Sliced-Wasserstein Estimation with Spherical Harmonics as Control Variates

Authors: Rémi Leluc, Aymeric Dieuleveut, François Portier, Johan Segers, Aigerim Zhuman

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

Reproducibility Variable Result LLM Response
Research Type Experimental Several numerical experiments demonstrate the superior performance of SHCV against state-of-the-art methods for SW distance computation.
Researcher Affiliation Academia 1CMAP, Ecole Polytechnique, Institut Polytechnique de Paris, France 2CREST, ENSAI, France 3ISBA/LIDAM, UCLouvain, Belgium.
Pseudocode Yes Algorithm 1 Sliced Wasserstein Monte Carlo
Open Source Code Yes For ease of reproducibility, numerical details are in the appendix and the code is available here.
Open Datasets Yes Then we focus on SW2 2(µm, νm) between 3D point clouds from the Shape Net Core dataset of Chang et al. (2015)... we consider the image classification task on the digits dataset of Alpaydin & Kaynak (1998).
Dataset Splits No We use a train/test split of size 80/20 giving Ntrain = 1 437 training and Ntest = 360 testing images, respectively, where each set is balanced. No explicit validation split is mentioned.
Hardware Specification Yes The experiments were performed on a laptop Intel Core i7-10510U CPU 1.80GHz 8.
Software Dependencies Yes In Algorithm 8, random orthogonal matrices R U(Od(R)) are generated using the function ortho group of the Python package scipy (Virtanen et al., 2020).
Experiment Setup Yes The degrees of the spherical harmonics are reported in Table 9 in Appendix D.2. ... The SHCV estimate uses spherical harmonics up to degree 2L = 16 leading to s = 152 control variates.