When Will Gradient Regularization Be Harmful?

Yang Zhao, Hao Zhang, Xiuyuan Hu
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:61144-61158, 2024.

Abstract

Gradient regularization (GR), which aims to penalize the gradient norm atop the loss function, has shown promising results in training modern over-parameterized deep neural networks. However, can we trust this powerful technique? This paper reveals that GR can cause performance degeneration in adaptive optimization scenarios, particularly with learning rate warmup. Our empirical and theoretical analyses suggest this is due to GR inducing instability and divergence in gradient statistics of adaptive optimizers at the initial training stage. Inspired by the warmup heuristic, we propose three GR warmup strategies, each relaxing the regularization effect to a certain extent during the warmup course to ensure the accurate and stable accumulation of gradients. With experiments on Vision Transformer family, we confirm the three GR warmup strategies can effectively circumvent these issues, thereby largely improving the model performance. Meanwhile, we note that scalable models tend to rely more on the GR warmup, where the performance can be improved by up to 3% on Cifar10 compared to baseline GR. Code is available at https://github.com/zhaoyang-0204/gnp.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-zhao24t, title = {When Will Gradient Regularization Be Harmful?}, author = {Zhao, Yang and Zhang, Hao and Hu, Xiuyuan}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {61144--61158}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/zhao24t/zhao24t.pdf}, url = {https://proceedings.mlr.press/v235/zhao24t.html}, abstract = {Gradient regularization (GR), which aims to penalize the gradient norm atop the loss function, has shown promising results in training modern over-parameterized deep neural networks. However, can we trust this powerful technique? This paper reveals that GR can cause performance degeneration in adaptive optimization scenarios, particularly with learning rate warmup. Our empirical and theoretical analyses suggest this is due to GR inducing instability and divergence in gradient statistics of adaptive optimizers at the initial training stage. Inspired by the warmup heuristic, we propose three GR warmup strategies, each relaxing the regularization effect to a certain extent during the warmup course to ensure the accurate and stable accumulation of gradients. With experiments on Vision Transformer family, we confirm the three GR warmup strategies can effectively circumvent these issues, thereby largely improving the model performance. Meanwhile, we note that scalable models tend to rely more on the GR warmup, where the performance can be improved by up to 3% on Cifar10 compared to baseline GR. Code is available at https://github.com/zhaoyang-0204/gnp.} }
Endnote
%0 Conference Paper %T When Will Gradient Regularization Be Harmful? %A Yang Zhao %A Hao Zhang %A Xiuyuan Hu %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-zhao24t %I PMLR %P 61144--61158 %U https://proceedings.mlr.press/v235/zhao24t.html %V 235 %X Gradient regularization (GR), which aims to penalize the gradient norm atop the loss function, has shown promising results in training modern over-parameterized deep neural networks. However, can we trust this powerful technique? This paper reveals that GR can cause performance degeneration in adaptive optimization scenarios, particularly with learning rate warmup. Our empirical and theoretical analyses suggest this is due to GR inducing instability and divergence in gradient statistics of adaptive optimizers at the initial training stage. Inspired by the warmup heuristic, we propose three GR warmup strategies, each relaxing the regularization effect to a certain extent during the warmup course to ensure the accurate and stable accumulation of gradients. With experiments on Vision Transformer family, we confirm the three GR warmup strategies can effectively circumvent these issues, thereby largely improving the model performance. Meanwhile, we note that scalable models tend to rely more on the GR warmup, where the performance can be improved by up to 3% on Cifar10 compared to baseline GR. Code is available at https://github.com/zhaoyang-0204/gnp.
APA
Zhao, Y., Zhang, H. & Hu, X.. (2024). When Will Gradient Regularization Be Harmful?. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:61144-61158 Available from https://proceedings.mlr.press/v235/zhao24t.html.

Related Material