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].
Fundamental Bias in Inverting Random Sampling Matrices with Application to Sub-sampled Newton
Authors: Chengmei Niu, Zhenyu Liao, Zenan Ling, Michael W. Mahoney
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5. Numerical Experiments of De-biased SSN Figure 2 assesses the impact of the sketch size m on both relative error ( ˇβT β 2 H/ ˇβ0 β 2 H) and running time of de-biased SSN employing approximate ridge leverage score sampling (SSN-ARLev), in comparison to the Newton LESS method (Derezi nski et al., 2021a) based on random projection. The results in Figure 2 demonstrate that the proposed de-biased SSN consistently outperforms Newton LESS across all tested sketch size m, exhibiting a superior convergence complexity trade-off. |
| Researcher Affiliation | Academia | 1School of Electronic Information & Communications, Huazhong University of Science & Technology, Wuhan, Hubei, China. 2ICSI, LBNL, and Department of Statistics, University of California, Berkeley, USA. Correspondence to: Zhenyu Liao <EMAIL>. |
| Pseudocode | No | The paper describes steps and methods in regular paragraph text and mathematical derivations. There are no explicitly labeled sections such as 'Algorithm' or 'Pseudocode' nor any structured, code-like formatted procedures. |
| Open Source Code | No | The paper states: 'Further implementation details, including those for the SRHT, are available in the public repository provided by Derezi nski et al. (2021a) at https://github.com/lessketching/newtonsketch.' However, this link refers to a third-party tool ('Newton-LESS') that the authors used as a baseline, not their own implementation code for the novel methodology presented in this paper. |
| Open Datasets | Yes | We solve the following logistic regression problem... sampled from both MNIST (Le Cun et al., 1998) and CIFAR-10 (Krizhevsky, 2009) datasets |
| Dataset Splits | No | The paper states: 'For MNIST data matrix, we have n = 213 and d = 27, and for CIFAR-10 data, we have n = 214 and d = 28.' It describes dataset sizes and a binary classification setup, but does not provide specific training, validation, or test dataset splits. |
| Hardware Specification | No | The paper mentions 'wall-clock time' in its numerical experiments but does not provide specific hardware details such as GPU or CPU models, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions using 'torchvision.transforms from Py Torch' for pre-processing images, but it does not specify the version number for PyTorch or any other software components. |
| Experiment Setup | Yes | And we fix in Figures 2 and 3 the ridge regularization parameter to λ = 10 2 for both MNIST and CIFAR-10 data in Section 5." and "In our experiments, the first-order methods, Gradient Descent and Stochastic Gradient Descent, are used with a fixed step size. ... second-order methods... employ step sizes that are dynamically adjusted using a line search algorithm based on the Armijo condition (Bonnans et al., 2006, Chapter 3)." |