All Labels Are Not Created Equal: Enhancing Semi-supervision via Label Grouping and Co-training

Pseudo-labeling is a key component in semi-supervised learning (SSL). It\nrelies on iteratively using the model to generate artificial labels for the\nunlabeled data to train against. A common property among its various methods is\nthat they only rely on the model's prediction to make labeling decisions\nwithout considering any prior knowledge about the visual similarity among the\nclasses. In this paper, we demonstrate that this degrades the quality of\npseudo-labeling as it poorly represents visually similar classes in the pool of\npseudo-labeled data. We propose SemCo, a method which leverages label semantics\nand co-training to address this problem. We train two classifiers with two\ndifferent views of the class labels: one classifier uses the one-hot view of\nthe labels and disregards any potential similarity among the classes, while the\nother uses a distributed view of the labels and groups potentially similar\nclasses together. We then co-train the two classifiers to learn based on their\ndisagreements. We show that our method achieves state-of-the-art performance\nacross various SSL tasks including 5.6% accuracy improvement on Mini-ImageNet\ndataset with 1000 labeled examples. We also show that our method requires\nsmaller batch size and fewer training iterations to reach its best performance.\nWe make our code available at https://github.com/islam-nassar/semco.\n