İçeriğe geç
CANLI Dün Yerli yapay zeka İridaa, kendi modeli “Kuzey”i devreye aldı: Tahmini skorlar rakiplerine göre nerede? 6 gün önce Bu Hafta AI Model Değerlendirmesi (2026-W46) 8 Kas 2026 OpenAI Uyarı Vermeden Hesap Kapatıyor: Aylık 200 Dolarlık Pro Aboneler Bile Etkilendi 8 Kas 2026 Yapay Zekâ 18 Yıllık Kuantum Varsayımını Üç Günde Çözdü: Aaronson-Ambainis Artık Bir Teorem 8 Kas 2026 Gemini 4 Flash ve Yeni Claude İçin İşaretler Artıyor: LMArena’da ‘Hydrogen’, Sızıntıda ‘Hyperion’ Zekâ endeksi lideri Claude Opus 5 (Adaptive Reasoning, Max Effort) · 63,1 16.11.2026
Bülten

arXiv 2606.03148 cs.CV

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

  • Sreehari Rammohan
  • Huy Ha
  • Carl Vondrick

01 30 saniyede editör özeti

Robust visual classification often depends on localizing the main foreground objects in an image while ignoring contextual distractors.

02 Özet

Özetle: Robust visual classification often depends on localizing the main foreground objects in an image while ignoring contextual distractors.

Özet

Robust visual classification often depends on localizing the main foreground objects in an image while ignoring contextual distractors. Surprisingly, we find that the last layer 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.

Orijinal Özet (İngilizce)

Robust visual classification often depends on localizing the main foreground objects in an image while ignoring contextual distractors. Surprisingly, we find that the last layer 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.

Kaynak: arXiv:2606.03148 · PDF

Robust visual classification often depends on localizing the main foreground objects in an image while ignoring contextual distractors. Surprisingly, we find that the last layer 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.

03 PDF

Bu PDF başka bir sitede (arxiv.org) barındırılıyor ve sayfa içinde gösterilemiyor. Yukarıdaki düğmelerle yeni sekmede açabilir ya da indirebilirsiniz.

04 Kaynak göster

BibTeX
@article{rammohan2026a,
  title   = {$A^2$: Smaller Self-Supervised ViTs Localize Better than Larger Ones},
  author  = {Sreehari Rammohan and Huy Ha and Carl Vondrick},
  journal = {arXiv preprint arXiv:2606.03148},
  year    = {2026},
  url     = {https://arxiv.org/abs/2606.03148}
}
APA
Rammohan, S., Ha, H., & Vondrick, C. (2026). $A^2$: Smaller Self-Supervised ViTs Localize Better than Larger Ones. arXiv. https://arxiv.org/abs/2606.03148
Paylaş
X LinkedIn WhatsApp Telegram Reddit Bluesky Threads E-posta

Yorumlar 0

Yorum yazın

Karakter: 0/1000

Henüz yorum yok. İlk yorumu siz yazın.