Standard Convolutional Neural Networks (CNNs) possess a natural "superpower": translation equivariance. If an object shifts in an image, its internal representation shifts accordingly. However, as noted by Yuexi Du and colleagues in GMR-Conv, this superpower fails for rotations and reflections. While we can define perfect mathematical symmetries in continuous space, implementing them on a grid of square pixels introduces "jagged edges" and interpolation artifacts that break the very equivariance we seek to build.
A growing body of research focuses on bridging this gap between continuous group theory and discrete computation, ensuring that neural networks remain robust to transformations like rotation, scaling, and gauge changes.
The Discretization Dilemma
The fundamental challenge is that a square pixel grid does not look the same when rotated by an arbitrary angle. If you rotate a 3x3 kernel by 45 degrees, the new "pixels" fall between the original grid coordinates, requiring interpolation. As Luca Della Libera and others explain in Deep Learning for 2D and 3D Rotatable Data, these small interpolation errors are not harmless; they accumulate and amplify as data passes through deep layers, leading to inconsistent feature representations.
This problem is even more pronounced in 3D. Iván Díaz and Richard McKinley demonstrate in SO(3)-steerable convolutions for 3D medical data that voxel grids introduce discretization artifacts that cause performance to fluctuate based on a patient's orientation in a scanner.
Steerable Filters and Continuous Kernels
To solve this, researchers move away from learning discrete pixel weights. Instead, they parameterize filters using continuous basis functions.
In Learning Steerable Filters for Rotation Equivariant CNNs, Maurice Weiler and his team represent filters as linear combinations of circular harmonics. Because these harmonics are defined analytically, the filter can be "steered" to any orientation without the blurring associated with grid rotation.
A similar logic applies to PDO-eConvs. Zhengyang Shen and Zhouchen Lin model convolutions as Partial Differential Operators (PDOs). They provide a crucial theoretical bound: the equivariance error introduced by discretizing these operators scales quadratically with grid spacing. This means that by using higher-order numerical schemes, we can achieve robust equivariance even on standard small kernels.
Symmetries Beyond the Plane
When moving to non-Euclidean spaces, such as the surface of a sphere or a curved manifold, the discretization problem becomes even more complex.
- Spherical Grids: Taco Cohen and Max Welling introduced Spherical CNNs, which use generalized Fourier transforms to handle 3D rotations. However, as Carlos Esteves notes in Learning SO(3) Equivariant Representations, standard equiangular spherical grids have non-uniform sampling density (clustering at the poles), which breaks perfect symmetry.
- Manifolds and Meshes: On curved surfaces, there is no global "up" or "down." Pim de Haan and colleagues address this in Gauge Equivariant Mesh CNNs by using parallel transport to move data across triangle meshes. Here, the discretization artifact stems from the mesh geometry itself, which can break the exact equivariance defined in the continuous "gauge" theory.
Achieving Numerical Stability
Recent work focuses on making these models "representation equivalent." In Are Neural Operators Really Neural Operators?, Francesca Bartolucci and her team use Frame Theory to ensure that discrete implementations remain mathematically consistent with their continuous counterparts across different grid resolutions, effectively eliminating aliasing.
Similarly, Adaptive Sampling for Continuous Group Equivariant Neural Networks proposes predicting the sampling grid from the data itself. This ensures that even when using a small number of samples to save memory, the network's output remains perfectly equivariant to continuous rotations.
By treating the grid not as the ground truth, but as a discrete sample of a continuous underlying signal, these architectures—from SFNOs for weather modeling to EST for molecular dynamics—allow deep learning to respect the fundamental physical symmetries of the real world.