Researchers have identified a problem with contrastive learning methods, commonly used in radiology language-image pretraining. They found that these methods lead to a degradation in zero-shot evaluation performance and fail with large batch sizes, despite decreasing training loss. The issue is hypothesized to stem from unpaired reports sharing clinical findings with an image, creating a biased target. To address this, a new method called CorLIP was developed, which corrects for target misspecification by treating text as positive when its clinical assertions align with the image. CorLIP demonstrates improved benchmark performance, eliminates post-peak degradation, enables successful large-batch training, and amplifies gains from more data. AI
IMPACT Corrects contrastive learning failures, potentially improving performance and stability in radiology AI models.
RANK_REASON The cluster contains a research paper detailing a new method for improving language-image pretraining in radiology. [lever_c_demoted from research: ic=1 ai=1.0]
- alphaXiv
- arXiv
- CatalyzeX
- CLEVR-Radiology
- CORE Recommender
- CorLIP
- DagsHub
- Gotit.pub
- Hugging Face
- Influence Flower
- ScienceCast
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →