Timezone: »
Spotlight
Understanding Gradient Descent on the Edge of Stability in Deep Learning
Sanjeev Arora · Zhiyuan Li · Abhishek Panigrahi
Deep learning experiments by \citet{cohen2021gradient} using deterministic Gradient Descent (GD) revealed an {\em Edge of Stability (EoS)} phase when learning rate (LR) and sharpness (\emph{i.e.}, the largest eigenvalue of Hessian) no longer behave as in traditional optimization. Sharpness stabilizes around $2/$LR and loss goes up and down across iterations, yet still with an overall downward trend. The current paper mathematically analyzes a new mechanism of implicit regularization in the EoS phase, whereby GD updates due to non-smooth loss landscape turn out to evolve along some deterministic flow on the manifold of minimum loss. This is in contrast to many previous results about implicit bias either relying on infinitesimal updates or noise in gradient. Formally, for any smooth function $L$ with certain regularity condition, this effect is demonstrated for (1) {\em Normalized GD}, i.e., GD with a varying LR $\eta_t =\frac{\eta}{\norm{\nabla L(x(t))}}$ and loss $L$; (2) GD with constant LR and loss $\sqrt{L- \min_x L(x)}$. Both provably enter the Edge of Stability, with the associated flow on the manifold minimizing $\lambda_{1}(\nabla^2 L)$. The above theoretical results have been corroborated by an experimental study.
Author Information
Sanjeev Arora (Princeton University)
Zhiyuan Li (Princeton University)
Abhishek Panigrahi (Princeton University)
Related Events (a corresponding poster, oral, or spotlight)
-
2022 Poster: Understanding Gradient Descent on the Edge of Stability in Deep Learning »
Tue. Jul 19th through Wed the 20th Room Hall E #1219
More from the Same Authors
-
2023 : The Marginal Value of Momentum for Small Learning Rate SGD »
Runzhe Wang · Sadhika Malladi · Tianhao Wang · Kaifeng Lyu · Zhiyuan Li -
2023 : Sharpness Minimization Algorithms Do Not Only Minimize Sharpness To Achieve Better Generalization »
Kaiyue Wen · Tengyu Ma · Zhiyuan Li -
2023 : Chain of Thought Empowers Transformers to Solve Inherently Serial Problems »
Zhiyuan Li · Hong Liu · Denny Zhou · Tengyu Ma -
2023 : Fine-Tuning Language Models with Just Forward Passes »
Sadhika Malladi · Tianyu Gao · Eshaan Nichani · Jason Lee · Danqi Chen · Sanjeev Arora -
2023 : 🎤 Fine-Tuning Language Models with Just Forward Passes »
Sadhika Malladi · Tianyu Gao · Eshaan Nichani · Alex Damian · Jason Lee · Danqi Chen · Sanjeev Arora -
2023 : High-dimensional Optimization in the Age of ChatGPT, Sanjeev Arora »
Sanjeev Arora -
2023 Poster: Task-Specific Skill Localization in Fine-tuned Language Models »
Abhishek Panigrahi · Nikunj Saunshi · Haoyu Zhao · Sanjeev Arora -
2023 Poster: A Kernel-Based View of Language Model Fine-Tuning »
Sadhika Malladi · Alexander Wettig · Dingli Yu · Danqi Chen · Sanjeev Arora -
2022 : On the SDEs and Scaling Rules for Adaptive Gradient Algorithms »
Sadhika Malladi · Kaifeng Lyu · Abhishek Panigrahi · Sanjeev Arora -
2022 : Implicit Bias of Gradient Descent on Reparametrized Models: On Equivalence toMirror Descent »
Zhiyuan Li · Tianhao Wang · Jason Lee · Sanjeev Arora -
