Active Slices for Sliced Stein Discrepancy
Authors: Wenbo Gong, Kaibo Zhang, Yingzhen Li, Jose Miguel Hernandez-Lobato
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on goodness-of-fit tests and model learning show that our approach achieves both improved performance and faster convergence. |
| Researcher Affiliation | Academia | 1Department of Engineering, University of Cambridge, Cambridge, United Kingdom 2Department of Computing, Imperial College London, London, United Kingdom. |
| Pseudocode | Yes | Algorithm 1 Active slice algorithm |
| Open Source Code | No | The paper does not explicitly state that the source code for their methodology is made open or provide a link to it. |
| Open Datasets | Yes | The 3 GOF test benchmarks, with details in appendix H.1, are: (1) Laplace: p(x) = N(0, I), q(x) = QD d=1 Lap(xd|0, 1/ 2); (2) Multivariate-t: p(x) = N(0, 5 3I), q(x) is a fully factorized multivariate-t with 5 degrees of freedom, 0 mean and scale 1; (3) Diffusion: p(x) = N(0, I), q(x) = N(0, Σ1) where in q(x) the variance of 1st-dim is 0.3 and the rest is I. |
| Dataset Splits | Yes | For all methods requiring GO or active slices, we split the 1000 test samples from q into 800 test and 200 training data, where we run GO or active slice method on the training set. |
| Hardware Specification | No | The paper does not provide specific hardware specifications (e.g., GPU models, CPU types) used for running the experiments. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers. |
| Experiment Setup | Yes | For all methods requiring GO or active slices, we split the 1000 test samples from q into 800 test and 200 training data, where we run GO or active slice method on the training set. SKSD-g+GO with 1000 training epochs still exhibits a decreasing test power in Laplace and multivariate-t. On the other hand, SKSD-g+KE+GO with 50 training epochs has nearly optimal performance. [...] SKSD-rg+GO runs 50 training epochs with r and gr initialized to I. |