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].
Wasserstein Coreset via Sinkhorn Loss
Authors: Haoyun Yin, Yixuan Qiu, Xiao Wang
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our approach significantly outperforms existing methods in terms of sample selection quality, computational efficiency, and achieving a smaller Wasserstein distance. We have demonstrated the superior performance of our method in extensive simulation studies and practical applications with image data. |
| Researcher Affiliation | Academia | Haoyun Yin EMAIL Department of Statistics Purdue University; Yixuan Qiu EMAIL School of Statistics and Data Science Shanghai University of Finance and Economics; Xiao Wang EMAIL Department of Statistics Purdue University |
| Pseudocode | Yes | Algorithm 1 Algorithm for Wasserstein coreset via Sinkhorn loss (WCSL); Algorithm 2 Sinkhorn Loss Computation and Differentiation (SLCD) |
| Open Source Code | No | The paper refers to third-party tools like 'Pot: Python optimal transport' (Flamary et al., 2021) and 'OTT: A JAX toolbox for all things Wasserstein' (Cuturi et al., 2022), but does not provide an explicit statement or link for the source code of the methodology described in this paper. |
| Open Datasets | Yes | To further demonstrate the practical applicability of the WCSL method, we applied it to image datasets, the MNIST and Fashion MNIST datasets. |
| Dataset Splits | No | The paper describes generating coresets of various sizes (e.g., 'n = 1000 points each, reduced to coresets of B = 100 points' or 'coresets of sizes from 25 to 400'), which are subsets of the data used for training. However, it does not provide explicit training/test/validation splits with percentages, absolute counts, or references to standard splits for reproducible model evaluation, but rather uses the coresets for training and implies evaluation on unspecified remaining data. |
| Hardware Specification | No | The paper mentions 'modern computing hardware, such as GPUs' and 'modern hardware such as GPUs' as general capabilities, but does not provide specific details (e.g., model names, memory, CPU specifications) about the hardware used to run the experiments. |
| Software Dependencies | No | The paper mentions software components like 'L-BFGS', 'Adam update', 'OTT-JAX library', and 'Pot: Python optimal transport' but does not specify their version numbers, which is necessary for reproducible ancillary software details. |
| Experiment Setup | Yes | To evaluate the performance of the WCSL method, we compared it against several alternative sampling techniques: random sampling, the centroids obtained via k-means clustering, the SCCP method (Mak & Joseph, 2018), and kernel thinning (Dwivedi & Mackey, 2021)... The experiment settings include datasets of n = 1000 points each, reduced to coresets of B = 100 points. The three configurations tested are a normal distribution, a 2-dimensional spiral, and a circular distribution... We use a 100-dimensional t-distribution and a 100-dimentional normal distribution with n = 10000 points as our datasets, and reduce to coresets of various sizes, ranging from B = 25 to B = 400... We fix the regularization parameter λ 1 to be 0.001... The maximum number of iterations is 10000 for Greenkhorn and 1000 for other algorithms... we fix n = 150, m = 200, and consider varying dimensions p = 1, 10, 50. The Sinkhorn regularization parameters compared are λ 1 = 0.01, 0.001... We set εlbfgs = 10 6, and let εott = p max{n, m} εlbfgs. |