A Dynamic Mixup Approach Towards Improved Robustness of Classifiers

Sanskriti, Seungwoo Jang, Donghee Kim, Soobin Cha, Dongwon Kim, Kwangsu Kim

Keywords: robustness, data shift, distribution shift, data augmentation

Abstract: The robustness of image classifiers has been an extensive area of growing research. Current methods typically rely on data augmentation techniques to simulate distribution shifts based on image corruptions. These techniques mainly consider linearity to generate synthetic samples. However, corruptions that occur in the real world are more complex and unlikely to follow a linear drift. We introduce an adaptation of the mixup approach, Dynamic Mixup, as our data processing technique. Dynamic Mixup uses a simple mixing strategy to combine augmented versions considering the non-linearity that exists between them. Training on these samples encourages learning representations robust to new or unseen distortions. Our experimental findings reveal that Dynamic Mixup outperforms the previous methods with improved robustness in image and object classification tasks.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5355