Equivariance by Contrast: Identifiable Equivariant Embeddings from Unlabeled Finite Group Actions
Tobias Schmidt, Steffen Schneider, Matthias Bethge · 2025
Read Equivariance by Contrast: Identifiable Equivariant Embeddings from Unlabeled Finite Group Actions by Tobias Schmidt, Steffen Schneider, Matthias Bethge (2025) 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-2510-21706.txt.
Description
Tobias Schmidt, Steffen Schneider, and Matthias Bethge teach a neural encoder to respect symmetry without building the symmetry in. Many data sets transform under groups, rotations, shifts, flips. Models that internalize those actions generalize better, yet standard designs hard-wire one chosen group and cannot discover the symmetry structure from raw observation pairs.
Their method, Equivariance by Contrast, learns the embedding and the group's action together from pairs of an observation and its transformed twin. In the learned space, each group element acts as an invertible linear map, with no group-specific assumptions baked into the architecture. On shape data with combined rotations and cyclic shifts, the embeddings reproduce group operations faithfully, and synthetic tests cover non-abelian groups too. A proof guarantees the structure is identifiable, the first such general encoder-only result.
Catalog record
| CREATOR | Tobias Schmidt, Steffen Schneider, Matthias Bethge |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 123 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-21706