CORE: Mitigating Catastrophic Forgetting in Continual Learning through Cognitive Replay

TL;DR

Catastrophic forgetting remains a core obstacle in continual learning. CORE tackles this by mimicking human memory review through Adaptive Quantity Allocation and Quality-Focused Data Selection, using forgetting and interference signals to drive buffer management and sample curation. The method demonstrates stronger average performance and more balanced task retention across MNIST and CIFAR benchmarks, highlighting the value of cognitive-inspired replay strategies. Together, these contributions establish CORE as a robust benchmark for replay-based continual learning and offer practical improvements for long-horizon learning systems.

Abstract

This paper introduces a novel perspective to significantly mitigate catastrophic forgetting in continuous learning (CL), which emphasizes models' capacity to preserve existing knowledge and assimilate new information. Current replay-based methods treat every task and data sample equally and thus can not fully exploit the potential of the replay buffer. In response, we propose COgnitive REplay (CORE), which draws inspiration from human cognitive review processes. CORE includes two key strategies: Adaptive Quantity Allocation and Quality-Focused Data Selection. The former adaptively modulates the replay buffer allocation for each task based on its forgetting rate, while the latter guarantees the inclusion of representative data that best encapsulates the characteristics of each task within the buffer. Our approach achieves an average accuracy of 37.95% on split-CIFAR10, surpassing the best baseline method by 6.52%. Additionally, it significantly enhances the accuracy of the poorest-performing task by 6.30% compared to the top baseline. Code is available at https://github.com/sterzhang/CORE.

Figures (6)

• Figure 1: Illustration of Catastrophic Forgetting, which shows Task 1's rapid decline in accuracy when Task 2 starts to train (marked by the red arrow).
• Figure 2: Illustration of continuous task acquisition and forgetting mechanisms in lifelong learning.
• Figure 3: Pipeline of CORE. Box.a-g illustrate the process of strategically constructing the replay buffer after the introduction of current task $\mathcal{T}_\tau$ to prepare for the subsequent task $\mathcal{T}_{\tau+1}$.
• Figure 4: A dimension-reduced feature space for five classes.
• Figure 5: Comparison of data selection methods. The red triangle: the feature center of the class; Orange dots: samples randomly selected; Green dots: samples chosen to balance the sample distribution, which maintain a symmetrical relation (blue dotted line) around the feature center.