Presentation Information
[4Yin-A-32]Soft Label Construction via Integration of Confidence and Neighborhood Information
〇Mao Iwasaki1, Kosuke Sugiyama1, Masato Uchida1 (1. Waseda University)
Keywords:
deep learning,weakly supervised learning,soft label,label quality,additional supervision
In real-world machine learning applications, ideal-sized datasets are not always available. When sufficient training data is scarce, enriching the information content of individual labels is crucial. This study proposes a soft label construction method that integrates annotation confidence and neighborhood information in the feature space to represent labels as probability distributions. Experimental results using benchmark datasets confirmed that the proposed method improves classification accuracy and can reduce the error relative to the true distribution compared to existing methods such as Label Smoothing.