Browse Teach Curate

Equidensity Scores and the Eden Metric for Generative-Model Evaluation

Why many popular goodness-of-fit scores for synthetic-vs-real data distributions (correlation, Jaccard, earth-mover, KL) systematically overstate quality: they weight every datapoint equally, which the authors call an equipoint score. Students learn the alternative equidensity scoring principle, its first instance the Eden score, how it tracks human perception of fit and avoids this grade inflation, and its connection to Renyi entropy of negative order.

KITT Text

Grade Inflation in Generative Models

Generative models hold great potential, but only if one can trust the evaluation of the data they generate. We show that many commonly used quality scores for comparing two-dimensional distributions …