2020/12/01 by Boaz Lerner, Lerner, Boaz, Guy Shiran +3
Computer Science · #Advanced Neural Network Applications #Computer Vision and Pattern Recognition (cs.CV) #Domain Adaptation and Few-Shot Learning #FOS: Computer and information sciences #Machine Learning (cs.LG) #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.2012.00504
openalex publication_date 2020/12/01 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Recently, Semi-Supervised Learning (SSL) has shown much promise in leveraging\nunlabeled data while being provided with very few labels. In this paper, we\nshow that ignoring the labels altogether for whole epochs intermittently during\ntraining can significantly improve performance in the small sample regime. More\nspecifically, we propose to train a network on two tasks jointly. The primary\nclassification task is exposed to both the unlabeled and the scarcely annotated\ndata, whereas the secondary task seeks to cluster the data without any labels.\nAs opposed to hand-crafted pretext tasks frequently used in self-supervision,\nour clustering phase utilizes the same classification network and head in an\nattempt to relax the primary task and propagate the information from the labels\nwithout overfitting them. On top of that, the self-supervised technique of\nclassifying image rotations is incorporated during the unsupervised learning\nphase to stabilize training. We demonstrate our method's efficacy in boosting\nseveral state-of-the-art SSL algorithms, significantly improving their results\nand reducing running time in various standard semi-supervised benchmarks,\nincluding 92.6% accuracy on CIFAR-10 and 96.9% on SVHN, using only 4 labels per\nclass in each task. We also notably improve the results in the extreme cases of\n1,2 and 3 labels per class, and show that features learned by our model are\nmore meaningful for separating the data.\n