Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

$A^2$: Smaller Self-Supervised ViTs Localize Better than Larger Ones

About

Robust visual classification often depends on localizing the main foreground objects in an image while ignoring contextual distractors. Surprisingly, we find that the attention maps of smaller self-supervised ViTs localize foreground objects better than those of larger ViTs. However, we still need large ViTs, because they extract richer representations from each patch. To get the best of both worlds, good localization and rich representations, we propose $A^2$, a simple method that leverages this inverse scaling finding by decoupling where to look (a small attention model) from what to extract (a large embedding model): we crop around the attention peaks of a small model and embed the crops with a larger model. $A^2$ uses entirely pretrained features, requires no group labels, and does not require per-dataset attention or backbone training. Across 5 benchmarks, $A^2$ is competitive with backbone-matched loss-level methods like DFR, and outperforms end-to-end attention training under stronger distribution shifts.

Sreehari Rammohan, Huy Ha, Carl Vondrick• 2026

Related benchmarks

TaskDatasetResultRank
Image ClassificationWaterbirds
WG Accuracy94.4
283
Image ClassificationWaterbirds (test)
Worst-Group Accuracy94.4
214
Image ClassificationSpawrious O2O-Hard
Accuracy95.2
47
Image ClassificationSpawrious M2M-Hard
Accuracy (M2M-Hard)87.1
47
Image ClassificationMetaShift Cat vs. Dog 18 (test)
Test Accuracy77.3
22
Domain GeneralizationMetaShift Animals 8-class (test)
Test Accuracy88.9
22
Binary ClassificationWaterbirds (OOD)
Accuracy98.2
18
Image ClassificationSpawrious (test)--
15
Image ClassificationSpawrious M2M
WGA85.4
13
O.O.D. ClassificationSpawrious M2M
Accuracy94
13
Showing 10 of 17 rows

Other info

Follow for update