MAGIC-SSCIL: Manifold Anchoring and Geometric Incremental Calibration for Semi-Supervised Class Incremental Learning

Continual learning aims to enable models to learn sequentially from new data without forgetting previous knowledge. Early approaches like Elastic Weight Consolidation (EWC) and knowledge distillation addressed catastrophic forgetting through parameter regularization and functional preservation. With the advent of large pre-trained models, methods like SLCA and prototype-based approaches shifted focus to statistical summaries, reducing memory costs. Semi-supervised techniques, such as FixMatch, introduced pseudo-labeling and consistency regularization to leverage unlabeled data. However, these methods rely on confidence thresholds, which become less effective as class numbers grow, leading to feature drift and unstable class representations. Recent works incorporate geometric and topological constraints to improve stability, but they often neglect the structure of unlabeled data, limiting plasticity. The challenge remains to develop methods that utilize data geometry for both stability and plasticity in a scalable, exemplar-free setting.

Existing SSCIL methods struggle with feature drift caused by unreliable pseudo-labels and statistical fragility under sparse supervision. Confidence thresholding discards valuable hard samples and fails to adapt as the label space expands. Class statistics estimated from limited labels are noisy, leading to inaccurate generative replay and increased forgetting. Moreover, the absence of explicit geometric constraints allows the feature space to deform, causing class relationships to distort over time. These issues hinder the ability to maintain a stable, discriminative feature space while learning new classes continuously, especially without storing past exemplars, which is critical for privacy and resource constraints.

MAGIC introduces a threshold-free pseudo-labeling mechanism via graph-based label propagation, leveraging data manifold structure to improve sample utilization. SWGC employs Bayesian shrinkage to produce stable class statistics from sparse labels, reducing estimation noise and feature drift. GSA enforces topological consistency by aligning the relational structure of features between student and teacher models, preventing semantic drift. This integrated approach allows the model to adapt to new classes while preserving the geometry of old classes, addressing both stability and plasticity without storing exemplars. The framework combines statistical calibration with geometric regularization, a novel synergy in SSCIL, especially effective in low-label, fine-grained scenarios.

• �� Construct a k-nearest neighbor graph in the feature space of the frozen backbone, representing data points and their similarities.
• �� Perform label propagation on this graph to diffuse label information from labeled and pseudo-labeled nodes, generating soft label distributions for unlabeled data.
• �� Use the reliability scores from LP to weight class mean and covariance estimates via SWGC, applying Bayesian shrinkage to stabilize statistics with a single interpretable parameter ν₀.
• �� Sample phantom features from the calibrated Gaussian distributions for replay, avoiding raw exemplar storage.
• �� During training, optimize the student model with a combined loss: classification loss on fixed prototypes, LP-based soft targets, and the GSA loss that aligns feature relations with the teacher.
• �� At task boundaries, update label propagation and statistics, ensuring the feature space topology remains consistent.
• �� Throughout, lock the backbone parameters, adapt only the plastic head, and enforce geometric alignment to prevent drift.

Experiments conducted on CIFAR-100, CUB-200, and ImageNet-R datasets, with label ratios of 1%, 5%, and 10%. Baselines include FixMatch, SLCA, SimpleCIL, and other state-of-the-art SSCIL methods. Metrics focus on average incremental accuracy and forgetting rate. Hyperparameters such as k in LP, α in propagation, and ν₀ in SWGC are tuned via validation. Ablation studies isolate the effects of LP, SWGC, and GSA. Results show MAGIC consistently outperforms baselines, especially in low-label regimes, with accuracy gains of 5-8%. Statistical robustness and topology preservation are validated through feature visualization and class relation metrics.

MAGIC achieves up to 85% accuracy on CIFAR-100 at 10% labels, surpassing SLCA by 7%. At 1% labels, accuracy remains above 70%, outperforming all baselines. Similar improvements are observed on CUB-200 and ImageNet-R, with robustness in class relation preservation. Ablation confirms each component's contribution: LP improves sample utilization, SWGC stabilizes statistics, GSA maintains topology. The method reduces catastrophic forgetting significantly, especially in fine-grained, low-label scenarios, demonstrating superior stability and plasticity balance.

Applicable in privacy-sensitive domains like medical diagnostics, autonomous systems, and edge AI, where raw data storage is restricted. The framework enables continuous learning from sparse labels and unlabeled streams, reducing annotation costs. It supports deployment in resource-constrained environments, facilitating scalable, lifelong learning in real-world applications. Future integration with multi-modal data and self-supervised signals could further enhance its versatility.

MAGIC's performance may degrade under extremely sparse labels (<1%) or severe class imbalance, due to unreliable statistical estimates. The graph construction and label propagation can be sensitive to noise, affecting pseudo-label quality. Scalability to very large datasets or numerous tasks may require optimization. The fixed backbone assumption limits adaptation to evolving feature representations, suggesting future work on dynamic backbone updates.