representation-drift

该论文提出 SAGA 方法，将任务边界处训练的稀疏自编码器转换为对表示漂移的低秩局部约束，以缓解持续学习中的灾难性遗忘，并在三个 ViT-B/16 基准上取得最佳的 Class-IL 准确率与最低遗忘率。

This NYU/Xi'an Jiaotong paper shows that early representational drift forms a low-dimensional 'scaffold' subspace that networks preferentially reuse when learning later tasks, indicating that early experience leaves a persistent geometric imprint on neural network adaptation.