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].

Trust-Region Variational Inference with Gaussian Mixture Models

Authors: Oleg Arenz, Mingjun Zhong, Gerhard Neumann

JMLR 2020 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate VIPS++ on several domains and compare it to state-of-the-art methods for variational inference and Markov-chain Monte Carlo. We demonstrate that we can learn high quality approximations of several challenging multimodal target distributions that are significantly better than those learned by competing methods for variational inference. Compared to sampling methods, we show that we can achieve similar sample quality while using several orders of magnitude less function evaluations. Section 5 is explicitly titled "Experiments" and contains sub-sections detailing "Sampling Problems", "Ablations", "Illustrative Experiment", and "Results", including analysis of metrics like "maximum mean discrepancy (MMD)" and figures showing performance over "function evaluations".
Researcher Affiliation Collaboration Oleg Arenz from Technische Universität Darmstadt, Mingjun Zhong from University of Aberdeen are academic institutions. Gerhard Neumann from Karlsruhe Institute of Technology is an academic institution, but he is also affiliated with Bosch Center for Artificial Intelligence, which is an industry entity. Since there is a mix of academic and industry affiliations, the paper is a collaboration.
Pseudocode Yes The paper includes several algorithms: "Algorithm 1 Updating a Gaussian variational approximation based on surrogate", "Algorithm 2 Variational Inference by Policy Search (Basic Variant)", and "Algorithm 3 Ensure that every component has sufficiently many effective samples." These are clearly structured and labeled algorithm blocks.
Open Source Code Yes The implementation can be found at https://github.com/Oleg Arenz/VIPS.
Open Datasets Yes We perform two experiments for binary classification that have been taken from Nishihara et al. (2014) using the German credit and breast cancer data sets (Lichman, 2013). ... We perform Bayesian Gaussian process regression on the ionosphere data set (Lichman, 2013) as described by Nishihara et al. (2014).
Dataset Splits Yes We consider 81 noisy observations o1...81 of x1 and x2 using steps of dt = 1. We assume Gaussian observation noise with zero mean and variance σ2 = 0.2 and discard the first 40 observations.
Hardware Specification No The paper states: "Calculations for this research were conducted on the Lichtenberg high performance computer of the TU Darmstadt." and "we ran the experiments on a fast quad-core laptop and made use of multi-threading." While a named computer and a type of laptop are mentioned, specific hardware details like CPU/GPU models, processor types, or memory amounts are not provided for either.
Software Dependencies No The paper mentions several software implementations by reference (e.g., "Python implementation by Bovy (2013)", "pyhmc (Nabney et al., 2018)", "implementation based on tensorflow (Abadi et al., 2015)"). However, it does not provide specific version numbers for general software components like Python or TensorFlow itself, only citations to papers describing the tools or implementations.
Experiment Setup Yes Appendix L, titled "VIPS++ Hyper-Parameters", contains a table listing specific values for various parameters such as "KL bound for components 1e-2", "number of desired samples (per dimension and component) 20", "adding rate for components 30 or 1", "minimum weight 1e-6", and "ℓ2-regularization for WLS 1e-14".