST) methods train neural networks by maintaining sparsity while dynamically adapting the network topology. Despite the promise of reduced computation, DST methods converge significantly slower than dense training, often requiring comparable training time to achieve similar accuracy. We demonstrate both analytically and empirically that Batch Normalization (BN) adversely affects sparse training, and propose SparseOpt, a sparsity-aware optimizer, to address this. Experiments on ResNet models across CIFAR-100 and ImageNet demonstrate consistently faster convergence and improved generalization with our proposed method. Our work highlights the limitations of current normalization layers in sparse training and provides the first systematic study of the interaction between Batch Normalization, sparse layers, and DST, taking a significant step toward making DST practically competitive with dense training.
title SparseOpt: Addressing Normalization-induced Gradient Skew in Sparse Training
topic Machine Learning
url https://arxiv.org/abs/2605.27541