Effective labeling on unlabeled data for imbalanced training dataset

Abstract

When a small amount of labeled data and a large amount of unlabeled data exist, the research on semi-supervised learning that trains a network by efficiently utilizing the unlabeled data is actively conducting. Existing studies about semi-supervised learning only deal with situations in which the class distribution of training dataset is balanced. If class distribution of training dataset is imbalanced, those method makes the network over-fit to a class with a lot of data (major class), so the network produces wrong pseudo-label with high probability for a class with less data (minor class). In this thesis, we propose a training framework that prevents the over-fitting problem when the class distribution is imbalanced by making reliable pseudo-label of unlabeled data. We separately design the auxiliary labeling network to focus only on making pseudo-label. In order to utilize the labeled data and unlabeled data together, the labeling network has an auto-encoder-based structure. The proposed method can be applied without changing the main classification network because it increases the labeled data by producing high-reliability pseudo-labeled data through a auxiliary labeling network. In addition, a more accurate pseudo-label is produced by proposing the distance loss function of the latent feature considering the class distribution. Part of the prepared pseudo-labeled data is progressively added to the training dataset to mitigate the imbalance of class distribution.