Bootstrap Generalization Ability from Loss Landscape Perspective

Abstract

Domain generalization aims to learn a model that can generalize well on the unseen test dataset, i.e., out-of-distribution data, which has different distribution from the training dataset. To address domain generalization in computer vision, we introduce the loss landscape theory into this field. Specifically, we bootstrap the generalization ability of the deep learning model from the loss landscape perspective in four aspects, including backbone, regularization, training paradigm, and learning rate. We verify the proposed theory on the NICO++, PACS, and VLCS datasets by doing extensive ablation studies as well as visualizations. In addition, we apply this theory in the ECCV 2022 NICO Challenge1 and achieve the 3rd place without using any domain invariant methods.

Cite

Text

Chen et al. "Bootstrap Generalization Ability from Loss Landscape Perspective." European Conference on Computer Vision Workshops, 2022. doi:10.1007/978-3-031-25075-0_34

Markdown

[Chen et al. "Bootstrap Generalization Ability from Loss Landscape Perspective." European Conference on Computer Vision Workshops, 2022.](https://mlanthology.org/eccvw/2022/chen2022eccvw-bootstrap/) doi:10.1007/978-3-031-25075-0_34

BibTeX

@inproceedings{chen2022eccvw-bootstrap,
  title     = {{Bootstrap Generalization Ability from Loss Landscape Perspective}},
  author    = {Chen, Huanran and Shao, Shitong and Wang, Ziyi and Shang, Zirui and Chen, Jin and Ji, Xiaofeng and Wu, Xinxiao},
  booktitle = {European Conference on Computer Vision Workshops},
  year      = {2022},
  pages     = {500-517},
  doi       = {10.1007/978-3-031-25075-0_34},
  url       = {https://mlanthology.org/eccvw/2022/chen2022eccvw-bootstrap/}
}