New Framework for Equivariant Neural Networks via Irreducible Representations

A recent study published on arXiv (2410.06665v4) introduces a novel approach for defining equivariant linear layers in neural networks through the use of irreducible representations and Schur's lemma, moving away from conventional parameter-sharing methods. The researchers successfully derive established models including DeepSets, 2-IGN graph equivariant networks, and notably simplify the derivation for Deep Weight Space (DWS) networks. Furthermore, they broaden their methodology to encompass unaligned symmetric sets that necessitate wreath product equivariance, offering a comprehensive characterization that uncovers many additional non-Siamese layers beyond earlier limitations. This paper falls under the 'replace-cross' category and can be accessed via the provided URL.