Original title: Understanding Data Augmentation from a Robustness Perspective
Authors: Zhendong Liu, Jie Zhang, Qiangqiang He, Chongjun Wang
In visual recognition, data augmentation greatly enhances model robustness, yet its mechanisms remain unclear in many existing methods. This article takes a dual approach—theoretical and empirical—to unravel this phenomenon. Theoretically, it frames data augmentation within game theory, diving into empirical evaluations of emblematic strategies. Findings reveal these techniques fuel mid- and high-order game interactions universally across datasets and augmentation methods. Introducing a streamlined robustness proxy simplifies assessment, offering deep insights into model dynamics and system robustness. These insights provide a fresh perspective on evaluating model safety and robustness in visual recognition tasks, uncovering intricate correlations and unveiling the underlying dynamics of model-game interactions.
Original article: https://arxiv.org/abs/2311.12800