Researchers have developed a new approach to characterizing equivariant linear layers for representations of permutations and related groups. Unlike traditional methods, this approach uses irreducible representations and Schur's lemma to derive alternative models like DeepSets and Deep Weight Space (DWS) networks. The team also extended their method to unaligned symmetric sets, where they found a vast number of additional non-Siamese layers that can improve performance in tasks such as graph anomaly detection and learning Wasserstein distances. --- Why it matters: This research matters because it provides new insights into the design of equivariant neural networks, which are essential for applications like graph analysis and processing. The ability to derive alternative models using irreducible representations could lead to more efficient and effective neural network architectures. Source: https://arxiv.org/abs/2410.06665