Adaptive Gradient Methods at the Edge of Stability

• URL: http://arxiv.org/abs/2207.14484v2
• Date: Mon, 15 Apr 2024 22:52:51 GMT
• Title: Adaptive Gradient Methods at the Edge of Stability
• Authors: Jeremy M. Cohen, Behrooz Ghorbani, Shankar Krishnan, Naman Agarwal, Sourabh Medapati, Michal Badura, Daniel Suo, David Cardoze, Zachary Nado, George E. Dahl, Justin Gilmer,
• Abstract summary: We shed light on the training dynamics of adaptive gradient methods like Adam in deep learning. Our findings can serve as a foundation for the community's future understanding of adaptive gradient methods in deep learning.
• Score: 23.246757545508444
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Very little is known about the training dynamics of adaptive gradient methods like Adam in deep learning. In this paper, we shed light on the behavior of these algorithms in the full-batch and sufficiently large batch settings. Specifically, we empirically demonstrate that during full-batch training, the maximum eigenvalue of the preconditioned Hessian typically equilibrates at a certain numerical value -- the stability threshold of a gradient descent algorithm. For Adam with step size $\eta$ and $\beta_1 = 0.9$, this stability threshold is $38/\eta$. Similar effects occur during minibatch training, especially as the batch size grows. Yet, even though adaptive methods train at the ``Adaptive Edge of Stability'' (AEoS), their behavior in this regime differs in a significant way from that of non-adaptive methods at the EoS. Whereas non-adaptive algorithms at the EoS are blocked from entering high-curvature regions of the loss landscape, adaptive gradient methods at the AEoS can keep advancing into high-curvature regions, while adapting the preconditioner to compensate. Our findings can serve as a foundation for the community's future understanding of adaptive gradient methods in deep learning.

Related papers

• Adaptive Federated Learning Over the Air [108.62635460744109]
  We propose a federated version of adaptive gradient methods, particularly AdaGrad and Adam, within the framework of over-the-air model training. Our analysis shows that the AdaGrad-based training algorithm converges to a stationary point at the rate of $mathcalO( ln(T) / T 1 - frac1alpha ).
  arXiv  Detail & Related papers  (2024-03-11T09:10:37Z)
• Adaptive Strategies in Non-convex Optimization [5.279475826661643]
  An algorithm is said to be adaptive to a certain parameter if it does not need a priori knowledge of such a parameter. This dissertation presents our work on adaptive algorithms in three scenarios.
  arXiv  Detail & Related papers  (2023-06-17T06:52:05Z)
• On the SDEs and Scaling Rules for Adaptive Gradient Algorithms [45.007261870784475]
  Approxing Gradient Descent (SGD) as a Differential Equation (SDE) has allowed researchers to enjoy the benefits of studying a continuous optimization trajectory. This paper derives the SDE approximations for RMSprop and Adam, giving theoretical guarantees of correctness as well as experimental validation of their applicability.
  arXiv  Detail & Related papers  (2022-05-20T16:39:03Z)
• The Power of Adaptivity in SGD: Self-Tuning Step Sizes with Unbounded Gradients and Affine Variance [46.15915820243487]
  We show that AdaGrad-Norm exhibits an order optimal convergence of $mathcalOleft. We show that AdaGrad-Norm exhibits an order optimal convergence of $mathcalOleft.
  arXiv  Detail & Related papers  (2022-02-11T17:37:54Z)
• Adaptive Differentially Private Empirical Risk Minimization [95.04948014513226]
  We propose an adaptive (stochastic) gradient perturbation method for differentially private empirical risk minimization. We prove that the ADP method considerably improves the utility guarantee compared to the standard differentially private method in which vanilla random noise is added.
  arXiv  Detail & Related papers  (2021-10-14T15:02:20Z)
• Adam revisited: a weighted past gradients perspective [57.54752290924522]
  We propose a novel adaptive method weighted adaptive algorithm (WADA) to tackle the non-convergence issues. We prove that WADA can achieve a weighted data-dependent regret bound, which could be better than the original regret bound of ADAGRAD.
  arXiv  Detail & Related papers  (2021-01-01T14:01:52Z)
• MaxVA: Fast Adaptation of Step Sizes by Maximizing Observed Variance of Gradients [112.00379151834242]
  We propose adaptive learning rate principle, in which the running mean of squared gradient in Adam is replaced by a weighted mean, with weights chosen to maximize the estimated variance each coordinate. This results in faster adaptation, which leads more desirable empirical convergence behaviors.
  arXiv  Detail & Related papers  (2020-06-21T21:47:43Z)
• Towards Better Understanding of Adaptive Gradient Algorithms in Generative Adversarial Nets [71.05306664267832]
  Adaptive algorithms perform gradient updates using the history of gradients and are ubiquitous in training deep neural networks. In this paper we analyze a variant of OptimisticOA algorithm for nonconcave minmax problems. Our experiments show that adaptive GAN non-adaptive gradient algorithms can be observed empirically.
  arXiv  Detail & Related papers  (2019-12-26T22:10:10Z)
• On the Convergence of Adaptive Gradient Methods for Nonconvex Optimization [80.03647903934723]
  We prove adaptive gradient methods in expectation of gradient convergence methods. Our analyses shed light on better adaptive gradient methods in optimizing non understanding gradient bounds.
  arXiv  Detail & Related papers  (2018-08-16T20:25:28Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.