| 引用本文: | 李昱洁,吴晗,孟丹,李天瑞,杨新.面向开放世界持续学习的任务敏感提示驱动混合专家模型.软件学报,2026,37(4):1531-1547 |
| |
|
| |
|
|
| 本文已被：浏览 2025次 下载 1266次 |
码上扫一扫！ |
|
|
| 面向开放世界持续学习的任务敏感提示驱动混合专家模型 |
|
李昱洁1,2, 吴晗3, 孟丹1,2, 李天瑞4, 杨新3
|
|
1.西南财经大学 管理科学与工程学院, 四川 成都 611130;2.智能金融教育部工程研究中心 (西南财经大学), 四川 成都 611130;3.西南财经大学 计算机与人工智能学院, 四川 成都 611130;4.西南交通大学 计算机与人工智能学院, 四川 成都 614202
|
|
| 摘要: |
| 开放世界持续学习(OWCL)旨在模拟现实环境中任务不断演化、类别动态变化且遇到未经训练的未知样本的情景. 一个良好的开放世界持续学习模型不仅需要在学习新任务的同时保持对已学任务的记忆, 还需具备识别未知类别的能力, 进而实现持续且鲁棒的知识积累与泛化. 然而, 现有持续学习方法普遍建立在封闭世界假设之上, 无法有效应对开放类别带来的类别不确定性与任务间干扰, 尤其在知识稳定性与知识可塑性之间的权衡上表现出明显不足. 因此, 在开放世界持续学习问题的形式化定义基础上, 提出一种任务敏感提示驱动的混合专家模型TP-MoE (task-aware prompt-driven mixture of experts), 以实现对任务语义的动态建模与专家模块的高效调度, 从而帮助模型进行知识传输和知识更新. 具体而言, TP-MoE引入一种即插即用的任务提示聚合机制并改进门控机制用以专家网络路由, 在任务增量过程中持续融合历史与当前任务知识; 同时结合一种自适应开放边界阈值策略, 可根据新旧知识的迁移动态调整开放类别的判别边界, 从而提升开放类别检测能力与已知类别分类准确性. 实验结果表明, TP-MoE在Split-CIFAR100和Open-CORe25基准数据集上对各类指标的测试均取得领先性能, 展现出良好的稳健性与泛化性, 为开放世界持续学习任务中的知识建模与任务调度提供了一种可扩展、可迁移的新框架. |
| 关键词: 开放世界持续学习 持续学习 任务敏感 混合专家模型 知识传输 |
| DOI：10.13328/j.cnki.jos.007525 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金(62476228) |
|
| Task-aware Prompt-driven Mixture-of-experts Model for Open-world Continual Learning |
|
LI Yu-Jie1,2, WU Han3, MENG Dan1,2, LI Tian-Rui4, YANG Xin3
|
|
1.School of Management Science and Engineering, Southwestern University of Finance and Economics, Chengdu 611130, China;2.Engineering Research Center for Intelligent Finance (Southwestern University of Finance and Economics), Ministry of Education, Chengdu 611130, China;3.School of Computing and Artificial Intelligence, Southwestern University of Finance and Economics, Chengdu 611130, China;4.School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 614202, China
|
| Abstract: |
| Open-world continual learning (OWCL) aims to simulate real-world scenarios where tasks evolve continuously, categories change dynamically, and unseen samples are encountered. A well-designed OWCL model is expected not only to retain knowledge of learned tasks while acquiring new tasks but also to recognize unknown categories, thus achieving continuous and robust knowledge accumulation and generalization. However, most existing continual learning methods are built upon the closed-world assumption and cannot effectively cope with the category uncertainty and inter-task interference introduced by open categories. In particular, they show clear limitations in balancing knowledge stability and plasticity. Therefore, based on the formal definition of the OWCL problem, this study proposes a task-aware prompt-driven mixture-of-experts model (TP-MoE), which realizes dynamic modeling of task semantics and efficient scheduling of expert modules, thus supporting knowledge transfer and knowledge update. Specifically, TP-MoE introduces a plug-and-play task prompt aggregation mechanism and improves the gating strategy for expert routing, enabling the continual integration of historical and current task knowledge during task increments. At the same time, an adaptive open-boundary thresholding strategy is incorporated, which dynamically adjusts the decision boundaries of open categories according to the transfer between new and old knowledge, thus enhancing both open-category detection capability and known-category classification accuracy. Experimental results demonstrate that TP-MoE achieves state-of-the-art performance across various metrics on the Split-CIFAR100 and Open-CORe25 benchmarks, exhibiting strong robustness and generalization. This study provides a scalable and transferable framework for knowledge modeling and task scheduling in open-world continual learning. |
| Key words: open-world continual learning (OWCL) continual learning task-aware mixture-of-experts (MoE) model knowledge transfer |
|
|
|
|
