Equivariance (Machine Learning)

Equivariance is a relationship between a transformation applied to data and the way a model processes that data. When a model is equivariant, applying a specific change—such as shifting an image left or rotating a molecule—before feeding it into the model results in an output that has been changed in exactly the same prescribed manner after the model runs. In other words, the order of transformation and modelling can be swapped without altering the overall effect; the model’s response mirrors the structure of the input change.

This property matters because many real‑world problems have built‑in symmetries. If a task is known to behave consistently under certain motions—like an object looking the same whether it appears higher or lower in a photograph—a model that respects those motions does not need to relearn the same pattern for every possible position. Equivariance therefore lets a system learn more efficiently, improves its ability to generalize beyond the examples it has seen, and often yields predictions that obey physical or geometric laws.

Equivariance shows up wherever symmetry is central to the data. Convolutional neural networks exploit translation equivariance, ensuring that sliding a picture across the pixel grid shifts the feature maps in the same way. Graph neural networks are built to be permutation‑equivariant so reshuffling node order does not change their output apart from the corresponding reshuffle. In scientific machine learning, models of fluids or crystals are designed to be equivariant under rotations and reflections, guaranteeing that predictions rotate or reflect exactly as the underlying system would.