Semi-supervised Learning by Coaching

Hieu Pham, Quoc V. Le

Abstract:

Recent semi-supervised learning (SSL) methods often have a teacher to train a student in order to propagate labels from labeled data to unlabeled data. We argue that a weakness of these methods is that the teacher does not learn from the student’s mistakes during the course of student’s learning. To address this weakness, we introduce Coaching, a framework where a teacher generates pseudo labels for unlabeled data, from which a student will learn and the student’s performance on labeled data will be used as reward to train the teacher using policy gradient.

Our experiments show that Coaching significantly improves over state-of-the-art SSL baselines. For instance, on CIFAR-10, with only 4,000 labeled examples, a WideResNet-28-2 trained by Coaching achieves 96.11% accuracy, which is better than 94.9% achieved by the same architecture trained with 45,000 labeled. On ImageNet with 10% labeled examples, Coaching trains a ResNet-50 to 72.94% top-1 accuracy, comfortably outperforming the existing state-of-the-art by more than 4%. Coaching also scales successfully to the high data regime with full ImageNet. Specifically, with additional 9 million unlabeled images from OpenImages, Coaching trains a ResNet-50 to 82.34% top-1 accuracy, setting a new state-of-the-art for the architecture on ImageNet without using extra labeled data.

Keywords: semi-supervised, teacher, student, label propagation, image classification