ReLSL: Reliable Label Selection and Learning Based Algorithm for Semi-Supervised Learning

Deep neural networks have achieved remarkable success in many visual representation fields, such as object detection, recognition, etc. However, requiring the large quantity of well labeled data for training is one of their most prevalent limitations. Many real-world classification applications are concerned with samples that are not presented in standard benchmark datasets, and building large labeled dataset for each new task to be learned is not practically feasible. Although enormous quantities of unlabeled data are accessible and can be collected with minimal effort, the data labeling process is still extremely expensive. Semi-supervised learning (SSL) provides a way to improve a model's performance with the surplus of unlabeled data when only limited labeled data are available. However, when the labeled data is extremely scarce, the performance of the existing SSL algorithms can be severely affected. For example, on the prevalent CIFAR-10 dataset, when each class is supported by only one label sample, the accuracy of most SSL algorithms degrades seriously. The problem is mainly manifested as: the initial informative information for classification is extremely limited, the model faces cold-start problem; in the process of training, the proportion of pseudo-label noise is difficult to control and the model has a much larger potential risk to be collapsed. In this paper, we propose a Reliable Label Selection and Learning (ReLSL) framework, which tackles the problem semi-supervised deep learning facing when only few-shot labeled image data is available. In brief, we exploit synergies among unsupervised learning, SSL and robust learning to bootstrap additional reliable labels for robust network training. For the unsupervised learning, it is used to ease the problem of cold-start under scarce labeled conditions. For SSL and robust learning, they are used to obtain good learning performance in the presence of noise labels. To be specific, for our whole ReLSL, we first implement Anchor Neighborhood Discovery (AND), an unsupervised learning algorithm to extract features of all training samples, and then obtain their pseudo-label by applying graph-based label propagation algorithm. Then, in order to screen out more reliable and informative samples, a pseudo-label learning and calibration strategy is proposed that comprehensively considers the mean and consistency of the sample's output, and conduct effective screening of samples through Small-Loss theory. After obtaining the dataset with extended labels, considering that a certain proportion of label noise is inevitably introduced into the training set, we therefore propose two strategies to train a robust SSL model, namely, a Label-Smoothing strategy (LS) for regularizing labels from being too sharp, thus reducing noise label interference to loss function; Mean-Shifting Correction strategy (MSC) for reducing the risk of sample output deviation. As a result, the proposed ReLSL achieves state-of-the-art performance on CIFAR-10/100, SVHN, STL-10 and Mini-ImageNet across a variety of SSL conditions with the CNN-13, WRN-28-2 and ResNet-18 networks. In particular, our framework achieves a 6.78% accuracy boosting on CIFAR-10 with only 10 labeled data under WRN-28-2. Moreover, our algorithm can achieve the test error of 6.39±0.47% with only 100 labeled data under CNN-13, which is comparable to the one with typical SSL under 4000 labeled conditions. © 2022, Science Press. All right reserved.