KSD Aggregated Goodness-of-fit Test

Authors: Antonin Schrab, Benjamin Guedj, Arthur Gretton

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

Reproducibility Variable Result LLM Response
Research Type Experimental We find on both synthetic and real-world data that KSDAGG outperforms other state-of-the-art quadratic-time adaptive KSD-based goodness-of-fit testing procedures. We discuss the implementation of KSDAGG and experimentally validate our proposed approach on benchmark problems, not only on datasets classically used in the literature but also on original data obtained using state-of-the-art generative models (i.e. Normalizing Flows).
Researcher Affiliation Academia Antonin Schrab Centre for Artificial Intelligence Gatsby Computational Neuroscience Unit University College London & Inria London a.schrab@ucl.ac.uk Benjamin Guedj Centre for Artificial Intelligence University College London & Inria London b.guedj@ucl.ac.uk Arthur Gretton Gatsby Computational Neuroscience Unit University College London arthur.gretton@gmail.com
Pseudocode Yes Algorithm 1 KSDAGG
Open Source Code Yes Contributing to the real-world applications of these goodness-of-fit tests, we provide publicly available code to allow practitioners to employ our method: https://github.com/antoninschrab/ksdagg-paper.
Open Datasets Yes MNIST dataset (Le Cun et al., 1998, 2010)
Dataset Splits No The paper uses various datasets (Gamma, GBRBM, MNIST Normalizing Flow) but does not explicitly provide details about train/validation/test splits for its experiments. For MNIST, it mentions a pre-trained model but not the experimental splits for the KSDAGG tests.
Hardware Specification Yes All experiments have been run on an AMD Ryzen Threadripper 3960X 24 Cores 128Gb RAM CPU at 3.8GHz
Software Dependencies No The paper mentions using third-party implementations ('Jitkrittum et al. (2017)' and 'Phillip Lippe’s implementation') but does not specify any software dependencies with version numbers for its own code or key libraries.
Experiment Setup Yes All our experiments are run with level = 0.05 using the IMQ kernel defined in Equation (7) with parameter βk = 0.5. We use a parametric bootstrap with B1 = B2 = 500 bootstrapped KSD values to compute the adjusted test thresholds, and B3 = 50 steps of bisection method to estimate the correction u in Equation (6).