Orthogonal Estimation of Wasserstein Distances

Citations Over Time

Abstract

Wasserstein distances are increasingly used in a wide variety of applications in machine learning. Sliced Wasserstein distances form an important subclass which may be estimated efficiently through one-dimensional sorting operations. In this paper, we propose a new variant of sliced Wasserstein distance, study the use of orthogonal coupling in Monte Carlo estimation of Wasserstein distances and draw connections with stratified sampling, and evaluate our approaches experimentally in a range of large-scale experiments in generative modelling and reinforcement learning.

Related Papers

• Wasserstein Generative Adversarial Networks(2017)
• Wasserstein Auto-Encoders(2018)
• → Sliced Wasserstein Generative Models(2019)113 cited
• → Fused Gromov-Wasserstein Distance for Structured Objects(2020)78 cited