ENLAS

Abstract

Nowadays, Deep Neural Networks (DNNs) have achieved numerous success in various computer vision tasks such as image classification, image segmentation and object detection. Since DNNs are over-parameterized, they are prone to overfitting when corrupted labels are included. This phenomenon is named as memorization of noisy labels. In this paper, we propose ENLAS (Ensemble of Loss Magnitude and Alignment Score): Robust Learning via Ensemble Framework for Handling Noisy Labels. ENLAS is mainly composed of three parts: 1) Feature extractor, 2) Gaussian-Mixture-Models, 3) Importance factor. ENLAS takes advantage of both loss magnitude and alignment score by ensemble without additional hyperparameter. In addition, the quality of feature extractor is maintained regardless of noise ratio by utilizing gradually increasing threshold. Under the proposed framework, we conduct experiments with respect to various noise rate in CIFAR-10 and CIFAR-100. Experimental results validate that ENLAS outperforms other baseline methods on various benchmark datasets.