준지도학습을 이용한 노이즈 데이터 학습 방법

Abstract

One of the major problems of modern neural networks is that models are vulnerable to data noise. In order to resolve this concern, research on removing noisy data has been actively conducted. However, there is a limitation that information in the removed noisy cannot be utilized for learning. In this paper, we propose an effective learning method based on FixMatch, one of the widely-used semi-supervised learning methods, and devise additional techniques that are effective for noisy labels such as model ensemble and parameter scheduling. Our experiments show that the proposed method achieves the best accuracy under every noise rate condition verifying that the proposed model is robust to data noise.