EVENT DOSSIER
To address the issues of misalignment between pre-trained models in general continuous learning and downstream tasks, as well as unreliable alignment of data flow output, researchers proposed the MePo++ unified training framework. This framework utilizes two complementary components: MetaPrep and StreamAlign. MetaPrep employs unsupervised meta-refinement of pseudo-continuous sequences to enhance representation plasticity, while StreamAlign coordinates the evolution of online features with stable pre-trained geometries to improve representation stability. Experiments verified the continuous effectiveness and versatility of MePo++ across various pre-trained models, datasets, and continuous learning baselines; the related code has been open-sourced.
通用持续学习（GCL）旨在在不依赖任务标识、明确边界或历史数据的情况下从演变数据流中学习。针对现有预训练模型方法忽视的预训练与下游适应错位及模糊数据流下输出对齐不可靠问题，研究者提出了 MePo++ 统一后训练框架。该框架通过表示精炼与协调两个互补组件 MetaPrep 和 StreamAlign，利用无监督元精炼伪持续序列提升表示可塑性，并通过协调演变在线特征与稳定预训练几何增强表示稳定性。实验在多样预训练模型、数据集及持续学习基线中验证了 MePo++ 的持续有效性与通用性，其代码已开源。