Research

Testing whether a model fits data on curved spaces, published at ICML 2021

Plenty of real data doesn't live in flat Euclidean space: angles, rotations, directions. We built a goodness-of-fit test that works natively on that geometry, and that shows where a model is wrong, not just that it is.

← Back to Research

Our paper, "Interpretable Stein Goodness-of-fit Tests on Riemannian Manifold," was published at ICML 2021 (the 38th International Conference on Machine Learning). It's co-authored by Sentinel Assurance co-founder Wenkai Xu, together with Takeru Matsuda (University of Tokyo & RIKEN Center for Brain Science).

What the paper does

  • Extends kernel Stein discrepancy theory to build goodness-of-fit tests for distributions defined on Riemannian manifolds, such as tori and rotation groups, instead of assuming flat Euclidean data.
  • Works even when the reference model's normalizing constant is unknown or intractable to compute, the common case for realistic models on these spaces.
  • Produces an interpretable diagnostic rather than a bare pass/fail statistic: it can point to where and how a model's fit breaks down.
  • Backed by asymptotic efficiency analysis and validated on simulations and real datasets.

Our take

A model is only as trustworthy as the tests you can run against it, and most standard testing tools quietly assume the data lives in flat space. A lot of real-world data doesn't. Building tests that work on the actual geometry of the data, and that explain their verdict instead of just returning a number, is the same discipline worth bringing to verifying more complex systems: don't just say something's wrong, show where.

Want to talk to the research team?

We're always happy to discuss the work behind Sentinel, or how it applies to your AI agents.