Iterative Semi-Supervised Learning Using Softmax Probability

For the classification problem in practice, one of challenging issues is to obtain enough labeled data for training. Moreover, even if such has been sufficiently accumulated, most datasets often exhibit long-tailed distribution with heavy class imbalance, which results a biased model towards majority class. To alleviate semi-supervised learning methods using additional unlabeled have considered. However, as matter course, accuracy much lower than that from supervised learning. In this study, under assumption available, we propose iterative algorithms, iteratively correct labeling extra based on softmax probabilities. The show proposed algorithms provide high validate tested two scenarios: balanced dataset and imbalanced dataset. Under both scenarios, our provided higher previous state-of-the-arts. Code available at .