How robust are discriminatively trained zero-shot learning models?

作者：

Highlights：

• Novel analyses on corruption robustness of discriminative ZSL models.

• Release of three new benchmark datasets for ZSL robustness analyses.

• Class imbalance and model strength of ZSL methods leads to severe robustness issues.

• The pseudo-robustness effect is present for adversaries but not corruptions.

• Corruption defense methods improve clean accuracy, provide new strong baselines.

摘要

Highlights•Novel analyses on corruption robustness of discriminative ZSL models.•Release of three new benchmark datasets for ZSL robustness analyses.•Class imbalance and model strength of ZSL methods leads to severe robustness issues.•The pseudo-robustness effect is present for adversaries but not corruptions.•Corruption defense methods improve clean accuracy, provide new strong baselines.

论文关键词：Zero-shot learning,Robust generalization,Adversarial robustness

论文评审过程：Received 26 September 2021, Revised 4 December 2021, Accepted 18 January 2022, Available online 23 January 2022, Version of Record 1 February 2022.

论文官网地址：https://doi.org/10.1016/j.imavis.2022.104392