Abstract:

Task-free online continual learning is a task-agnostic and autonomous machine learning approach, where model dynamic update is achieved through continuous adaptation to new data and forgetting suppression. In the existing online continual learning methods, model accuracy is prioritized typically at the expense of computational efficiency, making it difficult for the model to respond promptly to changes in the data stream due to the lag in training speed in high-speed data stream scenarios. To address the above challenge, an efficiency-performance co-optimized online sparse continual learning framework was proposed to overcome the limitations of conventional approaches through constructing a bidirectional sparse adaptive regulation mechanism. Firstly, a dynamic sparse topology optimization framework for parameter importance measurement was designed, so that unstructured parameter pruning was achieved by incorporating parameter sensitivity analysis. Secondly, a memory-efficiency dual-objective optimization model was established, in which computational budget allocation was adjusted dynamically based on online class distribution estimation, so as to realize optimal computational resource configuration. Finally, a gradient decoupling optimization strategy was developed to employ gradient masking to enable bidirectional optimization of both old and new knowledge, thereby accelerating model updates and preserving the integrity of the knowledge topology at the same time. The benchmark tests results show that the proposed framework has significant advantages. Compared to the baseline ER (Experience Replay), with a memory buffer of 100, on CIFAR-10 dataset the framework achieves average improvements of 4.86% in Average Online Accuracy (AOA) and 6.25% in Test Accuracy (TA) ; on CIFAR-100 dataset, the framework obtains enhancements of 13.77% in AOA and 3.08% in TA; on Mini-ImageNet dataset, it shows performance gains of 17.83% and 25.00% in AOA and TA respectively. Visualization analysis shows that the proposed framework captures underlying concept drift patterns in data streams successfully while maintaining real-time response ability. It can be seen that the proposed framework breaks through the traditional methods’ dilemma of trade-off between computational efficiency and model performance, and establishes a new paradigm for online continual learning systems in open environments.

Key words: Continual Learning (CL), sparsity, online learning, parameter masking, computational efficiency

摘要：

无边界在线连续学习是一种任务无关的自主高效机器学习方法，通过连续适应新数据和抑制遗忘实现模型的动态更新。现有在线连续学习方法通常追求模型准确率而牺牲计算效率，这使得在高速数据流场景中，模型因训练速度滞后，难以及时响应数据流的变化。针对以上问题，提出一种面向效率?性能协同优化的在线稀疏连续学习框架，通过构建双向稀疏自适应调控机制突破传统方法的瓶颈。首先，设计参数重要性度量的动态稀疏拓扑优化框架，融合参数敏感性分析，实现非结构化参数剪枝；其次，建立记忆?效率双目标优化模型，基于在线类别分布估计动态调节计算预算分配，实现计算资源的最优配置；最后，构建梯度解耦优化策略，采用梯度掩码方法实现新旧知识的双向优化，在加速模型更新的同时保持知识拓扑的完整性。基准测试结果表明，所提框架展现出显著优势，当内存缓冲区大小为100时，相较于基线ER（Experience Replay），在CIFAR-10数据集上，平均在线准确率（AOA）和测试准确率（TA）分别提升了4.86%和6.25%；在CIFAR-100数据集上，AOA和TA分别提升了13.77%和3.08%；在Mini-ImageNet数据集上，AOA和TA的提升幅度达17.83%和25.00%。可视化分析表明，所提框架在保持实时响应能力的同时，成功捕获了数据流中的潜在概念漂移模式。可见，所提框架突破了传统方法在计算效率与模型性能间的权衡困境，为开放环境下的在线连续学习系统建立了新范式。

关键词: 连续学习, 稀疏化, 在线学习, 参数掩码, 计算效率