Screening Sinkhorn Algorithm for Regularized Optimal Transport

Authors: Mokhtar Z. Alaya, Maxime Berar, Gilles Gasso, Alain Rakotomamonjy

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate the efficiency of SCREENKHORN on complex tasks such as dimensionality reduction and domain adaptation involving regularized optimal transport.
Researcher Affiliation Collaboration Mokhtar Z. Alaya LITIS EA4108 University of Rouen Normandy mokhtarzahdi.alaya@gmail.com Maxime Bérar LITIS EA4108 University of Rouen Normandy maxime.berar@univ-rouen.fr Gilles Gasso LITIS EA4108 INSA, University of Rouen Normandy gilles.gasso@insa-rouen.fr Alain Rakotomamonjy LITIS EA4108 University of Rouen Normandy and Criteo AI Lab, Criteo Paris alain.rakoto@insa-rouen.fr
Pseudocode Yes Algorithm 1: SCREENKHORN(C, η, µ, ν, nb, mb)
Open Source Code No The paper states that the authors 'have implemented our SCREENKHORN algorithm in Python' and 'based our code on the ones of Python Optimal Transport toolbox (POT) [15]' but does not provide a link or explicit statement that their specific implementation of SCREENKHORN is open-source or publicly available.
Open Datasets Yes We have implemented our SCREENKHORN algorithm in Python and used the L-BFGS-B of Scipy. Regarding the machine-learning based comparison, we have based our code on the ones of Python Optimal Transport toolbox (POT) [15] and just replaced the sinkhorn function call with a screenkhorn one. (...) Here, we analyse the impact of using SCREENKHORN instead of SINKHORN in a complex machine learning pipeline. Our two applications are a dimensionality reduction technique, denoted as Wasserstein Discriminant Analysis (WDA), based on Wasserstein distance approximated through Sinkhorn divergence [16] and a domain-adaptation using optimal transport mapping [10], named OTDA. (...) We have used a toy problem involving Gaussian classes with 2 discriminative features and 8 noisy features and the MNIST dataset.
Dataset Splits No The paper mentions 'different values of η, the regularization parameter, the allowed budget nb ranging from 0.01 to 0.99, different values of n and m' and measures 'marginal violations', 'running time', and 'relative divergence difference'. However, it does not specify explicit train/validation/test splits with percentages, absolute counts, or references to predefined standard splits for the datasets used.
Hardware Specification No The paper does not provide any specific hardware details such as CPU/GPU models, memory specifications, or types of computing resources used for running the experiments.
Software Dependencies No The paper mentions 'Python', 'L-BFGS-B of Scipy', and 'Python Optimal Transport toolbox (POT) [15]' but does not provide specific version numbers for any of these software components.
Experiment Setup Yes For all applications, we have set η = 1 unless otherwise specified. (...) for SCREENKHORN, the L-BFGS-B algorithm is stopped when the largest component of the projected gradient is smaller than 10 6, when the number of iterations or the number of objective function evaluations reach 105.