Researchers have introduced matrix spectral functions as a broader class of objectives for dataset appraisal, encompassing the Vendi Score and determinantal point processes (DPPs). They demonstrated that these functions, along with common neural scaling law objectives, are submodular. The study also developed efficient optimization methods, reducing evaluation time by a factor of 35,000, making direct optimization of the Vendi Score feasible for large datasets like ImageNet-1K. Experiments comparing various objectives revealed that facility location performed best in predicting held-out test performance, while the Vendi Score could become a poor proxy at higher values. The research also highlighted that random subsets are surprisingly concentrated in appraisal scores and performance, and that data value is determined by more than just size, class balance, and training budget. AI
IMPACT Introduces a more efficient and comprehensive framework for dataset evaluation, potentially improving model training and performance prediction.
RANK_REASON This is a research paper detailing a new theoretical framework and computational methods for evaluating datasets. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
