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..
Non-Asymptotic Analysis Of Data Augmentation For Precision Matrix Estimation
Authors: Lucas Morisset, Adrien Hardy, Alain Durmus
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We support our theoretical results with numerical experiments. In this section, we illustrate Theorem 1 and Theorem 2 on real datasets. We use MNIST and CIFAR10, consisting of 70,000 labeled 28 x 28 images and 60,000 labeled 32 x 32 images, respectively... |
| Researcher Affiliation | Collaboration | Lucas Morisset Qube Research and Technologies & École Polytechnique Adrien Hardy Qube Research and Technologies Alain Durmus École Polytechnique |
| Pseudocode | No | The paper describes methods and derivations but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured code-like steps for a procedure. |
| Open Source Code | No | Although we do not provide the code for our numerical experiments, the definition of ˆEAug(λ) in Section 3 is self sufficient to replicate our results. |
| Open Datasets | Yes | In this section, we illustrate Theorem 1 and Theorem 2 on real datasets. We use MNIST and CIFAR10, consisting of 70,000 labeled 28 x 28 images and 60,000 labeled 32 x 32 images, respectively... Our numerical experiments have been performed on the well known, open-access, dataset MNIST... |
| Dataset Splits | No | For both datasets, we denote by X = [X1, . . . , Xn] Rd n the matrix formed by the first n samples, for varying n > 0. To approximate EX(λ) and EAug(λ), we use the sample covariance matrix ˆΣX computed from all available samples (70,000 for MNIST, 60,000 for CIFAR10), and consider the proxies ED X(λ) := 1 d RX(λ) ˆΣ 1 X 2 F and ED Aug(λ) := 1 d RAug(λ) ˆΣ 1 X 2 F... The paper describes data usage for an estimation task, not traditional training/test/validation splits for model evaluation. |
| Hardware Specification | No | Although the aspect of computer ressources were not discussed in the paper, we believe that our estimator doesn t require any particular computing or memory ressources to be efficiently implemented. For this reason, we deemed this discussion as unnecessary. |
| Software Dependencies | No | The paper does not explicitly state any specific software dependencies or their version numbers, such as programming languages, libraries, or frameworks used for implementation. |
| Experiment Setup | Yes | MNIST. We discard the labels, normalize pixel values to [0, 1], and add pixel-level Gaussian noise with standard deviation σ = 0.1 to ensure that the covariance matrix ΣX is well-conditioned. CIFAR10. We discard the labels and convert images to grayscale. For both datasets, we denote by X = [X1, . . . , Xn] Rd n the matrix formed by the first n samples, for varying n > 0. It also shows λ [10^-3, 1] in figures and discusses α = m/(n + m). |