The Data Manifold under the Microscope

2026-06-14

AI Summary

This work addresses the gap in deep learning theory caused by the manifold hypothesis’s lack of precise characterization and reliable benchmarks for the geometric properties of data manifolds. We introduce a controllable benchmark framework that extends the dSprites and COIL-20 datasets through dense, axis-aligned sampling and employs finite-difference estimators to recover geometric quantities—such as curvature, reach, and volume—with high accuracy. This framework provides the first testbed combining known ground-truth geometry with the complexity of real-world data, enabling near-ground-truth geometric estimation even in regimes where generic estimators fail. Leveraging this platform, we calibrate and evaluate the scaling behavior of generalization bounds proposed by Genovese, Fefferman, and colleagues, thereby exposing fundamental limitations in current theoretical analyses.

Abstract

A significant gap exists between theory and practice in deep learning. Generalization and approximation error bounds are often derived for simplified models or are too loose to be informative. Many rely on the manifold hypothesis and on geometric regularity such as intrinsic dimension, curvature, and reach. Progress requires insight into data-manifold geometry and suitable benchmarks, yet existing options are polarized: analytic manifolds with known geometry but limited applicability, or real-world datasets where geometry is only coarsely estimable. We introduce a benchmarking framework for studying data geometry. We repurpose and extend dSprites and COIL-20 with additional transformation dimensions and dense, axis-aligned sampling, and pair them with finite-difference estimators that recover curvature, reach, and volume at near-ground-truth accuracy in a regime where general-purpose estimators are unreliable or difficult to deploy. The framework is intended as a controlled testbed, useful as a calibration environment for geometric estimators and a sandbox for probing theoretical assumptions. To illustrate its use, we present two application studies, namely assessing the scaling behavior of the bounds of Genovese et al. and Fefferman et al., and tracking the layer-wise geometry of a $β$-VAE, highlighting the behavior of current bounds and the value of controlled benchmarks for guiding and validating future theory. A reference implementation is available at https://github.com/koulakis/manifold-microscope.

Problem

• manifold hypothesis
• data geometry
• benchmarking
• generalization bounds
• intrinsic dimension

Innovation

• data manifold
• geometric estimation
• benchmarking framework
• finite-difference estimators
• controlled testbed