Neural Relational Inference with Efficient Message Passing Mechanisms
Authors: Siyuan Chen, Jiahai Wang, Guoqing Li7055-7063
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The experimental results on simulated physics systems show that the proposed method outperforms existing state-of-the-art methods. |
| Researcher Affiliation | Academia | Siyuan Chen, Jiahai Wang , Guoqing Li School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China {chensy47, ligq7}@mail2.sysu.edu.cn, wangjiah@mail.sysu.edu.cn |
| Pseudocode | Yes | Notations used in this paper, details of the computation of L and the pseudo code of NRI-MPM are shown in the supplementary material1. |
| Open Source Code | No | The codes of NRI2 and Modular Meta3 are public and thus directly used in our experiments. SUAGR is coded by ourselves according to the original paper. (Note: The paper states that code for *baselines* is public, but not for *their own method*.) |
| Open Datasets | Yes | All methods are tested on three types of simulated physical systems: particles connected by springs, charged particles and phase-coupled oscillators, named Springs, Charged and Kuramoto, respectively... Further details on data generation can be found in (Kipf et al. 2018). |
| Dataset Splits | Yes | All datasets contain 50k training samples, and 10k validation and test samples. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory amounts) used for running experiments were mentioned. |
| Software Dependencies | No | The paper mentions using GRUs and attention mechanisms as components, but does not provide specific software versions for libraries (e.g., PyTorch 1.9) or frameworks (e.g., TensorFlow 2.x) used for implementation. Mind Spore is mentioned in acknowledgements, but without a version number. |
| Experiment Setup | Yes | The penalty factor λ is set to 102 for all datasets except that it is set to 1 and 103 for the 5-object and 10-object Kuramoto datasets, respectively. Details of experimental settings are shown in the supplementary material. |