Introduces a training-free method to visualize and validate the invariances of any feature extractor using diffusion priors.
March 24, 2026
Original Paper
Show Me What You Don't Know: Efficient Sampling from Invariant Sets for Model Validation
arXiv · 2603.21782
The Takeaway
Practitioners can now 'see' what a model is ignoring (or incorrectly grouping) without training separate generative models for each detector. This exposes critical failure modes, such as medical models failing to distinguish between standard anatomy and rare congenital reversals.
From the abstract
The performance of machine learning models is determined by the quality of their learned features. They should be invariant under irrelevant data variation but sensitive to task-relevant details. To visualize whether this is the case, we propose a method to analyze feature extractors by sampling from their fibers -- equivalence classes defined by their invariances -- given an arbitrary representative. Unlike existing work where a dedicated generative model is trained for each feature detector, o