OpenHRD: Hierarchical representation decoupling for open-world semi-supervised learning

Abstract

In realistic open-world semi-supervised scenarios, novel classes always emerge from unlabeled data, which leads to the performance degradation of existing semi-supervised learning (SSL) methods. The absence of any supervisory signals for novel classes hinders the model from learning disentangled representations, causing the model to confuse known classes with novel classes and generate significant prediction bias towards known classes. In this paper, we propose a hierarchical representation decoupling approach, named OpenHRD, which jointly decouples representations at the instance level and class level for different samples to address this challenge. Specifically, at the instance level, we impose representation constraints on the most similar instance pairs with highest representation similarity to mitigate representation confusion between samples. Furthermore, we also propose an adaptive pseudo-label debiasing regularization method for unlabeled instances at the instance level, which effectively alleviate the prediction bias toward known classes during the model training. At the class level, we introduce an inter-class contrastive learning strategy for novel classes to enlarge the representation distinction between each novel class and other classes. Extensive experimental results on various settings over CIFAR-10, CIFAR-100, and ImageNet-100 demonstrate the superior performance of the proposed OpenHRD. We will release the code at: https://github.com/srxhlife/OpenHRD.