arXiv:2608.07586 · v2
MAGIC-SSCIL: Manifold Anchoring and Geometric Incremental Calibration for Semi-Supervised Class Incremental Learning
MAGIC-SSCIL:面向半监督类增量学习的流形锚定与几何增量校准
摘要
Exemplar-free semi-supervised class-incremental learning (SSCIL) asks a model to learn new classes from a few labeled and many unlabeled images without storing past data. A common recipe replays old classes from stored Gaussian statistics and pseudo-labels unlabeled data with a confidence threshold. With one or two labels per class, however, the model's confidence is a poor guide to which pseudo-labels are right, so the statistics that replay depends on rest on few or noisy samples. We propose MAGIC (Manifold Anchoring and Geometric Incremental Calibration), which propagates labels over a k-nearest-neighbor graph of frozen features and uses the propagated confidences as weights in the unlabeled loss and, through Soft-Weighted Geometry Calibration (SWGC), in the class statistics from which phantom features replay old classes. With a frozen ResNet-18 on CIFAR-100, CUB-200, and ImageNet-R at 1%, 5%, and 10% labels, we compare MAGIC with nine supervised class-incremental methods equipped with FixMatch and two native SSCIL methods. MAGIC gains from unlabeled data in all nine settings, most at 1%, and leads on fine-grained CUB-200 at every ratio; at 1% it is level with the best on ImageNet-R and 2.2 points behind the native SSCIL method TACLE on CIFAR-100. Experiments inside MAGIC show that most of this gain flows through the weighted statistics, not the loss, that at 1% labels graph pseudo-labels are more accurate than the model's own at equal coverage, and that a fixed threshold of 0.95 removes most of the gain. At higher ratios, baselines that adapt the backbone or model full covariances overtake MAGIC on CIFAR-100 and ImageNet-R.
作者备注:v2: substantially revised. The method is simplified (Geometric Structural Alignment, manifold mixup, warm-up and variance shrinkage are removed), all MAGIC experiments are new (five seeds), the TACLE baseline numbers are corrected, and new analyses of pseudo-label accuracy, class statistics and ablations are added. 52 pages