A Simulated Annealing Based Inexact Oracle for Wasserstein Loss Minimization

Authors: Jianbo Ye, James Z. Wang, Jia Li

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We applied the method to optimal transport with Coulomb cost and the Wasserstein non-negative matrix factorization problem, and made comparisons with the existing method of entropy regularization. By experiments, we demonstrate the effectiveness of Gibbs-OT for solving optimal transport with Coulomb cost (Benamou et al., 2016) and the Wasserstein non-negative matrix factorization (NMF) problem (Sandler & Lindenbaum, 2009; Rolet et al., 2016).
Researcher Affiliation Academia 1College of Information Sciences and Technology, The Pennsylvania State University, University Park, PA. 2Department of Statistics, The Pennsylvania State University, University Park, PA.. Correspondence to: Jianbo Ye <jxy198@ist.psu.edu>.
Pseudocode Yes Algorithm 1 Gibbs Sampling for Optimal Transport
Open Source Code No The paper mentions 'a C/C++ implementation' but does not provide any statement about open-sourcing the code or a link to a repository for the described methodology.
Open Datasets Yes one is a subset of MNIST handwritten digit images which contains 200 digits of 5 , and the other is the ORL 400-face dataset.
Dataset Splits No The paper mentions using datasets (MNIST, ORL) but does not specify any explicit train/validation/test dataset splits or reference predefined splits with citations.
Hardware Specification Yes On a single-core of a 3.3 GHz Intel Core i5 CPU, the average time spent for each epoch for these two datasets are 0.84 seconds and 16.8 seconds, respectively.
Software Dependencies No The paper mentions 'a C/C++ implementation' and 'Mosek' as a solver but does not provide specific version numbers for any software components.
Experiment Setup Yes For Gibbs-OT, we use a geometric temperature scheme such that T = 2.0(1/l4)n/l/N at the n-th iteration, where l is the max iteration number. In particular, we set K = 40, γ = 2.0 for both datasets.