Uncertainty aware validation improves generative model trustworthiness
An end-to-end framework for uncertainty-aware validation of generative models
Machine Learning
Summary
It's hard to know how much to trust generative models, which create data based on patterns they learn, because current tests treat these models as if they were perfect. The authors propose a method that measures uncertainty in these models' predictions, so tests can distinguish between real errors and uncertainty from limited data. This way, the model is only rejected when it truly fails, not just because the test is too sensitive. They show this method works on both simple examples and real physics problems, allowing more confident use of these models.
What this means in practice
- •For physics data analysts: Validate generative background models with uncertainty estimates to reduce false rejections and improve analysis reliability in collider experiments.
- •For quality assurance teams: Incorporate uncertainty-aware testing of generative models to distinguish true errors from sampling noise when evaluating model performance on large datasets.
Authors
Giada Badaracco, Sean Benevedes, Christina Reissel, Philip Harris, Gaia Grosso
Abstract
Density estimation with generative models is a common task in the physical sciences, for example in background estimation in particle physics. Existing machine learning techniques, however, lack methods to quantify uncertainties that can be propagated into a statistical test. That limits the validation of generative models to standard goodness-of-fit tests, in which the generative model is treated as exact (point-null hypothesis). But a model fitted to a finite training set is never exact, so if its uncertainties are neglected, any sufficiently powerful goodness-of-fit test will eventually reject it once the test set is large enough to resolve this estimation error. We propose SIGMA, an end-to-end framework combining density estimation with principled uncertainty quantification. We use Gaussian mixture models, for which uncertainties are tractable, and further construct a new goodness-of-fit test in which these uncertainties are profiled as nuisance parameters to account for the statistical error of the fit, promoting the conventional point-null test to a composite one. A rejection then signals genuine mismodeling rather than a test set large enough to resolve the error of the fit, provided the propagated uncertainties faithfully cover that error. We demonstrate the framework on a controlled two-dimensional toy model and on a collider physics benchmark modeling the QCD background process. With uncertainties included, the same model can be trusted on test sets up to an order of magnitude larger than the standard test allows. Our work thus provides a statistically grounded basis for a robust use of generative models in the physical sciences.