Misclassification Detection via Class Augmentation

Fei Zhu, Xu-yao Zhang, Chuang Wang, Cheng-lin Liu

Keywords: misclassification detection, augmentation, uncertainty, confidence

Abstract: Despite the impressive performance in various pattern recognition tasks, deep neural networks (DNNs) are typically overconfident in their predictions, making it difficult to determine whether a test example is misclassified. In this paper, we propose a simple yet effective method of class augmentation (classAug) to address the challenge of misclassification detection in DNNs. Specifically, we increase the number of classes during training by assigning new classes to the samples generated using between-class interpolation. In spite of the simplicity, extensive experiments demonstrate that the misclassification detection performance of DNNs can be significantly improved by seeing more generated pseudo-classes during training. Additionally, we observe that DNNs trained with classAug are more robust on out-of-distribution examples and better calibrated. Finally, as a general regularization strategy, classAug can also enhance the original classification accuracy and few-shot learning performance.

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

One-sentence Summary: Class augmentation is proposed to address the misclassification detection in DNNs by learning additional synthetic classes during training.

Reviewed Version (pdf): https://openreview.net/references/pdf?id=PiNmmHl_n