Scholarlink

<p>Few-Shot Class-Incremental Learning (FSCIL) aims to enable deep neural networks to learn new tasks incrementally from a few samples while retaining prior knowledge. While dual-prompt tuning has shown promise in this domain, existing methods often struggle with prompt interference and suboptimal feature space structure. In this paper, we propose a synergistic framework named PDR-FSCIL, which systematically tackles these challenges through three interconnected components. First, our Prompt Decoupling Regularization (PDR) module enforces semantic orthogonality between the Overall Prompt (OP) and the Incremental Prompt (IP) using cosine similarity. This ensures the OP preserves general knowledge while the IP focuses on task-specific learning, fundamentally mitigating prompt-level interference. Second, we introduce a regularization strategy to cohesively refine the feature manifold by combining contrastive learning and center regularization. This approach employs contrastive learning to maximize inter-class separation while simultaneously using center regularization to enhance intra-class compactness. Third, a global contextual weighting mechanism is integrated into the prototype classifier to construct more robust and representative class prototypes by dynamically weighting samples. Built upon a Vision Transformer(ViT) backbone, our integrated approach was validated on five standard benchmarks (CIFAR-100, CUB-200, MiniImageNet, EuroSAT, and DTD), demonstrating its tremendous potential. Crucially, our approach addresses the significant computational challenges of applying large-scale ViT in continual learning. By freezing the vast majority of the model’s parameters and only fine-tuning lightweight prompts, our method drastically reduces the demand for computational resources and training time, making it a viable solution for real-time industrial and medical applications that rely on HPC infrastructure. This work underscores the importance of parameter-efficient strategies in deploying powerful, supercomputer-trained models in dynamic, real-world environments. Code address: <a href="https://github.com/C201918029/FSCIL/tree/master">https://github.com/C201918029/FSCIL/tree/master</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

PDR-FSCIL: prompt decoupling regularization with an optimized classifier for few-shot class-incremental learning

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Meilan Hao,
  • Yiren Cai,
  • Yizhan Gu,
  • Xin Ning,
  • Deepak Kumar Jain

摘要

Few-Shot Class-Incremental Learning (FSCIL) aims to enable deep neural networks to learn new tasks incrementally from a few samples while retaining prior knowledge. While dual-prompt tuning has shown promise in this domain, existing methods often struggle with prompt interference and suboptimal feature space structure. In this paper, we propose a synergistic framework named PDR-FSCIL, which systematically tackles these challenges through three interconnected components. First, our Prompt Decoupling Regularization (PDR) module enforces semantic orthogonality between the Overall Prompt (OP) and the Incremental Prompt (IP) using cosine similarity. This ensures the OP preserves general knowledge while the IP focuses on task-specific learning, fundamentally mitigating prompt-level interference. Second, we introduce a regularization strategy to cohesively refine the feature manifold by combining contrastive learning and center regularization. This approach employs contrastive learning to maximize inter-class separation while simultaneously using center regularization to enhance intra-class compactness. Third, a global contextual weighting mechanism is integrated into the prototype classifier to construct more robust and representative class prototypes by dynamically weighting samples. Built upon a Vision Transformer(ViT) backbone, our integrated approach was validated on five standard benchmarks (CIFAR-100, CUB-200, MiniImageNet, EuroSAT, and DTD), demonstrating its tremendous potential. Crucially, our approach addresses the significant computational challenges of applying large-scale ViT in continual learning. By freezing the vast majority of the model’s parameters and only fine-tuning lightweight prompts, our method drastically reduces the demand for computational resources and training time, making it a viable solution for real-time industrial and medical applications that rely on HPC infrastructure. This work underscores the importance of parameter-efficient strategies in deploying powerful, supercomputer-trained models in dynamic, real-world environments. Code address: https://github.com/C201918029/FSCIL/tree/master.