A Kernelised Stein Statistic for Assessing Implicit Generative Models

Authors: Wenkai Xu, Gesine D Reinert

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on synthetic distributions and trained generative models on synthetic and real datasets illustrate that the method shows improved power performance compared to existing approaches.
Researcher Affiliation Academia Wenkai Xu Department of Statistics University of Oxford wenkai.xu@stats.ox.ac.uk, Gesine Reinert Department of Statistics University of Oxford and Alan Turing Institute reinert@stats.ox.ac.uk
Pseudocode Yes Algorithm 1 Estimating the conditional probability via summary statistics, Algorithm 2 Assessment procedures for implicit generative models
Open Source Code Yes The code is available at https://github.com/wenkaixl/npksd.git.
Open Datasets Yes MNIST Dataset This dataset contains 28 28 grey-scale images of handwritten digits [Le Cun et al., 1998]; CIFAR10 Dataset This dataset contains 32 32 RGB coloured images [Krizhevsky, 2009]
Dataset Splits No The paper mentions training and test samples but does not explicitly provide details about a validation set split or how to create one for reproducibility.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments, such as GPU models, CPU models, or specific cloud instance types.
Software Dependencies No The paper does not explicitly state specific software dependencies with version numbers, such as libraries, frameworks, or operating systems used for the experiments.
Experiment Setup Yes For our experiments, 600 samples are generated from the generative models and 100 samples are used for the test; the samples are then down-sampled into 7 7 images, as in Schrab et al. [2021]. The test level is α = 0.05.; For our experiments, 800 samples are generated from the generative models and 200 samples are used for the test; the samples are then down-sampled into 8 8 images, in a similar fashion as in Schrab et al. [2021]. The test level is α = 0.05.