Sliced Optimal Transport Distances and Their Applications in Machine Learning
Optimal transport theory and its associated statistical distances, such as Wasserstein distances, play a pivotal role in machine learning, contributing significantly to advancements in areas like generative modeling, domain adaptation, inverse problems, and self-supervised learning. However, the computational cost of transport-based distances remains a significant obstacle, restricting their broader application. Sliced optimal transport-based distances have emerged as a viable and computationally efficient alternative, offering distinct statistical advantages. This presentation will explore the concept of generalized sliced optimal transport, discuss its unbalanced variants and its extensions to non-Euclidean spaces. Lastly, we will provide an overview of how these sliced distances are applied across various machine learning domains.
—