AI robustness analysis with consideration of corner cases

Corner cases are a set of high-risk data in deep learning (DL) systems, which could lead to incorrect and unexpected behaviors. To study corner cases' influence on DL models' robustness and stability, this paper implemented a research with corner case description and detection, as well as DL model's robustness measurement. Firstly, a corner case descriptor based on surprise adequacy was introduced for corner case data detection, its high values are proved useful to reflect incorrect behaviors in classification. Then, based on the proposed corner case descriptor, training dataset was updated by removing data having high possibility to be corner case, and utilized for model retraining. A practical robustness analysis method was applied to measure the robustness radius of both the original and the retrained DL models. Through experiments on MNIST data and industrial vision inspection data, an interesting phenomenon is found that, with consideration of corner case data removing in training data, DL models' robustness can be improved to some extent.