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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Variational Inference with Mixtures of Isotropic Gaussians
Authors: Marguerite Petit-Talamon, Marc Lambert, Anna Korba
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our contributions include the development of a variational framework and algorithms tailored to isotropic Gaussian mixtures, as well as an empirical evaluation across synthetic and real-world datasets, demonstrating that our approach achieves a compelling balance between modeling accuracy for multimodal targets and computational efficiency. This paper is organized as follows. Section 6. In this section, we evaluate our proposed methods summarized in Algorithm 1, namely IBW and MD, on different experiments on both toy and real data. |
| Researcher Affiliation | Academia | Marguerite Petit-Talamon CREST, ENSAE Institut Polytechnique de Paris EMAIL Marc Lambert INRIA, Ecole Normale Supérieure DGA, French Procurement Agency EMAIL Anna Korba CREST, ENSAE Institut Polytechnique de Paris EMAIL |
| Pseudocode | Yes | Algorithm 1 MIG optimization with IBW or MD |
| Open Source Code | Yes | Our code is available at https://github.com/ margueritetalamon/VI-MIG. |
| Open Datasets | Yes | We evaluate our methods on two probabilistic inference tasks using classical datasets. The first one is Bayesian logistic regression (BLR) for two UCI datasets: breast_cancer (2 labels, d = 30) and wine (3 labels, d = 39). The second one aims to compute a Bayesian neural network (BNN) posterior on a regression task on the boston dataset using a single hidden layer neural network of 50 units (d = 601), and on the MNIST with a one layer neural network with 256 units (d = 203530). |
| Dataset Splits | Yes | The training ratio has been set to 0.5 for UCI datasets and 0.8 for MNIST. |
| Hardware Specification | Yes | All experiments (except MNIST) were conducted on a Mac Book Air (M3, 2024) with an Apple M3 processor and 16 GB of RAM. The MNIST experiments were run on an NVIDIA 50-90 GPU. |
| Software Dependencies | No | The paper mentions "Python scripts" and refers to "Real NVP architecture [Dinh et al., 2017]" and the corresponding code repository. However, it does not provide specific version numbers for Python, any deep learning framework (e.g., PyTorch, TensorFlow), or other key software libraries used in the implementation. |
| Experiment Setup | Yes | Initialization of the variational mixture: For N N a given number of components, we initialize the variational mixture by sampling the means in a ball of size [ s, s]d, where s R+ , and setting each covariance matrix to r Id, where r R+ . For the GD algorithm (mean optimization only, following Huix et al. [2024]), the variances are initialized in the same way but kept fixed during optimization. Optimization hyperparameters: We set the step-size γ, the number of iteration niter, the number of Monte Carlo samples to Bgrad = 10 for gradient estimation, and to BKL = 1000 for the KL objective estimation. Table 1: Hyperparameters |