Equivariance
A function property where input transformations produce corresponding, predictable transformations in the output.
Equivariance describes a property of a function or model whereby applying a transformation to the input produces a corresponding, well-defined transformation in the output. Formally, a function f is equivariant with respect to a transformation g if f(g·x) = g·f(x) — the function and the transformation commute. This stands in contrast to invariance, where the output remains entirely unchanged under input transformations. In machine learning, equivariance is most naturally illustrated in image processing: a translation-equivariant model that detects an edge in one region of an image will detect the same edge if it shifts to another region, with the detection shifting correspondingly rather than disappearing or requiring relearning.
The practical importance of equivariance in neural networks lies in its ability to encode known symmetries directly into model architecture, rather than learning them implicitly from data. Standard convolutional neural networks (CNNs) are translation-equivariant by construction — a core reason for their success in vision tasks. Extending this principle to other symmetry groups (rotations, reflections, permutations, or more abstract group actions) gave rise to the field of geometric deep learning. Group equivariant convolutional networks (G-CNNs), introduced around 2016, generalized the convolution operation to act over broader symmetry groups, dramatically improving sample efficiency and generalization on tasks with known geometric structure.
Equivariance has since become a foundational design principle across many domains. In molecular property prediction and drug discovery, equivariant graph neural networks respect the rotational and translational symmetries of 3D atomic structures, enabling physically consistent predictions without requiring exhaustive data augmentation. In climate modeling, robotics, and particle physics, similar principles allow models to respect the underlying symmetries of the problem domain. By baking symmetry constraints into architecture rather than learning them from scratch, equivariant models typically require less data, generalize more reliably, and produce outputs that are consistent with known physical or geometric laws.
