Pith. sign in

REVIEW 7 cited by

Surrogate Gap Minimization Improves Sharpness-Aware Training

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2203.08065 v2 pith:UMLRTMTL submitted 2022-03-15 cs.LG cs.AI

classification cs.LGcs.AI
keywords lossgsamperturbedsurrogategeneralizationtextbftextitimproves
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

The recently proposed Sharpness-Aware Minimization (SAM) improves generalization by minimizing a \textit{perturbed loss} defined as the maximum loss within a neighborhood in the parameter space. However, we show that both sharp and flat minima can have a low perturbed loss, implying that SAM does not always prefer flat minima. Instead, we define a \textit{surrogate gap}, a measure equivalent to the dominant eigenvalue of Hessian at a local minimum when the radius of the neighborhood (to derive the perturbed loss) is small. The surrogate gap is easy to compute and feasible for direct minimization during training. Based on the above observations, we propose Surrogate \textbf{G}ap Guided \textbf{S}harpness-\textbf{A}ware \textbf{M}inimization (GSAM), a novel improvement over SAM with negligible computation overhead. Conceptually, GSAM consists of two steps: 1) a gradient descent like SAM to minimize the perturbed loss, and 2) an \textit{ascent} step in the \textit{orthogonal} direction (after gradient decomposition) to minimize the surrogate gap and yet not affect the perturbed loss. GSAM seeks a region with both small loss (by step 1) and low sharpness (by step 2), giving rise to a model with high generalization capabilities. Theoretically, we show the convergence of GSAM and provably better generalization than SAM. Empirically, GSAM consistently improves generalization (e.g., +3.2\% over SAM and +5.4\% over AdamW on ImageNet top-1 accuracy for ViT-B/32). Code is released at \url{ https://sites.google.com/view/gsam-iclr22/home}.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. OmniOpt: Taxonomy, Geometry, and Benchmarking of Modern Optimizers

    cs.LG 2026-07 conditional novelty 6.0 of 10

    A meta-pipeline plus LMO four-axis view yields a dual taxonomy of 108 optimizers, and a multi-objective LLM/vision benchmark shows no single family dominates the quality–cost–memory frontier.

  2. LightSAM: Parameter-Agnostic Sharpness-Aware Minimization

    cs.LG 2025-05 reject novelty 6.0 of 10

    An adaptive SAM variant using AdaGrad and Adam steps for both perturbation and update is claimed to converge at O(ln T / T^{1/4}) without tuning, but the Adam version still needs decaying hyperparameters and the proof...

  3. Gradient-Energy Guided Block-Wise Perturbations for Sharpness-Aware Minimization

    cs.LG 2026-07 conditional novelty 5.0 of 10

    GEAR-SAM re-allocates SAM's fixed perturbation radius across network blocks in proportion to an EMA of squared block-gradient norms, improving generalization on CIFAR, transfer, and label-noise benchmarks.

  4. GLAD: Generalizable Tuning for Vision-Language Models

    cs.CV 2025-07 conditional novelty 5.0 of 10

    GLAD improves few-shot CLIP generalization by fusing original and sharpness-aware gradients during LoRA tuning and adding image-conditional text alignment.

  5. Adversarial Data Augmentation for Single Domain Generalization via Lyapunov Exponent-Guided Optimization

    cs.CV 2025-07 reject novelty 5.0 of 10

    LEAwareSGD modulates the learning rate with a Lyapunov exponent estimate and claims state-of-the-art accuracy on three single-domain generalization benchmarks, but the method is underspecified and its hyperparameters ...

  6. C-Flat++: Towards a More Efficient and Powerful Framework for Continual Learning

    cs.LG 2025-08 conditional novelty 4.0 of 10

    Adding zeroth- and first-order flatness penalties to continual learning losses yields small consistent accuracy gains across seven methods, with the gated C-Flat++ variant at roughly 30% of the update cost.

  7. Domain-Generalization to Improve Learning in Meta-Learning Algorithms

    cs.LG 2025-08 reject novelty 4.0 of 10

    DGS-MAML layers gradient matching onto SharpMAML and claims O(1/T) convergence and tighter PAC-Bayes bounds, but the displayed theorems give O(1/sqrt T) under the paper's own parameter choices.

Pith tools