Smart memory replay: Harnessing unlabeled data for efficient class-incremental learning

Current continual learning methods can utilize labeled data to alleviate catastrophic forgetting effectively. However, obtaining labeled samples can be difficult and tedious as it may require expert knowledge. In many practical application scenarios, labeled and unlabeled samples exist simultaneously, with more unlabeled than labeled samples in streaming data. Unfortunately, existing class-incremental learning methods face limitations in effectively utilizing unlabeled data, thereby impeding their performance in incremental learning scenarios.

To solve the problems, a research team led by Qiang Wang published their new research on 15 December 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.

We delve into a more challenging scenario: Semi-Supervised Class-Incremental Learning (SSCIL), aiming to leverage unlabeled data to alleviate catastrophic forgetting in classification tasks. The SSCIL model is restricted to the current task's dataset, comprising a limited number of labeled samples and a substantial amount of unlabeled samples.

In the research, we propose a novel SSCIL framework tailored to classification tasks, based on Fixmatch. This framework facilitates the gradual acquisition of new class knowledge while maintaining a balance between the stability and plasticity on previously learned classes. Apart from it, to regularize the incremental learning process, we propose a novel strategy for measuring the temporal consistency with the unlabeled data memory replay. The extensive experiments on two benchmark datasets are concluded and experimental results showcase notable performance advantages over other competing methods.