Doing Fast Adaptation Fast: Conditionally Independent Deep Ensembles for Distribution Shifts
Lists
Only you see this
Abstract
Classifiers in a diverse ensemble capture distinct predictive signals, which is valuable for datasets containing multiple strongly predictive signals. Performing fast adaptation at test time allows us to generalize to distributions where certain signals are no longer predictive, or to avoid relying on sensitive or protected attributes. However, ensemble learning is often expensive, even more so when we need to enforce diversity constraints between the high-dimensional representations of the classifiers. Instead, we propose an efficient and fast method for learning ensemble diversity. We minimize conditional mutual information of the output distributions between classifiers, a quantity which can be cheaply and exactly computed from empirical data. The resulting ensemble contains individually strong predictors that are only dependent because they predict the label. We demonstrate the efficacy of our method on shortcut learning tasks. Performing fast adaptation on our ensemble selects shortcut-invariant models that generalize well to test distributions where the shortcuts are uncorrelated with the label.
BibTeX
@inproceedings{yang2023doing,
author = {Wanqian Yang and Aahlad Manas Puli and Andrew Gordon Wilson and Rajesh Ranganath},
title = {{Doing Fast Adaptation Fast: Conditionally Independent Deep Ensembles for Distribution Shifts}},
booktitle = {International Conference on Learning Representations},
year = {2023},
url = {https://openreview.net/forum?id=17RDXeF-skZ},
}Loading related content...