New MaRS method improves out-of-distribution detection for foundation models

Researchers have developed a new method called MaRS (Mahalanobis Residual Scoring) for detecting out-of-distribution (OOD) data in foundation models, particularly for medical imaging. Unlike previous methods that struggled with distribution shifts, MaRS uses a lightweight autoencoder to learn an in-distribution manifold and measures deviations using Mahalanobis distance on reconstruction residuals. This approach yields variance-aware OOD scores and has demonstrated superior performance across various imaging modalities and model types compared to existing baselines. AI

IMPACT Enhances the reliability of foundation models in critical applications like medical imaging by improving out-of-distribution detection.

RANK_REASON The cluster describes a new research paper detailing a novel method for out-of-distribution detection. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• arXiv
• autoencoder
• CatalyzeX
• DagsHub
• foundation models
• Francesco Di Salvo
• Hugging Face
• k-nearest neighbors algorithm
• Mahalanobis Residual Scoring
• MaRS
• medical images

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →