School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore 639798, Singapore.
School of Computing Science, University of Glasgow, Singapore 567739, Singapore.
Neural Netw. 2021 Jul;139:24-32. doi: 10.1016/j.neunet.2021.02.015. Epub 2021 Feb 23.
Semi-supervised learning has largely alleviated the strong demand for large amount of annotations in deep learning. However, most of the methods have adopted a common assumption that there is always labeled data from the same class of unlabeled data, which is impractical and restricted for real-world applications. In this research work, our focus is on semi-supervised learning when the categories of unlabeled data and labeled data are disjoint from each other. The main challenge is how to effectively leverage knowledge in labeled data to unlabeled data when they are independent from each other, and not belonging to the same categories. Previous state-of-the-art methods have proposed to construct pairwise similarity pseudo labels as supervising signals. However, two issues are commonly inherent in these methods: (1) All of previous methods are comprised of multiple training phases, which makes it difficult to train the model in an end-to-end fashion. (2) Strong dependence on the quality of pairwise similarity pseudo labels limits the performance as pseudo labels are vulnerable to noise and bias. Therefore, we propose to exploit the use of self-supervision as auxiliary task during model training such that labeled data and unlabeled data will share the same set of surrogate labels and overall supervising signals can have strong regularization. By doing so, all modules in the proposed algorithm can be trained simultaneously, which will boost the learning capability as end-to-end learning can be achieved. Moreover, we propose to utilize local structure information in feature space during pairwise pseudo label construction, as local properties are more robust to noise. Extensive experiments have been conducted on three frequently used visual datasets, i.e., CIFAR-10, CIFAR-100 and SVHN, in this paper. Experiment results have indicated the effectiveness of our proposed algorithm as we have achieved new state-of-the-art performance for novel visual categories learning for these three datasets.
半监督学习在很大程度上缓解了深度学习对大量注释的强烈需求。然而，大多数方法都采用了一种常见的假设，即始终有来自同一类未标记数据的标记数据，这在实际应用中是不切实际和受限的。在这项研究工作中，我们专注于当未标记数据和标记数据的类别彼此不相交时的半监督学习。主要挑战是如何在它们彼此独立且不属于同一类别的情况下，有效地利用标记数据中的知识来标记未标记的数据。以前的最先进方法已经提出构建成对相似性伪标签作为监督信号。然而，这些方法通常存在两个问题：（1）所有以前的方法都由多个训练阶段组成，这使得很难以端到端的方式训练模型。（2）对成对相似性伪标签的质量强烈依赖限制了性能，因为伪标签容易受到噪声和偏差的影响。因此，我们建议在模型训练期间利用自监督作为辅助任务，以便标记数据和未标记数据将共享同一组替代标签，并且总体监督信号可以具有强正则化。通过这样做，所提出算法中的所有模块都可以同时进行训练，从而提高学习能力，因为可以实现端到端学习。此外，我们建议在构建成对伪标签时利用特征空间中的局部结构信息，因为局部属性对噪声更稳健。本文在三个常用的视觉数据集，即 CIFAR-10、CIFAR-100 和 SVHN 上进行了广泛的实验。实验结果表明了我们提出的算法的有效性，因为我们在这三个数据集的新视觉类别学习方面取得了新的最先进的性能。
