New Directional Influence Function improves data attribution in constrained learning

Researchers have developed a new method called the Directional Influence Function (DIF) to accurately estimate the impact of individual training data points on machine learning models, particularly in constrained learning scenarios. Unlike traditional influence functions, DIF explicitly accounts for the constraints imposed during training, such as fairness or safety requirements. The new method has been validated on tasks like fairness-constrained CNNs and constrained linear regression, demonstrating its effectiveness in predicting changes in model performance and aligning closely with retraining results. AI

IMPACT Provides a more reliable tool for understanding model behavior and identifying influential data points in complex, constrained learning systems.

RANK_REASON Academic paper introducing a novel methodology. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• CNNs
• Directional Influence Function
• Influence Function

• paper
• safety

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