Faster Wasserstein Distance Estimation with the Sinkhorn Divergence

Authors: Lénaïc Chizat, Pierre Roussillon, Flavien Léger, François-Xavier Vialard, Gabriel Peyré

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We finally demonstrate the efficiency of the proposed estimators with numerical experiments.
Researcher Affiliation Academia 1: Laboratoire de Mathématiques d Orsay, CNRS, Université Paris-Saclay, Orsay, France 2: ENS, PSL University, Paris, France 3: Univ. Gustave Eiffel, CNRS, ESIEE Paris, Marne-la-Vallée, France
Pseudocode No The paper describes Sinkhorn's algorithm using mathematical equations, but it does not present it in a structured pseudocode or algorithm block.
Open Source Code Yes 2The code to reproduce these experiments is available at this webpage https://gitlab.com/ proussillon/wasserstein-estimation-sinkhorn-divergence.
Open Datasets No The paper assesses the estimators on 'synthetic problems' using 'n independent samples from µ and ν' or 'densities discretized on grids', but does not provide concrete access information (link, DOI, formal citation) for a publicly available or open dataset.
Dataset Splits No The paper discusses 'n independent samples' or 'discretized densities' for the distributions µ and ν, but it does not specify explicit dataset splits (e.g., percentages or counts) for training, validation, or testing to reproduce the experiments. The concept of train/validation/test splits in the context of machine learning model training is not applied to the data used in this paper.
Hardware Specification No No specific hardware details (e.g., GPU/CPU models, memory amounts, or detailed computer specifications) used for running the experiments were provided in the paper.
Software Dependencies No The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiments.
Experiment Setup Yes In Section 5, the paper states: 'for a target L1 error on the potential, we chose the largest λ and smallest n that achieve this error, with λ [0.1, 1] and n [10, 100000]' and 'We report the computational time using the Sinkhorn s iterations of Eq. (6) stopped when the ℓ1-error on the marginals is below 10 5.'