视觉CIL方法能否迁移到蛋白质？冻结蛋白质编码器的诊断基准
原标题：Do Vision CIL Methods Transfer to Protein? A Diagnostic Benchmark of Frozen Protein Encoders
AI 摘要
该研究对九种代表性类增量学习方法在五种冻结蛋白质编码器和两个酶功能数据集上进行了受控基准测试，发现没有任何方法能可靠地跨编码器迁移。诊断分析表明，不稳定性源于仅训练分类头时基于梯度的稳定性信号微弱、嵌入几何差异大，以及协方差容量需求随数据集规模和表示几何变化。作者认为表示本身应被视为持续学习设计的核心组件，而非被动的特征提取器。
正文节选
Continual Learning in the Era of Foundation Models and Embodied Agents Do Vision CIL Methods Transfer to Protein? A Diagnostic Benchmark of Frozen Protein Encoders Abstract Protein function annotation operates over an expanding label space, making class-incremental learning (CIL) a natural framework for incorporating newly discovered functions without full retraining. Yet existing CIL methods have been developed largely in vision and have not been systematically evaluated over frozen protein rep