When does loss-based prioritization fail?

TL;DR

Treating examples with higher loss as more important doesn't work in some cases.

Venue
In ICML 2021 Workshop of SubSetML: Subset Selection in Machine Learning: From Theory to Practice.
BibTeX
@article{hu2021does,
  title={When does loss-based prioritization fail?},
  author={Niel Teng Hu and Xinyu Hu and Rosanne Liu and Sara Hooker and Jason Yosinski},
  year={2021},
  eprint={2107.07741},
  archivePrefix={arXiv},
  primaryClass={cs.LG}
}
Date