노이즈와 라벨 스무딩 사이의 지수 보간법을 이용한 인공지능 강인성 향상

Abstract

While modern deep learning methods exhibit cutting-edge performances on a variety of machine learning tasks, they are often prone to real-world driven data corruptions. To cope with such problem, different augmentation methods and model architectures have been proposed to make neural network robust against common data corruptions. In this paper, we propose a simple and efficient data augmentation strategy that increases model’s robustness on visual recognition tasks. On several common corruption benchmarks including MNIST-C, CIFAR10-C, and CIFAR-100-C, our method increased model’s robustness both as a standalone method and on the top of previous state-of-the-art augmentation methods. Since EINS is a general augmentation method that is neither model-specific nor data-specific, it can be easily exploited with other methods to further increase model’s robustness in most circumstances.

본 논문은 인공지능의 강인성을 향상시키는 데이터 증강 방법인 EINS를 제안한 다. 이론적 특성 유도과 인공지능 강인성 벤치마크위에서의 실험을 통해 EINS의 실효성을 입증한다.