arXiv：cs.LG· BoRen Deng, Xiangyue Ma, Chenglong Li, Xiaoting Du·· 4 小时前AI 评分37
回放必须保留什么？区分可校正偏差与类别对应关系
What Must Replay Preserve? Separating Correctable Bias from Class Correspondence
AI 导读
该研究提出一个诊断框架，把缓存预测视为时间异质监督，区分样本存储时已知的类别与之后学到的类别。在 CIFAR-100 上配合 DER++，用固定常数替换后学类别的未刷新存储分数，误差在 1 个百分点等价范围内，且任务级偏移把删除这些匹配的代价从 14.9 降到 1.8 个百分点；而重排存储时已知类别的非目标分数在偏移前后分别代价 4.3 和 4.0 个百分点，图像蒸馏中也存在类似代价。
正文
Abstract:Class-incremental learning must recognize all classes seen so far without task labels. Logit replay methods such as DER and DER++ mitigate forgetting by matching the model's past predictions on stored examples. Deleting this matching reveals its benefit, but the resulting accuracy cost cannot show whether the stored scores themselves are needed, or whether the cost survives correction of the classifier's bias toward recent classes. We propose a diagnostic framework that treats a cached prediction as temporally heterogeneous supervision: it separates classes known when an example was stored from classes learned afterward, edits each group, and evaluates every model before and after a task-level offset that leaves within-task predictions unchanged. On CIFAR-100 with DER++, suitable fixed constants replace the unrefreshed stored scores of later-learned classes within an equivalence margin of 1 percentage point, and the offset reduces the cost of deleting their matching from 14.9 to 1.8 points. Reassigning the non-gold scores of classes known at storage, which preserves their values and each task's target probability, costs 4.3 points before and 4.0 after the offset, and a parallel cost persists in image distillation. In the tested fixed-head setting, the large cost of deleting later-class matching is thus mostly correctable by this offset, whereas the smaller cost of disrupting class correspondence persists. Code and data are available at this http URL.
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2610.07077 [cs.LG] |
| (or arXiv:2610.07077v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.07077 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Xiaoting Du [view email]
[v1]
Mon, 5 Oct 2026 10:03:55 UTC (3,302 KB)
来源：arXiv：cs.LG · arxiv.org