Aggressive Growing Mixup: A Faster and Better Semi-Supervised Learning Approach
Zhenlei Li, Zhaolin Hong, Haoran Zheng · 2021 IEEE 3rd International Conference on Frontiers Technology of Information and Computer (ICFTIC) · 2021
Semi-supervised learning has proven to be a powerful paradigm to avoid overfitting small labeled training sets. In this work, we propose a more robust algorithm to make better leverage unlabeled samples for generalization models compared to the latest semi-supervised learning approaches. Our method, named Aggressive Growing Mixup or AGMixup, works for mixing labeled and unlabeled samples using Mixup-the batch size of unlabeled samples is aggressively increased. We show that AGMixup achieves state-of-the-art results on standard semi-supervised learning benchmarks CIFAR-10 and SVHN. For example, AGMixup outperforms all previous approaches and achieves an error rate of 4.25% on CIFAR-10 with only 250 examples and an error rate of 2.25% on SVHN with only 4000 examples, nearly matching the performance of the fully supervised model trained with all labeled examples.