PowerNet: Truncated Matrix Power Series as Quasi-Equivariant Layers
Keywords: symmetry, deep learning, geometry, Lie theory
Abstract: Despite being theoretically well-grounded, enforcing strict equivariance in deep learning models has shown to be harmful in some cases. The problem is that most available data does not follow mathematically precise rules, is noisy, and is not strictly group-structured. While soft equivariance approaches attempt to address these issues, they often struggle to maintain group structure and lack strong theoretical guarantees, potentially compromising the benefits of equivariance. Here we introduce the concept of \textit{quasi equivariances}, where group structure is maintained but the associated parameters become distributions, and implement it in the proposed \textit{PowerNet} architecture. Similar to CNNs, PowerNet is constructed by interlacing truncated matrix power series with non-linearities. We show how the base matrix used to define the power series can instill quasi-equivariance in a natural way. Finally, we provide results for augmented MNIST classification and transformation magnitude regression in addition to classification of CIFAR-10.
Primary Area: learning theory
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4405
Loading