Brief

last 24h

[2/2] 221 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

TOOL · arXiv stat.ML English(EN) · 1w

On the Accuracy of Newton Step and Influence Function Data Attributions

Researchers have developed a new mathematical analysis for data attribution methods like Influence Functions (IF) and Newton Step (NS) in convex learning problems. This analysis does not rely on strong convexity assumptions and provides tighter bounds, addressing limitations of previous work. The findings explain why NS data attribution is often more accurate than IF and establish asymptotic scaling laws for their errors. AI

IMPACT Provides a more accurate understanding of data attribution methods, crucial for model interpretability and responsible AI development.
RESEARCH · arXiv cs.CL English(EN) · 6d · [2 sources]

CLIF: Concept-Level Influence Functions for Transparent Bottleneck Models

Researchers have developed CLIF, a new method using influence functions to improve the interpretability of NLP models. This approach can identify influential training data points, both beneficial and detrimental, and allows for performance restoration without retraining by adjusting these samples. CLIF also analyzes concept-level influences within Concept Bottleneck Models, offering insights into decision-making processes. AI

IMPACT Enhances transparency in AI models, potentially enabling wider adoption in sensitive domains like finance and medicine.