2022 Poster: Understanding Contrastive Learning Requires Incorporating Inductive Biases »
Nikunj Umesh Saunshi · Jordan Ash · Surbhi Goel · Dipendra Kumar Misra · Cyril Zhang · Sanjeev Arora · Sham Kakade · Akshay Krishnamurthy -
2022 Poster: Robust Training of Neural Networks Using Scale Invariant Architectures »
Zhiyuan Li · Srinadh Bhojanapalli · Manzil Zaheer · Sashank Jakkam Reddi · Sanjiv Kumar -
2022 Spotlight: Understanding Contrastive Learning Requires Incorporating Inductive Biases »
Nikunj Umesh Saunshi · Jordan Ash · Surbhi Goel · Dipendra Kumar Misra · Cyril Zhang · Sanjeev Arora · Sham Kakade · Akshay Krishnamurthy -
2022 Oral: Robust Training of Neural Networks Using Scale Invariant Architectures »
Zhiyuan Li · Srinadh Bhojanapalli · Manzil Zaheer · Sashank Jakkam Reddi · Sanjiv Kumar -
2021 Poster: Risk Bounds and Rademacher Complexity in Batch Reinforcement Learning »
Yaqi Duan · Chi Jin · Zhiyuan Li -
2021 Spotlight: Risk Bounds and Rademacher Complexity in Batch Reinforcement Learning »
Yaqi Duan · Chi Jin · Zhiyuan Li -
2020 Poster: Provable Representation Learning for Imitation Learning via Bi-level Optimization »
Sanjeev Arora · Simon Du · Sham Kakade · Yuping Luo · Nikunj Umesh Saunshi -
2020 Poster: InstaHide: Instance-hiding Schemes for Private Distributed Learning »
Yangsibo Huang · Zhao Song · Kai Li · Sanjeev Arora -
2020 Poster: A Sample Complexity Separation between Non-Convex and Convex Meta-Learning »
Nikunj Umesh Saunshi · Yi Zhang · Mikhail Khodak · Sanjeev Arora -
2019 : Is Optimization a sufficient language to understand Deep Learning? »
Sanjeev Arora -
2019 Poster: A Theoretical Analysis of Contrastive Unsupervised Representation Learning »
Nikunj Umesh Saunshi · Orestis Plevrakis · Sanjeev Arora · Mikhail Khodak · Hrishikesh Khandeparkar -
2019 Oral: A Theoretical Analysis of Contrastive Unsupervised Representation Learning »
Nikunj Umesh Saunshi · Orestis Plevrakis · Sanjeev Arora · Mikhail Khodak · Hrishikesh Khandeparkar -
2019 Poster: Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networks »
Sanjeev Arora · Simon Du · Wei Hu · Zhiyuan Li · Ruosong Wang -
2019 Oral: Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networks »
Sanjeev Arora · Simon Du · Wei Hu · Zhiyuan Li · Ruosong Wang -
2018 Poster: Stronger Generalization Bounds for Deep Nets via a Compression Approach »
Sanjeev Arora · Rong Ge · Behnam Neyshabur · Yi Zhang -
2018 Oral: Stronger Generalization Bounds for Deep Nets via a Compression Approach »
Sanjeev Arora · Rong Ge · Behnam Neyshabur · Yi Zhang -
2018 Poster: On the Optimization of Deep Networks: Implicit Acceleration by Overparameterization »
Sanjeev Arora · Nadav Cohen · Elad Hazan -
2018 Oral: On the Optimization of Deep Networks: Implicit Acceleration by Overparameterization »
Sanjeev Arora · Nadav Cohen · Elad Hazan -
2018 Tutorial: Toward Theoretical Understanding of Deep Learning »
Sanjeev Arora -
2017 Poster: Generalization and Equilibrium in Generative Adversarial Nets (GANs) »
Sanjeev Arora · Rong Ge · Yingyu Liang · Tengyu Ma · Yi Zhang -
2017 Talk: Generalization and Equilibrium in Generative Adversarial Nets (GANs) »
Sanjeev Arora · Rong Ge · Yingyu Liang · Tengyu Ma · Yi Zhang