SDD: Synergistic Dual-Memory Debiasing for Semi-Supervised Continual Learning

Semi-Supervised Continual Learning (SSCL) aims to enable deep learning models to learn incrementally from label-scarce streaming data without catastrophic forgetting. Existing SSCL methods face two core issues: severe overfitting to limited labeled samples, and accumulated feature representation bias across incremental tasks. To address these issues, we propose Synergistic Dual-Memory Debiasing (SDD), a semi-supervised continual learning framework that integrates Dual Memory Replay with unbiased contrastive debiasing. It includes two tightly coupled core components: a Dual Memory Replay (DMR) strategy with confidence-aware sampling, which avoids aggravating overfitting to limited labeled samples and an Unbiased Contrastive Learning (UCL) module with balanced KL regularization, which mitigates accumulated feature bias. Extensive experiments on three standard benchmarks (CIFAR-10, CIFAR-100, ImageNet-100) show that SDD achieves competitive performance compared with state-of-the-art SSCL methods, with a maximum 4.98% accuracy gain over the NNCSL baseline.