Latent Lie Group Representations
Keywords: deep learning, symmetry, lie groups
Abstract: Symmetry detection tasks rely on identifying transformations of data points that keep some task-related quality, such as classification label, identical. These symmetries are useful during model selection for neural networks, as even a conceptually simple symmetry (e.g., translation invariance) can lead to superior performance-efficiency tradeoffs (e.g., CNNs). Leveraging neural networks to learn these transformations can lead to approaches that yield representations of the transformations in latent space, rather than just the data itself.
In this work, we propose a latent variable framework for learning one-parameter subgroups of Lie group symmetries from observations, improving the accuracy of the learned transformation with respect to the one in pixel-space, even including situations in which this might not even be desirable.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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/2024/AuthorGuide.
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: 9095
Loading