Hierarchical Gaussian Processes with Wasserstein-2 Kernels
Machine Learning
2022-02-02 v2 Machine Learning
Abstract
Stacking Gaussian Processes severely diminishes the model's ability to detect outliers, which when combined with non-zero mean functions, further extrapolates low non-parametric variance to low training data density regions. We propose a hybrid kernel inspired from Varifold theory, operating in both Euclidean and Wasserstein space. We posit that directly taking into account the variance in the computation of Wasserstein-2 distances is of key importance towards maintaining outlier status throughout the hierarchy. We show improved performance on medium and large scale datasets and enhanced out-of-distribution detection on both toy and real data.
Cite
@article{arxiv.2010.14877,
title = {Hierarchical Gaussian Processes with Wasserstein-2 Kernels},
author = {Sebastian Popescu and David Sharp and James Cole and Ben Glocker},
journal= {arXiv preprint arXiv:2010.14877},
year = {2022}
}