Unsupervised derivation of a perceptual metric from unlabeled data

Develop a perceptual distance function that can be derived solely from unlabeled data, without reliance on human-labeled judgments or supervised training, and that provides a mathematically interpretable measure of similarity for natural signals and images.

Background

The paper motivates the need for a perceptual distance measure that does not depend on costly and noisy human annotations. While many supervised metrics exist, their reliance on labeled perceptual judgments limits interpretability and scalability. The authors frame the challenge of obtaining a perceptual metric directly from unlabeled data as a longstanding problem and introduce the Information-Estimation Metric (IEM) as an unsupervised approach grounded in information-estimation relationships.

References

Deriving a perceptual metric solely based on unlabeled data remains a fundamental open problem of both scientific and practical importance.

— Learning a distance measure from the information-estimation geometry of data  (2510.02514 - Ohayon et al., 2 Oct 2025) in Section 1 (Introduction)

A good metric for image generation is still an open problem, and we hope our results on ImageNet encourage more work on it.

— Sphere Encoder 2  (2610.02208 - Yue et al., 1 Oct 2026) in Appendix, Section 2.1, “Distribution Shift”