In traditional deep learning, we bake our knowledge of the world into the architecture. We use Convolutional Neural Networks (CNNs) because we know that an object’s identity doesn't change when it shifts a few pixels to the left. As Joan Bruna, Arthur Szlam, and Yann LeCun formalized in Learning Stable Group Invariant Representations with Convolutional Networks, these networks succeed because they act as "stable group invariant" operators, decomposing global transformations into a hierarchy of local symmetries.
However, this approach hits a wall when we encounter data where the symmetries are unknown or too complex to define by hand—such as the laws governing a new physical system or the subtle variations in speech. The emerging research direction focuses on automated symmetry discovery: moving from prescribing symmetries to learning them directly from the data.
The Intuition: Symmetries as Redundancy
A symmetry is essentially a transformation that leaves the "essence" of a data point unchanged. If you rotate a picture of a cat, it is still a cat. If a model can identify which transformations are "nuisance variables," it can ignore them, leading to massive gains in sample efficiency. Alberto Bietti and colleagues demonstrated in On the Sample Complexity of Learning with Geometric Stability that knowing these symmetries reduces the amount of data needed by a factor proportional to the size of the symmetry group.
The challenge is that most current "steerable" or "equivariant" networks, like those proposed by Taco S. Cohen and Max Welling in Steerable CNNs, require researchers to manually specify these groups (e.g., "rotation" or "reflection") before training. Automated discovery seeks to remove this human-in-the-loop requirement.
Learning the Generators of Change
How do you find a symmetry you haven't seen before? One powerful method is to look at how data points change in pairs. If you have two observations of the same system at different times, the transformation between them likely follows a specific mathematical structure.
Chris Ick and Vincent Lostanlen in Learning a Lie Algebra from Unlabeled Data Pairs showed that a neural network can recover the "infinitesimal generator" of a transformation group—the fundamental mathematical rule of how things change—just by minimizing the distance between a predicted transformation and the actual data. Similarly, Prathosh AP introduced $H_\gamma$-Net in Automatic Discovery of One Parameter Subgroups of SO(n), which builds these discovered symmetries directly into the architecture to provide provable invariance.
Factoring "What" from "How"
Another way to discover symmetry is to force the model to separate the content of an image from its pose. Aaron Courville and Yoshua Bengio explored this in The Statistical Inefficiency of Sparse Coding for Images. Instead of learning a thousand different edge detectors at every possible angle, a "factored" model learns one master edge (the "what") and a set of transformation parameters (the "how").
This concept has been extended to unsupervised action estimation. Adam Prügel-Bennett and Jonathon Hare in Linear Disentangled Representations and Unsupervised Action Estimation used policy networks to select from a set of learnable transformation matrices. By seeing which matrix best explains the transition between two frames, the model "discovers" the underlying algebraic structure of the environment's symmetries without being told they exist.
Discovering Physical Laws
Perhaps the most exciting application of this field is in the natural sciences. When a model is forced to be "fully covariant"—respecting the arbitrary choices we make about units or coordinate systems—it can actually uncover hidden physical constants.
Soledad Villar argues in Towards fully covariant machine learning that if a model is constrained by symmetry but the data doesn't seem to fit, the model can autonomously identify missing latent vectors, such as gravity. This turns symmetry discovery into a tool for scientific inquiry, allowing models to extract governing exponents and scaling laws from noisy experimental data, as seen in the work of Watanabe Ryota on Data-driven discovery of self-similarity using neural networks.
Summary of the Shift
The research trajectory across these 393 papers moves through three stages:
- Hard-coded Invariance: Using pooling and convolutions to handle known translations (Bruna et al., 2013).
- Prescribed Equivariance: Designing layers for specific groups like rotations or reflections (Cohen & Welling, 2017).
- Automated Discovery: Learning the transformation groups and Lie algebras directly from raw, unlabeled data (Ick & Lostanlen, 2020; AP, 2025).
By automating this discovery, we enable machines to find the "hidden rules" of any domain, from the phonemes of a language (Zhang et al., 2014) to the fundamental constants of physics.