Trivialized Momentum Facilitates Diffusion Generative Modeling on Lie Groups

Authors: Yuchen Zhu, Tianrong Chen, Lingkai Kong, Evangelos Theodorou, Molei Tao

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

Reproducibility Variable Result LLM Response
Research Type Experimental 3 EXPERIMENTAL RESULTS. We will demonstrate accurate generative modeling of Lie group data corresponding to 1) complicated and/or high-dim distribution on torus, 2) protein and RNA structures, 3) sophisticated synthetic datasets on possibly high-dim Special Orthogonal Group, and 4) an ensemble of quantum systems... The resulting method achieves state-of-the-art performance on protein and RNA torsion angle generation and sophisticated torus datasets. We also, arguably for the first time, tackle the generation of data on high-dimensional Special Orthogonal and Unitary groups, the latter essential for quantum problems. ... We outperform baselines by a large margin on protein/RNA torsion angle datasets.
Researcher Affiliation Academia Yuchen Zhu , Tianrong Chen , Lingkai Kong, Evangelos A. Theodorou, Molei Tao Georgia Institute of Technology EMAIL
Pseudocode Yes Algorithm 1 TDM (Trivialized Diffusion Model)... Algorithm 2 Forward Operator Splitting Integration (FOSI)... Algorithm 3 Backward Operator Splitting Integration (BSOI)... Algorithm 4 Probability Flow ODE
Open Source Code Yes Code is available at https://github.com/yuchen-zhu-zyc/TDM.
Open Datasets Yes Protein and RNA Torsion Angles: We access the dataset prepared by Huang et al. (Huang et al., 2022) from the repository of (Chen & Lipman, 2024).
Dataset Splits Yes All datasets were meticulously partitioned into training and testing sets using a 9:1 ratio.
Hardware Specification Yes Hardware: All the experiments are running on one RTX TITAN, one RTX 3090 and one 4090.
Software Dependencies No The paper mentions using 'Adam W optimizer' but does not specify software dependencies like programming languages or libraries with version numbers.
Experiment Setup Yes Throughout our experiments, we maintained the diffusion coefficient γ(t) constant at 1, while the total time horizon T varied depending on the task, with a good choice ranging from T = 5 to T = 15. We use Adam W optimizer to train the neural networks with an initial learning rate of 5 10 4 with a cosine annealing learning rate scheduler. We train for at most 200k iterations with a batch size of 1024 for each task, and we observe that the model usually converges within 100k iterations. For low dimensional experiments such as Torus, SO(3), we set D = 256. For other experiments, we set D = 512. We choose varied k based on problem difficulties, ranging from k = 3 to k = 5. We use Si LU as the activation function for all the MLPs used in the neural network.