Explicit Discovery of Nonlinear Symmetries from Dynamic Data

Authors: Lexiang Hu, Yikang Li, Zhouchen Lin

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On top quark tagging and a series of dynamic systems, Lie NLSD shows qualitative advantages over existing methods and improves the long rollout accuracy of neural PDE solvers by over 20% while applying to guide data augmentation. In this section, we evaluate Lie NLSD on top quark tagging (Section 5.2) and a series of dynamic systems (Section 5.3).
Researcher Affiliation Academia 1State Key Lab of General AI, School of Intelligence Science and Technology, Peking University 2Institute for Artificial Intelligence, Peking University 3Pazhou Laboratory (Huangpu), Guangzhou, Guangdong, China. Correspondence to: Zhouchen Lin <EMAIL>.
Pseudocode Yes Algorithm 1 Lie NLSD
Open Source Code Yes Code and data are available at https://github.com/hulx2002/Lie NLSD.
Open Datasets Yes We first evaluate the ability of Lie NLSD to discover linear symmetries on top quark tagging (Kasieczka et al., 2019).
Dataset Splits Yes We sample M = 100 points from the training set to construct C R100 16 in Equation (13) (in practice, we construct C C R16 16).
Hardware Specification No No specific hardware details (GPU models, CPU types, or cloud computing platforms with specifications) are provided in the paper.
Software Dependencies No No specific version numbers for key software components or libraries are provided. The paper mentions the 'Adan optimizer' but without a version or broader software stack details.
Experiment Setup Yes We use an MLP with 3 hidden layers and hidden dimension 200 to fit the mapping. The sample size for symmetry discovery is M = 100. For Lie GAN, we set the dimension of the Lie algebra basis to 7, using an MLP with 2 hidden layers and hidden dimension 512 as the discriminator... For training, we set the batch size to 256 and use the Adan optimizer (Xie et al., 2024) with a learning rate of 10-3. For symmetry discovery, we consider the singular values smaller than ϵ2 = 10-2 as the effective information of the symmetry group... For basis sparsification, we set ϵ1 = 10-2 and ϵ2 = 10-1...