Robust semi-supervised learning to label bias

Abstract

It is well known that label bias hinders the practical applications of deep learning approaches in the real-world. Since deep learning models optimally learn the biases existing in the dataset, they cannot generalize well to fair, real-world requirement. In real-world, especially in applications where safety and reliability are required, it is important to ensure that the model produces fair predictions even if the model is trained on a biased data. In this dissertation, we present a semi-supervised learning method that allows the model to be fair even in the training data where bias exists. When biased data is trained with typical semi-supervised learning methods, performances severely degrades, in some cases far less than that of the supervised learning counterpart. In this paper, we design a framework called Prototypical Semantic Alignment (PSA) to effectively prevent this problem. To this end, we propose 1) Online Clustering, an algorithm that clusters labeled data in an online-manner and 2) Semantic Alignment Loss which enforces the unlabeled images to be attracted to the most similar prototype obtained through clustering. As a result, we show significant performance gains by consuming additional unlabeled images in long-tailed distribution data and label-scarce scenarios, which are real examples of label bias.