High-Dimensional Latents Should Be Diagnosed Through Phase Structure

2026

Read High-Dimensional Latents Should Be Diagnosed Through Phase Structure (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-02600.txt.

Description

Ascarate and co-authors read autoencoder latent spaces through spin-glass theory. A latent space holds the compressed codes a model learns. For a fixed decoder, reconstruction plus a hyperspherical prior forms a Hamiltonian. Latent coordinates act as continuous spins. The prior acts as an external magnetic field. This dictionary imports spin-glass diagnostics into machine learning.

Overlap distributions, susceptibility, and block-spin coarse-graining detect latent phases. Ordered, disordered, and edge-of-stability regimes become visible. Pushing latents toward the edge of stability has concrete payoffs. In generation, hyperspherical compression improves the reconstruction-generation trade-off. Tests use CIFAR-10 and CelebA64. In anomaly detection, the semi-ordered geometry improves out-of-distribution scores. Real Mars Rover and Galaxy Zoo data join the benchmarks. The authors advocate phase-aware evaluation beside standard metrics.

Catalog record

YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE88 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.02600

Browse the full library · About the Tower of Babel