Incremental Self-training for Semi-supervised Learning

• DOI: 10.48550/arxiv.2404.12398
• Publication Date: 2024-04-14
• Authors: Jifeng Guo; Zhulin Liu; Tong Zhang; C. L. Philip Chen
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Machine Learning (cs.LG)

Semi-supervised learning provides a solution to reduce the dependency of machine on labeled data. As one efficient semi-supervised techniques, self-training (ST) has received increasing attention. Several advancements have emerged address challenges associated with noisy pseudo-labels. Previous works acknowledge importance unlabeled data but not delved into their utilization, nor they paid attention problem high time consumption caused by iterative learning. This paper proposes Incremental Self-training (IST) for fill these gaps. Unlike ST, which processes all indiscriminately, IST in batches and priority assigns pseudo-labels samples certainty. Then, it around decision boundary after model is stabilized, enhancing classifier performance. Our simple yet effective fits existing self-training-based methods. We verify proposed five datasets two types backbone, effectively improving recognition accuracy speed. Significantly, outperforms state-of-the-art competitors three challenging image classification tasks.