Researchers have developed SurrogateSHAP, a novel framework designed to efficiently attribute contributions to data contributors in text-to-image models. This method avoids the computationally expensive process of retraining models for each data subset by using inference from a pretrained model. SurrogateSHAP employs a gradient-boosted tree to approximate the utility function and analytically derive Shapley values, significantly reducing computational overhead while identifying influential data sources. The framework has been validated across various tasks, including image quality, aesthetics, and product diversity, and shows promise for auditing safety-critical generative models. AI
IMPACT Provides a scalable method for valuing data contributors and auditing generative models, potentially impacting data marketplaces and model safety.
RANK_REASON This is a research paper detailing a new methodology for attribution in generative models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
