Researchers have investigated whether vision models can grasp abstract concepts, similar to how language models understand abstract directions in embedding space. Their study of 46 vision encoders revealed that while models often exhibit strong "local linearity" within specific visual domains (like digits or gestures), they frequently fail to maintain a consistent "global linearity" for the same concept across different domains. This phenomenon, termed "fragmented linearity," indicates that models may learn separate representations for the same abstract idea depending on the visual context. However, the research also suggests that models with better local linearity can progress towards global linearity, allowing concept directions to be transferred as linear probes across different visual regimes. AI
IMPACT Reveals limitations in vision model generalization, suggesting a need for improved methods to achieve cross-domain conceptual understanding.
RANK_REASON Academic paper on model capabilities.
Read on Hugging Face Daily Papers →
- alphaXiv
- arXiv
- CatalyzeX
- Connected Papers
- DagsHub
- Gotit.pub
- Hugging Face
- Litmaps
- MIT-States
- ScienceCast
- Scite
- Linear Representation Hypothesis
- vision models
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
