Safe and Sparse Newton Method for Entropic-Regularized Optimal Transport

Authors: Zihao Tang, Yixuan Qiu

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Various numerical experiments are conducted to demonstrate the effectiveness of the proposed algorithm in solving large-scale OT problems. In this section, we test the performance of the proposed SSNS algorithm on various numerical experiments.
Researcher Affiliation Academia Zihao Tang School of Statistics and Data Science Shanghai University of Finance and Economics tangzihao@stu.sufe.edu.cn Yixuan Qiu School of Statistics and Data Science Shanghai University of Finance and Economics qiuyixuan@sufe.edu.cn
Pseudocode Yes Algorithm 1 Sparsifying the Hessian matrix.
Open Source Code Yes An efficient implementation is included in the Reg OT Python package2. 2https://github.com/yixuan/regot-python.
Open Datasets Yes (Fashion-)MNIST: OT between a pair of images from the MNIST [15] or Fashion-MNIST [37] dataset. Image Net: OT between two categories of images from the Image Net dataset [9].
Dataset Splits No No specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology for training, validation, and test sets) was found. The paper describes problem sizes and data construction but not data partitioning for model training/evaluation.
Hardware Specification Yes All experiments in this article are conducted on a personal computer with an Intel i9-13900K CPU, 32 GB memory, and a Ubuntu 24.10 operating system.
Software Dependencies No The paper mentions 'Reg OT Python package' and 'Ubuntu 24.10 operating system' but does not specify version numbers for other critical software dependencies (e.g., Python, PyTorch, or other libraries).
Experiment Setup Yes For entropic-regularized OT, a commonly-used criterion to evaluate optimality is the marginal error of the estimated transport plan... Then we consider two settings of the regularization parameter, η = 0.01 and η = 0.001. Default values: µ0 = 1, ν0 = 0.01, cl = 0.1, cu = 1, κ = 0.001, γ = 1, ρ0 = 1/4 setting (ξ[0], ξ[1], ξ[2], ξ[3]) = (1, 0.5, 0.25, 0.1) gives reasonably good performance in most numerical experiments.