Scholarlink

With the growing adoption of AI models and the increasing scale and complexity of inference tasks, continual learning (CL) has gained prominence for edge devices. However, traditional CL methods rely on large replay buffers to mitigate forgetting, which conflicts with the limited resources of edge systems and struggles with task ambiguity and sudden data shifts. To address these issues, we proposes Dynamic Sparsity-driven Transition Feature Replay Continual Learning (STCL), a lightweight CL framework tailored for edge devices. STCL tackles three key challenges: (i) limited on-chip memory for CNN parameters, (ii) insufficient compute power for multi-epoch training, and (iii) the need for real-time adaptation to new environments. It achieves this through weight sparsity, transition feature activation, and adaptive dual-memory replay. Extensive experiments on large-scale CL benchmarks show that STCL achieves 95.6% adaptive accuracy on CORe50, outperforming state-of-the-art methods while significantly reducing memory usage.

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

STCL-Dynamic Sparsity-Driven Transition Feature Replay Continual Learning for Edge Devices

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Peng Zhang,
  • Jing Yang,
  • Xiaoli Ruan,
  • Qing Hou,
  • Xianghong Tang,
  • Jianhong Cheng

摘要

With the growing adoption of AI models and the increasing scale and complexity of inference tasks, continual learning (CL) has gained prominence for edge devices. However, traditional CL methods rely on large replay buffers to mitigate forgetting, which conflicts with the limited resources of edge systems and struggles with task ambiguity and sudden data shifts. To address these issues, we proposes Dynamic Sparsity-driven Transition Feature Replay Continual Learning (STCL), a lightweight CL framework tailored for edge devices. STCL tackles three key challenges: (i) limited on-chip memory for CNN parameters, (ii) insufficient compute power for multi-epoch training, and (iii) the need for real-time adaptation to new environments. It achieves this through weight sparsity, transition feature activation, and adaptive dual-memory replay. Extensive experiments on large-scale CL benchmarks show that STCL achieves 95.6% adaptive accuracy on CORe50, outperforming state-of-the-art methods while significantly reducing memory usage.