What Must Replay Preserve? Separating Correctable Bias from Class Correspondence

2026-10-05

AI Summary

This study investigates the essential components preserved during experience replay in class-incremental learning, which remain poorly understood. We propose a diagnostic framework that treats buffer predictions as heterogeneous supervisory signals and introduces a novel task-level shift mechanism to disentangle correctable bias from class correspondence. Leveraging DER++ replay, buffer prediction editing, and shift correction techniques, we systematically evaluate the independent contributions of each component. Our analysis reveals that the cost associated with late-stage class matching primarily stems from correctable bias, whereas disrupting class correspondence incurs persistent and substantial performance penalties. These findings provide critical theoretical insights for optimizing replay strategies in incremental learning scenarios.

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 anonymous.4open.science/r/replay-preserve-E22B.

Problem

• class-incremental learning
• logit replay
• catastrophic forgetting
• classifier bias
• class correspondence

Innovation

• Class-incremental learning
• Logit replay
• Diagnostic framework
• Bias correction
• Class correspondence