Benchmarking Framework Quantifies Data Manifold Geometry Using Extended dSprites and COIL-20
June 13, 2026
A new benchmarking framework extends dSprites and COIL-20 with additional transformation dimensions and dense axis-aligned sampling, paired with finite-difference estimators for intrinsic dimension, curvature, and reach. It targets the gap between analytic manifolds with known geometry and real datasets where geometry is only coarsely estimable.
HOW THIS AFFECTS YOU
●
researcherThis gives you controlled benchmarks with ground-truth geometric properties to test whether generalization bounds relying on manifold assumptions actually hold in practice.