AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Sharpness-Aware Minimization Leads to Low-Rank Features

Maksym Andriushchenko, Dara Bahri, Hossein Mobahi, Nicolas Flammarion

NeurIPS 2023 · Architectures and efficiency › Attention and transformers · approach: Long-context KV and memory attention · 3 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
add SAM (ASAM variant, rho 0 -> 4.0)ResNet on CIFAR-10 · CIFAR-10feature rank: 5048 -> 4383Improved
add SAM (GAM variant, rho 0 -> 0.8)ResNet on CIFAR-10 · CIFAR-10feature rank: 7633 -> 6897Improved
increase ASAM rho from 0.0 to 1.0ResNet on CIFAR-10 · CIFAR-10test error: 7.29 -> 6.38Improved
increase ASAM rho from 1.0 to 4.0ResNet on CIFAR-10 · CIFAR-10test error: 6.38 -> 10.64Hurt
increase GAM rho from 0.0 to 0.4ResNet on CIFAR-10 · CIFAR-10test error: 4.04 -> 3.64Improved
increase GAM rho from 0.4 to 1.6ResNet on CIFAR-10 · CIFAR-10test error: 3.64 -> 4.81Hurt
train ViT-L/16 with SAM instead of standard, after 3/4 blocksViT-L/16 · ImageNet-1kfeature rank: 595 -> 314Improved
train Mixer-B/16 with SAM, after 3/4 blocksMLP-Mixer-B/16 · ImageNet-1kfeature rank: 301 -> 477Hurt
train ViT-B/16 with SAM, last blockViT-B/16 · ImageNet-1kfeature rank: 680 -> 617Improved
train Mixer-B/16 with SAM, last blockMLP-Mixer-B/16 · ImageNet-1kfeature rank: 695 -> 620Improved
add linear bottleneck layer to image encoder (imagewise retrieval) (ablation)contrastive image-text model · MS-COCO imagewise retrievalretrieval error: 23.3 -> 22.7No clear effect
add linear bottleneck layer (textwise retrieval) (ablation)contrastive image-text model · MS-COCO textwise retrievalretrieval error: 22.2 -> 22.9Hurt

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