Implications of Gaussian Process Kernel Mismatch for Out-of-Distribution Data
Abstract
Gaussian processes provide reliable uncertainty estimates in nonlinear modeling, but a poor choice of the kernel can lead to poor generalization. Although learning the hyperparameters of the kernel typically leads to optimal generalization on in-distribution test data, we demonstrate issues with out-of-distribution test data. We then investigate three potential solutions-- (1) learning the smoothness using a discrete cosine transform, (2) assuming fatter tails in function-space using a Student-$t$ process, and (3) learning a more flexible kernel using deep kernel learning--and find some evidence in favor of the first two.
Cite
Text
Coker and Doshi-Velez. "Implications of Gaussian Process Kernel Mismatch for Out-of-Distribution Data." ICML 2023 Workshops: SPIGM, 2023.Markdown
[Coker and Doshi-Velez. "Implications of Gaussian Process Kernel Mismatch for Out-of-Distribution Data." ICML 2023 Workshops: SPIGM, 2023.](https://mlanthology.org/icmlw/2023/coker2023icmlw-implications/)BibTeX
@inproceedings{coker2023icmlw-implications,
title = {{Implications of Gaussian Process Kernel Mismatch for Out-of-Distribution Data}},
author = {Coker, Beau and Doshi-Velez, Finale},
booktitle = {ICML 2023 Workshops: SPIGM},
year = {2023},
url = {https://mlanthology.org/icmlw/2023/coker2023icmlw-implications/}
}