S2COPE discovers interpretable visual concepts from unlabeled images
burkov · x · 2026-10-31
Andriy Burkov introduces S2COPE, a method for discovering human-readable visual concepts directly from unlabeled images and testing whether those concepts support accurate, more robust image classification — addressing the interpretability limits of self-supervised image systems in scientific settings.
How it works:
- A vision-language model proposes multiple candidate descriptions per image;
- A separate image-text model scores how well each description matches its image and distinguishes it from others;
- The system prefers more informative descriptions based on these scores and repeatedly fine-tunes the proposal model.
The goal is making patterns learned from unlabeled data interpretable where driving features and categories aren't known in advance.
More from Research
- Neural Picture Alchemy: TMLR paper recasts deep learning as equations solved by backprop — CatAstro_Piyush · 2026-10-31
- KAIST ranks No.2 globally by first-author institutions at NeurIPS 2026 — JungWooHa2 · 2026-10-31
- Graph cuts' golden era in computer vision overlapped with AlexNet, not ancient history — _onionesque · 2026-10-31
- Anthropic study: experts defer to AI agents' leanings 86% of the time — akbirthko · 2026-10-31
- ChatGPT cites a 2025 paper by Kolmogorov, briefly confusing users about the famous mathematician — _onionesque · 2026-10-31
- High schooler's "Image-Space Rule Discovery" paper accepted to ACCV 2026 — HirokatuKataoka · 2026-10-31