Latent Space Symmetry Discovery
Authors: Jianke Yang, Nima Dehmamy, Robin Walters, Rose Yu
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentally, we demonstrate that our method can accurately discover the intrinsic symmetry in high-dimensional dynamical systems. La Li GAN also results in a wellstructured latent space that is useful for downstream tasks including equation discovery and long-term forecasting. |
| Researcher Affiliation | Collaboration | 1UCSD 2IBM Research 3Northeastern University |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | Yes | We make our code available at https://github.com/jiankeyang/La Li GAN. |
| Open Datasets | Yes | We use the script from SINDy Autoencoder 1 to generate the dataset. 1https://github.com/kpchamp/Sindy Autoencoders/tree/master/rd solver. ... We consider the Top Tagging dataset (Kasieczka et al., 2019)... |
| Dataset Splits | No | The paper specifies training and testing sets, but does not explicitly mention a separate validation set or a three-way split in the main text. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments, only general architecture types like MLPs and 1D convolution architectures. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | For the 2D latent space symmetry discovery, we train for 150 epochs with batch size 64. The learning rates for the autoencoder, the generator and the discriminator are 0.0003, 0.001, 0.001, respectively. The weights of the reconstruction loss and the GAN loss are set to wrecon = 1 and w GAN = 0.01. As in Lie GAN, we also include a regularization loss term lreg for Lie GAN generator, which pushes the Lie algebra basis away from zero, and the weight for the regularization is set to wreg = 0.1. We also apply sequential thresholding to the Lie GAN generator parameters. Every 5 epochs, matrix entries with absolute values less than 0.01 times the max absolute values across all entries are set to 0. |