Gradient-Free Kernel Stein Discrepancy

Authors: Matthew Fisher, Chris J. Oates

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

Reproducibility Variable Result LLM Response
Research Type Experimental The practical performance of GF-KSD remains to be assessed. Suitable choices for both q and the kernel parameters σ and β are proposed and investigated in Section 3, and practical demonstrations of GF-KSD are presented in Section 4.
Researcher Affiliation Academia 1Newcastle University, UK 2Alan Turing Institute, UK
Pseudocode No The paper describes methods using prose and mathematical notation but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Python code to reproduce the experiments reported below can be downloaded at [blinded].
Open Datasets Yes The data analysed are due to Hewitt [1921], and full details are contained in Appendix C.3.
Dataset Splits No The paper discusses 'sequences (πn)n N' and 'samples (xn)n N' in its experiments, and describes optimization processes, but does not provide explicit training/validation/test dataset splits for reproducibility.
Hardware Specification No The paper does not provide specific details about the hardware used for running its experiments.
Software Dependencies Yes In order to obtain independent samples from the posterior for comparison, we utilised Stan [Stan Development Team, 2022]... URL https://mc-stan.org/. R package version 2.21.5.
Experiment Setup Yes The Laplace approximation was obtained by the use of 48 iterations of the L-BFGS optimisation algorithm... The stochastic optimisation routine used was Adam [Kingma and Ba, 2015] with learning rate 0.001. Due to issues involving exploding gradients due to the q/p term in GF-KSD, we utilised gradient clipping... with the maximum 2-norm value taken to be 30. In the banana experiment, the dimensionality of the hidden units in the underlying autoregressive neural network was taken as 20.