Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Understanding MCMC Dynamics as Flows on the Wasserstein Space
Authors: Chang Liu, Jingwei Zhuo, Jun Zhu
ICML 2019 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments Detailed experimental settings are provided in Appendix C, and codes are available at https://github.com/ chang-ml-thu/FGH-flow. 5.1 Synthetic Experiment We show in Fig. 3 the equivalence of various dynamics simulations, and the advantages of p SGHMC-det and p SGHMC- f GH. [...] 5.2 Latent Dirichlet Allocation (LDA) We study the advantages of our p SGHMC methods in the real-world task of posterior inference for LDA. [...] 5.3 Bayesian Neural Networks (BNNs) We investigate our methods in the supervised task of training BNNs. Results in Fig. 5 is consistent with our claim: p SGHMC methods converge faster than Blob due to the usage of SGHMC dynamics. Their slightly better particle-efficiency can also be observed. |
| Researcher Affiliation | Academia | 1Dept. of Comp. Sci. & Tech., Institute for AI, BNRist Center, Tsinghua-Fuzhou Inst. for Data Tech., THBI Lab, Tsinghua University, Beijing, 100084, China. Correspondence to: Jun Zhu <EMAIL>. |
| Pseudocode | No | No structured pseudocode or algorithm blocks are present in the paper. |
| Open Source Code | Yes | codes are available at https://github.com/ chang-ml-thu/FGH-flow. |
| Open Datasets | Yes | 5.2 Latent Dirichlet Allocation (LDA) We study the advantages of our p SGHMC methods in the real-world task of posterior inference for LDA. We follow the same settings as Liu et al. (2018) and Chen et al. (2014). [...] 5.3 Bayesian Neural Networks (BNNs) We investigate our methods in the supervised task of training BNNs. We follow the settings of Chen et al. (2014) with slight modification explained in Appendix. Results in Fig. 5 is consistent with our claim: p SGHMC methods converge faster than Blob due to the usage of SGHMC dynamics. Their slightly better particle-efficiency can also be observed. [...] Performance on BNN with MNIST data set. |
| Dataset Splits | No | The paper mentions 'training BNNs' and 'holdout perplexity' but does not provide specific percentages or counts for training/validation/test splits, nor does it specify predefined splits used from citations. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory) are provided for the experimental setup. |
| Software Dependencies | No | No specific ancillary software details, such as library names with version numbers, are provided. |
| Experiment Setup | Yes | All methods adopt the same step size 0.01, and SGHMC-related methods share the same Σ 1 = 1.0, C = 0.5. [...] All methods share the same step size 0.001 and parameters Σ 1 = 300 and C = 0.1. [...] SGHMC-related methods share parameters. |