Self-Supervised Contrastive Regularization Algorithms For Robust Few Shot Classification

Abstract

With the limited amount of labeled data, few-shot classification is still a major problem in machine learning. In this paper, a novel Self-Supervised Contrastive Regularization (SSCR) framework is proposed, which combines self-supervised pretext tasks with contrastive regularization to boost the feature representation and generalization ability to unseen classes. The framework was tested on the miniImageNet dataset with 5-way, 1-shot, and 5-shot settings. To learn structural and semantic information, the model was pre-trained with the self-supervised tasks and then fine-tuned with the few-shot support set. Embeddings were then further refined by contrastive regularization, which reduces intra-class distances and increases inter-class distances. Experimental results show that SSCR achieves the best performance compared to baseline approaches such as prototypical networks and matching networks with 72.3% top-1 accuracy in the 1-shot setting and 83.5% in the 5-shot setting. Additionally, the framework demonstrated strong performance in the cross-domain evaluations using tiered ImageNet, with 1-shot accuracy of 69.4% and 5-shot accuracy of 82.3%, demonstrating its ability to handle shifts between different domains. t-SNE visualization of embeddings shows improved embedding of class-specific features. In summary, the proposed framework offers a scalable and robust solution to few-shot learning problems without an abundance of labeled data and can be applied in remote sensing, medical imaging, and industrial inspection, among others.