Equivariant Neural Networks

Equivariant neural networks build in known symmetries, so rotating or relabeling an input transforms the prediction in a predictable way rather than requiring the model to relearn the same pattern. Common approaches include group convolutions and steerable kernels, irreducible-representation tensor products, geometric-algebra layers, and equivariant transformers, diffusion models, and neural operators. They are used in settings where geometry or physical laws matter, including molecular modeling, fluid simulation, medical imaging, point clouds, and robotics, and are often evaluated for accuracy, data efficiency, robustness to transformations, and generalization across resolutions or domains. A central tension is choosing symmetries that match the data: exact constraints can improve reliability and efficiency, but may be too restrictive when symmetries are broken, while efficient implementations and reliable training remain important challenges.