Mutual learning of self-supervised ViTs for image classification with limited labeled data

With the rapid advancement of self-supervised learning in computer vision, pretrained models have shown remarkable transferability across a wide range of downstream tasks by effectively leveraging large-scale unlabeled data. However, prevailing approaches predominantly adopt supervised fine-tuning strategies on individual models, often overlooking the potential complementarity between different pretrained models during adaptation. This limitation constrains the further improvement of model performance. To address this, the authors propose a bidirectional mutual learning framework that enables two distinct self-supervised models to be jointly fine-tuned on a target task. By exchanging predictions during training, the models share knowledge and collaboratively enhance their representations. Extensive experiments on multiple benchmarks demonstrate that their method consistently outperforms conventional full-parameter fine-tuning and low-rank adaptation (LoRA) in classification performance. Notably, the superior results achieved on the Radiological Society of North America (RSNA) data set highlight the potential of their method in medical imaging analysis.