AI Paper Atlas

Research map / Optimization / Implicit bias and generalization

When Do Flat Minima Optimizers Work?

Jean Kaddour, Linqing Liu, Ricardo Silva, Matt Kusner

NeurIPS 2022 · Optimization › Implicit bias and generalization · approach: Sharpness-aware minimization theory · 6 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Average SAM iterates to form WASAMSAM · 42 benchmark experimentsgeneralization performance: 39/42 casesImproved
Use SAM instead of the non-flat baselineNLP tasks · NLP experimentstask performance: 7/10 experimentsImproved
Use SWA instead of the non-flat baselineNLP tasks · NLP experimentstask performance: 1/10; 4 tasksMixed
Use SWA instead of the non-flat baselineGraph representation learning · GRL experimentstask performance: 10/12 experimentsImproved
Use SAM instead of the non-flat baselineGraph representation learning · GRL experimentstask performance: 4/12 experimentsMixed
Use flat-minima optimizers instead of the baselineSAGE or DGCN · OGB-ProteinsROC-AUCNo clear effect
Use flat-minima optimizers instead of the baselineSAGE or DGCN · OGB-ProductsaccuracyImproved
Replace GCN layers with SAGE layersCitation2 link prediction · OGB-Citation2MRR: >1.15%; -0.31%Mixed
Average SAM iterates to form WASAMSAM · WRN on CIFAR100 and GIN on Code2median dominant Hessian eigenvalue: 673 -> 117; 16.65 -> 9.96Improved
Use SAM with AdamW instead of the AdamW baselineResNet34 · CIFAR100accuracy: -2.29Hurt
Switch the base optimizer from Adam to RMSprop (ablation)SWA and SAM · GIN on Code2F1: +0.83_{7 0.11} -> +0.67; +0.57_{7 0.09} -> +0.70Mixed
Use data augmentation rather than none (ablation)SWA and SAM · ResNet18 on CIFAR100accuracy improvement: None: +2.23, +2.20; Basic: +0.13, +0.74; AA: +3.03, +3.35Mixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy