New research challenges strong proxy-based explanations for AI model performance

A new paper published on arXiv by Hongmin Li presents a controlled counterexample to strong proxy-based explanations of out-of-distribution (OOD) performance in machine learning. The research demonstrates that a proxy ranking of pretraining datasets may not align with their actual performance on downstream tasks, even when using a fixed pretraining and probing setup. This finding suggests that while structure-based explanations are not entirely invalidated, there are limitations to their strength and applicability in predicting OOD accuracy. AI

IMPACT Challenges the reliability of certain AI model explanation techniques, suggesting caution in their application for predicting out-of-distribution performance.

RANK_REASON The cluster contains a research paper published on arXiv detailing a theoretical and experimental counterexample to a specific machine learning explanation method. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX Code Finder for Papers
• CORE Recommender
• DagsHub
• Gotit.pub
• Hongmin Li
• Hugging Face
• IArxiv Recommender
• ScienceCast

• paper
• other

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