Prototype Compensation Method for Small-Scale Incremental Learning Based on Attention-Guided Sample Augmentation and Causal Constraints

In small-shot, replay-free incremental learning scenarios, the scarcity of training samples can lead to weak model feature representation and insufficient causal K-nearest neighbor candidate feature pools, resulting in weight estimation bias and prototype drift, thereby exacerbating catastrophic forgetting. Existing causality-constrained methods have good stability, but they lack plasticity under small-shot conditions; ordinary data augmentation only expands the training samples and lacks causal representation constraints. To address this, this paper proposes a prototype compensation method that combines attention-guided sample augmentation with causal constraints. Attention-guided data augmentation (AGDA) is used to generate semantically invariant synthetic samples, expanding the causal K-nearest neighbor candidate pool to optimize weight estimation; combined with causal regularization to protect old class knowledge, supervised contrastive loss is used to optimize the feature space, and prototype drift is corrected through adversarial shift compensation. Experiments were conducted on 4-task and 8-task small-shot settings of Caltech 101 and Oxford Flowers 102. The results show that, compared to the contrastive baseline, this method improves average accuracy by 1.2%–3.3% and average incremental accuracy by 1.0%–3.0%, effectively reducing the forgetting rate, providing a feasible solution path for replay-free incremental learning with scarce samples.