잡음에 강인한 수면 단계 판별 네트워크를 위한 학습 방법

With development of deep learning technology, there has been increasing attention to exploit it on various domain such as healthcare, autonomous driving, etc. Especially for mobile devices for health management and treatment, deep models are vulnerable to noise, resulting in performance degradation, since they are trained with inputs acquired in the best environment. This paper proposes input transformation during model’s training to make it embrace bigger range of input’s perturbation. Experimental results show that the proposed method gives better robustness to not only adversarial noise but also other types of noises.