Learning Disentangled Representations for Environment Inference in Out-of-distribution Generalization
Dongqi Li (Beijing Jiaotong University), Zhu Teng (Beijing Jiaotong University), Li Qirui (AFCtech), Wang Ziyin (AFCtech), Baopeng Zhang (BJTU),* Jianping Fan (Lenovo)
The 34th British Machine Vision Conference
Abstract
Machine learning models often generalize poorly to out-of-distribution (OOD) data as a result of relying on features that are spuriously correlated with the label during training. To deal with this issue, environment inference methods are proposed to learn invariant predictors without environment labels. Previous environment inference works often employ Empirical risk minimization (ERM) as a reference model for environment inference because they assume ERM captures spurious features due to its inductive bias. In this work, we show that using ERM as a reference model has a pitfall in environment inference because it does not effectively capture spurious features. To this end, we propose a disentangled representation method by designing a variational auto-encoder to capture spurious features for environment inference without environment labels. Extensive experiments demonstrate that the proposed method outperforms other methods on both synthetic and real-world datasets.Video
Citation
@inproceedings{Li_2023_BMVC,
author = {Dongqi Li and Zhu Teng and Li Qirui and Wang Ziyin and Baopeng Zhang and Jianping Fan},
title = {Learning Disentangled Representations for Environment Inference in Out-of-distribution Generalization},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0538.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by