Automated Discovery of Data Symmetries

618 concepts393 papers1 proposal

The remarkable success of modern deep learning architectures often stems from hard-coded geometric priors. For instance, Joan Bruna et al. demonstrated in Learning Stable Group Invariant Representations with Convolutional Networks that Convolutional Neural Networks (CNNs) succeed because they cascade layers to decompose complex global transformations into a hierarchy of local symmetries.

However, traditional geometric deep learning requires human experts to manually specify these transformation groups before training, as seen in the rigid rotation or translation invariants formulated by Anant Raj et al. in Local Group Invariant Representations via Orbit Embeddings. or the continuous Lie group architectures designed by Erik J Bekkers in B-Spline CNNs on Lie groups. This manual approach fails when data variations are chaotic, unconstrained, or entirely unknown, as noted by Dipan K. Pal et al. in Discriminative Invariant Kernel Features....

To build truly autonomous systems, machine learning models must transition from hard-coded symmetry design to automated symmetry discovery—learning the underlying geometric and algebraic structures directly from raw data.

The Geometry of Symmetries: Learning Lie Algebras

How can a neural network search for an unknown continuous symmetry? A continuous symmetry group (a Lie group), such as the set of all 3D rotations, forms a curved manifold. Optimizing over a curved manifold directly via gradient descent is notoriously difficult.

The key insight that makes automated symmetry discovery work is shifting the optimization from the full Lie group to its Lie algebra—the flat, linear tangent space right at the identity transformation. Because this space is linear, any transformation can be represented as a linear combination of a few basis matrices, known as infinitesimal generators.

   Lie Group (Curved Manifold)
        /‾‾‾‾‾‾‾‾‾\
       |   • g     |  <-- Complex continuous transformation
        \_________/
          ^
          |  Matrix Exponential: exp(εW)
          |
   Lie Algebra (Flat Tangent Space)
        |    ↑
        |---• W       <-- Learnable Infinitesimal Generator (Linear Matrix)
       /              
      I (Identity)

Nima Dehmamy et al. operationalized this in Automatic Symmetry Discovery with Lie Algebra Convolutional Network. Instead of fixing the convolutions to a known group, their architecture (L-conv) parameterizes the filters using these learnable infinitesimal generators. By updating these generators via standard backpropagation, the network smoothly alters the directions of symmetry it respects.

To map these infinitesimal steps back onto the actual transformation manifold, the network uses the matrix exponential. As explored by Alex Gabel et al. in Learning Lie Group Symmetry Transformations with Neural Networks, exponentiating a learned algebraic matrix allows the model to reconstruct full, continuous transformations and quantify exactly how much a given data point has been transformed, all without prior human specification.

Discovering Symmetries from Data Dynamics

Symmetries leave distinct statistical footprints in how data changes. By observing these patterns over time or across distributions, models can infer underlying group structures without explicit labels.

Temporal Slowness and Topographic Organization

In an unsupervised setting, we can exploit the fact that the physical world transforms smoothly over time. T. Anderson Keller et al. demonstrated this in Topographic VAEs learn Equivariant Capsules. By organizing the latent space of a Variational Autoencoder (VAE) topographically—where neighboring neurons represent related features—and applying a "slowness" penalty across sequential video frames, the model naturally groups features into equivariant capsules. The network discovers real-world symmetries like rotation and scaling simply because those are the transformations that explain the slow, continuous variations in the data stream.

Action Estimation and Policy Gradients

When data consists of discrete sequential transitions, a model can discover symmetries by trying to "explain" the actions that bridge two states. Adam Prugel-Bennett et al. leveraged this in Linear Disentangled Representations and Unsupervised Action Estimation. They utilized a policy network to select from a set of learnable transformation matrices to predict a subsequent latent state. By optimizing this selection based on reconstruction accuracy, the model autonomously discovers the underlying algebraic structure of the environment's symmetries without human-defined group constraints.

Parameter-Sharing as a Learnable Constraint

From an architectural standpoint, equivariance is fundamentally a constraint on a layer's weight matrix. In Equivariance Through Parameter-Sharing, Siamak Ravanbakhsh et al. proved that a layer is equivariant to a discrete symmetry group if and only if its parameter-sharing pattern matches the structure of a colored bipartite graph defined by that group.

Instead of hard-coding this graph, automated symmetry discovery treats the parameter-sharing pattern itself as an optimization problem. For instance, the authors of Learning Invariance with Compact Transforms utilized the low displacement-rank (LDR) framework to learn displacement operators directly from data. By optimizing these operators rather than fixing them to standard structures (like the circulant matrices used in standard CNNs), the network can discover domain-specific structural constraints—such as the 2D grid structure of an image—entirely from training data.

Reverse-Engineering the Black Box

If a standard neural network is trained on a symmetric dataset without explicit architectural constraints, does it learn the symmetry on its own?

To answer this, Artem Moskalev et al. developed LieGG: Studying Learned Lie Group Generators. Rather than testing a model against pre-defined transformations, LieGG "interrogates" a trained model to reverse-engineer its internal invariances. By extracting the infinitesimal generators directly from the learned weights, they proved that standard architectures can autonomously discover geometric symmetries during training.

  [ Trained Model Weights ] ──> Extract Tangent Space ──> [ Learned Lie Generators ]
                                                                 │
                                                                 ▼
                                                      Quantify Internal Symmetries
                                                      (e.g., True Rotation vs. Learned)

Symmetries can also be discovered by analyzing the predictive ambiguity of a model. In Explaining the Ambiguity of Object Detection and 6D Pose from Visual Data, Fabian Manhardt et al. designed a system that predicts multiple pose hypotheses for an object. When an object possesses a geometric symmetry (like a symmetrical cylinder), the model's hypotheses naturally spread across all valid orientations. By analyzing the statistical distribution of these multiple predictions during inference, the system automatically maps out the object's underlying symmetries without requiring manual geometric annotations.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 428 papers
  2. Version 1 · 2026-10-04 · 489 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_automated_symmetry_discovery_ad64aa29,
  title = {automated symmetry discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-symmetry-discovery/rdr_ad64aa29e72e42b8925a186ad0a660ee},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_automated_symmetry_discovery_ad64aa29_v2,
  title = {automated symmetry discovery},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/automated-symmetry-discovery/rdr_ad64aa29e72e42b8925a186ad0a660ee/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake