Reversible and irreversible bracket-based dynamics for deep graph neural networks

Authors: Anthony Gruber, Kookjin Lee, Nathaniel Trask

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This section reports results on experiments designed to probe the influence of bracket structure on trajectory prediction and nodal feature classification. Additional experimental details can be found in Appendix B. In each Table, orange indicates the best result by our models, and blue indicates the best of those compared.
Researcher Affiliation Collaboration Anthony Gruber Center for Computing Research Sandia National Laboratories Albuquerque, NM. USA adgrube@sandia.gov Kookjin Lee School of Computing and Augmented Intelligence Arizona State University Tempe, AZ. USA kookjin.lee@asu.edu Nathaniel Trask School of Engineering and Applied Science University of Pennsylvania Philadelphia, PA. USA ntrask@seas.upenn.edu
Pseudocode No The paper describes methods and provides architectural diagrams but does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks.
Open Source Code Yes Code is available at the Github repository https://github.com/natrask/Bracket Graphs.
Open Datasets Yes Table 4 and Table 5 present results on common benchmark problems including the citation networks Cora [58], Citeseer [59], and Pubmed [60], as well as the coauthor graph, Coauthor CS [61], and the Amazon co-purchasing graphs, Computer and Photo [62].
Dataset Splits Yes a Bayesian search is conducted using Weights and Biases [76] for each bracket and each dataset using a random 80/10/10 train/valid/test split with random seed 123.
Hardware Specification Yes The experiments are conducted on systems that are equipped with NVIDIA RTX A100 and V100 GPUs.
Software Dependencies No The paper mentions software like PYTHON, PYTORCH [71], CUDA, and TORCHDIFFEQ [16] but does not provide specific version numbers for any of them.
Experiment Setup Yes The networks are trained to reconstruct the node/edge features in mean absolute error (MAE) using the Adam optimizer [73]. The NODEs and metriplectic bracket use an initial learning rate of 10 4, while the other models use an initial learning rate of 10 3. The width of the hidden layers in the message passing encoder/decoder is 64, and the number of hidden features for nodes/edges is 32. The time integrator used is simple forward Euler.