When we train a standard convolutional neural network (CNN), we take advantage of a built-in geometric guarantee: translation equivariance. If an edge or a pattern shifts across an image, the network's internal representations shift accordingly.
In recent years, geometric deep learning has successfully expanded these guarantees to rigid transformations like 2D and 3D rotations ($SO(2)$ and $SO(3)$). By hardcoding these symmetries into network layers using group convolutions (Cohen et al.) or steerable continuous filters (Kundu et al.), models achieve remarkable sample efficiency. They no longer need to see an object at every possible angle to recognize it.
However, traditional equivariant architectures suffer from a fundamental limitation: they are often trapped in a mathematical straightjacket. Most frameworks are strictly engineered for rigid Euclidean transformations in two or three dimensions. When confronted with real-world data variations—such as scaling, reflections, affine distortions, or higher-dimensional coordinate systems—these rigid constraints break down.
To solve this, a major research direction focuses on generalizing equivariant networks to broader, non-rigid, and high-dimensional symmetry groups. Instead of forcing networks to obey rigid geometric rules, researchers are developing flexible mechanisms that scale equivariance across complex transformations.
The Bottleneck of Rigid Equivariance
To understand why generalizing symmetry groups is difficult, consider how traditional rotation-equivariant networks operate. They typically rely on one of two methods:
- Discrete Group Convolutions ($G$-CNNs): Rotating filters by fixed angles (e.g., 90 degrees) on a square grid.
- Steerable Networks: Using complex mathematical expansions, such as spherical harmonics or Wigner matrices (Zhemchuzhnikov et al.), to handle continuous 3D rotations.
While mathematically rigorous, these approaches do not scale easily. Spherical harmonics are explicitly tied to three-dimensional space ($SO(3)$), making them useless for higher-dimensional physics simulations or abstract graphs. Furthermore, trying to hardcode continuous scaling (dilation) or affine transformations into a convolutional kernel introduces severe interpolation artifacts and immense computational overhead.
To escape these bottlenecks, the field has shifted toward three core paradigms: learned canonicalization, relaxed/partial equivariance, and coordinate-free representations.
Paradigm 1: Learned Canonicalization
Instead of redesigning the internal layers of a neural network to be mathematically equivariant to a complex group, why not transform the data itself into a standard orientation? This is the core intuition behind learned canonicalization.
Kaba et al. proposed a framework where a lightweight, separate network acts as a "canonicalizer." When an input arrives—regardless of its rotation, scale, or distortion—this network predicts the transformation required to map the data into a standardized, canonical frame. Once the data is normalized, it is passed to a standard, non-equivariant backbone network (like a standard ViT or MLP).
Input Data (Arbitrary Scale/Rotation)
│
▼
┌──────────────────────────────┐
│ Canonicalization Network │ ──► Predicts & applies inverse transform
└──────────────────────────────┘
│
▼
Canonicalized Data (Standardized Frame)
│
▼
┌──────────────────────────────┐
│ Standard Backbone (CNN/MLP) │
└──────────────────────────────┘
This approach completely decouples the symmetry group from the network architecture. If you want to support a new symmetry group (such as scaling or affine transformations), you do not need to rewrite the backbone's convolutional layers; you only need to allow the canonicalization network to output that specific class of transformations.
Paradigm 2: Relaxed and Partial Equivariance
In the real world, perfect mathematical symmetry is rare. A neural network designed to be perfectly rotation-equivariant might struggle with objects that exhibit "symmetry breaking"—where orientation actually matters. For example, in character recognition, a fully rotation-invariant model cannot tell the difference between a '6' and a '9'. Similarly, in object detection, aerial images might be arbitrarily rotated, but objects like cars still have a distinct front and back relative to their trajectory.
To bridge this gap, researchers have introduced relaxed and partial equivariance, allowing models to dynamically adjust their sensitivity to transformations.
Learnable Perturbations
Wu et al. introduced Relaxed Rotation-Equivariance (RRE) for object detection. Instead of enforcing strict group constraints, their approach introduces learnable perturbations into the rotation group. This allows the network to adapt to real-world imperfections and asymmetric variations in data, leading to faster convergence and fewer parameters than rigidly constrained models.
Input-Aware Symmetries
Kim et al. took this further by treating equivariance as a data-dependent latent variable through Variational Partial Group Convolutions. The network dynamically decides how much symmetry to enforce based on the specific input. If the model processes an upright digit '7', it can choose to remain sensitive to its orientation, but if it processes a symmetric texture, it can switch to full rotation invariance.
Soft Regularization
Another way to avoid rigid architectural constraints is to enforce symmetry through the loss function rather than the network's wiring. Bai et al. developed EQ-Reg, a regularization strategy that forces consistency across transformations (such as spatial rotations combined with cyclic channel shifts) during training. Because this is a soft constraint guided by regularization, it allows standard network architectures to benefit from equivariance while remaining flexible enough to handle non-rigid real-world data.
Paradigm 3: Coordinate-Free and Dimension-Agnostic Scaling
When dealing with graphs, point clouds, or high-dimensional scientific data, models cannot rely on fixed pixel grids. To generalize equivariance to these domains, architectures must become independent of specific coordinate systems and dimensions.
Dimension-Agnostic Networks
Instead of building models restricted to 2D or 3D spaces, Garcia Satorras introduced E(n) Equivariant Graph Neural Networks (EGNNs) (Garcia Satorras). EGNNs achieve equivariance across arbitrary dimensions ($n$) by operating entirely on relative distances between nodes rather than absolute coordinates. Because the distance between two points is invariant to rotation, translation, and reflection, the network can update coordinates and features without relying on complex, dimension-specific mathematical tools like spherical harmonics.
Complete Local Frames
For tasks requiring highly expressive geometric features without the heavy computational cost of tensor products, models like ClofNet (Du et al.) construct unique, local coordinate frames for every edge in a graph. By projecting raw geometric tensors into these local frames, the tensors are losslessly converted into invariant scalars. This allows standard, highly optimized message-passing networks to handle complex spatial interactions while maintaining strict geometric guarantees.
Summary
The evolution of equivariant neural networks is a journey from rigid geometric constraints to flexible, generalized transformations. By moving away from hardcoded, dimension-restricted layers and embracing learned canonicalization, partial equivariance, and coordinate-free graph representations, modern architectures can scale to handle the messy, multi-dimensional symmetries of the real world.