Jiahao Chen Zhiyuan Huang Yurou Liu Bing Su

Abstract

Long-tailed learning has garnered increasing attention due to its wideapplicability in real-world scenarios. Among existing approaches, Long-TailedSemi-Supervised Learning (LTSSL) has emerged as an effective solution byincorporating a large amount of unlabeled data into the imbalanced labeleddataset. However, most prior LTSSL methods are designed to train models fromscratch, which often leads to issues such as overconfidence and low-qualitypseudo-labels. To address these challenges, we extend LTSSL into the foundationmodel fine-tuning paradigm and propose a novel framework: LoFT (Long-tailedsemi-supervised learning via parameter-efficient Fine-Tuning). We demonstratethat fine-tuned foundation models can generate more reliable pseudolabels,thereby benefiting imbalanced learning. Furthermore, we explore a morepractical setting by investigating semi-supervised learning under open-worldconditions, where the unlabeled data may include out-of-distribution (OOD)samples. To handle this problem, we propose LoFT-OW (LoFT under Open-Worldscenarios) to improve the discriminative ability. Experimental results onmultiple benchmarks demonstrate that our method achieves superior performancecompared to previous approaches, even when utilizing only 1\% of the unlabeleddata compared with previous works.