Automated Symmetry Discovery for Equivariant Learning

This direction is merged into another direction: automated symmetry discovery. Its last version stays readable and citable here.

Version 1 of 121 papers24 conceptsfrozen 2026-10-04

In machine learning, symmetry is often described as a "cheat code." If a model knows that a physical system is invariant to rotation or that a robot's gait is symmetric across its left and right sides, it can achieve higher accuracy with significantly less data. Traditionally, these symmetries had to be hard-coded by human experts who understood the underlying group theory of the problem.

A growing body of research is shifting this paradigm from hard-coding symmetries to discovering them directly from raw data. By treating the symmetry group itself as a learnable parameter, researchers are building models that can "sense" the laws of physics or the geometric constraints of a dataset without prior instruction.

The Continuous Path: Learning Lie Algebras

Most physical symmetries—like rotating a sphere or sliding an object across a table—are continuous. These are mathematically described by Lie groups. The core intuition behind modern discovery frameworks is that we don't need to learn the entire group at once; we only need to learn its "infinitesimal generators," or its Lie algebra.

In Generative Adversarial Symmetry Discovery, Jianke Yang, Robin Walters, and their colleagues introduced LieGAN. The framework uses a generator to propose transformations and a discriminator to check if the data distribution remains invariant under those transformations. By parameterizing the generator with a Lie algebra basis, the model can discover complex equivariances like Lorentz transformations in particle physics data.

Similarly, Lexiang Hu and Zhouchen Lin developed LieSD (Symmetry Discovery for Different Data Types), which identifies these generators by analyzing the gradients of a pre-trained network. The intuition here is that if a network has already learned a task, the directions in which its predictions don't change (the "null space" of its sensitivity) reveal the underlying symmetries of the task.

The Discrete Gap: Reflections and Parity

While Lie algebras are powerful for smooth motion, they have a fundamental blind spot: they cannot represent discrete symmetries. A reflection (like a mirror image) or a parity flip is a "jump" that cannot be reached by a continuous flow from the identity.

This limitation is a recurring theme in the literature. As noted by Gyeonghoon Ko and Juho Lee in Learning Infinitesimal Generators of Continuous Symmetries from Data, continuous vector fields are inherently unable to capture transformations that lack a path to the identity.

To bridge this gap, researchers like Prathosh AP have proposed hybrid strategies. In A Unified Framework for Discovering Discrete Symmetries, they combine deep learning with multi-armed bandits to search through potential discrete group structures, such as cyclic or dihedral patterns, while simultaneously learning the target function.

From Global to Local Symmetries

Most early work assumed that symmetry is a global property—that the entire dataset obeys the same rotation or translation rules. However, real-world data is often more complex. A climate model might have different symmetries at the poles than at the equator; a robot might have symmetries that only apply to specific joints.

Manu Bhat, Nima Dehmamy, and Robin Walters addressed this with AtlasD (AtlasD: Automatic Local Symmetry Discovery). By formalizing symmetry as "atlas equivariance," they allow a model to discover symmetries within local neighborhoods of a manifold. Crucially, AtlasD uses discrete coset representatives to capture orientation-reversing transformations (like reflections) alongside continuous ones, providing a more complete picture of the data's geometric structure.

Why Discovery Works: Sample Efficiency and Constraints

The ultimate goal of this research is sample efficiency. In Adaptive Symmetry Discovery for Dynamical System Identification, Behrooz Tahmasebi and Melanie Weber demonstrate that identifying hidden symmetries allows a model to learn a dynamical system from much shorter trajectories.

The intuition is simple: if you know a system is symmetric, every data point you collect actually gives you multiple "virtual" data points for free. If you see a robot move forward, and you know it is left-right symmetric, you have effectively also seen it move forward in a mirrored configuration. By discovering these symmetries automatically, we can apply the power of equivariant neural networks to domains where the "rules" are not yet known, from complex metamaterials (Fernández et al.) to robotic harmonic analysis (Ordoñez Apraez et al.).

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_automated_symmetry_discovery_for_equivar_102b2522,
  title = {Automated Symmetry Discovery for Equivariant Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-symmetry-discovery-for-equivariant-learning/rdr_102b25229eaf40e18d51029ff3c44ab0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_automated_symmetry_discovery_for_equivar_102b2522_v1,
  title = {Automated Symmetry Discovery for Equivariant Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-symmetry-discovery-for-equivariant-learning/rdr_102b25229eaf40e18d51029ff3c44ab0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}