Основні поняття
ANEDL introduces EDL for outlier detection in Open-set SSL, improving uncertainty quantification and model performance.
Анотація
Introduction:
SSL methods assume labeled, unlabeled, and test data share the same distribution.
Open-set SSL deals with outliers not seen in labeled data.
ANEDL Framework:
Introduces EDL for outlier detection and adaptive negative optimization.
Utilizes Softmax and EDL for representation learning and uncertainty quantification.
Data Extraction:
"Our proposed method outperforms existing state-of-the-art methods across four datasets."
Related Works:
SSL methods categorized into consistency regularization and pseudo labeling.
Open-set SSL aims to detect outliers and classify inliers correctly.
Method:
ANEDL framework consists of shared feature extractor, Softmax head, and EDL head.
Adaptive Negative Optimization regulates EDL for inliers and outliers.
Experiments:
ANEDL outperforms SOTA methods in AUROC and error rate on CIFAR-10, CIFAR-100, and ImageNet-30.
Acknowledgments:
Supported by various research grants.
Статистика
"Our proposed method outperforms existing state-of-the-art methods across four datasets."