Abstract

Continual learning in data-scarce environments requires models to adapt to novel concepts without forgetting prior knowledge, a challenge exacerbated by strict computational constraints in real-world deployments. This thesis addresses the stability-plasticity-efficiency trilemma in both Few-Shot Class-Incremental Learning (FSCIL) and Incremental Few-Shot Object Detection (iFSOD). For classification, we propose Selective Backpropagation (SBP), a structured parameter budgeting framework that explicitly partitions network capacity into frozen, assigned, and free subsets to prevent catastrophic forgetting without experience replay. SBP maintains controlled plasticity through epoch-wise parameter re-initialization and is integrated with partial knowledge distillation to mitigate representation drift. For object detection, we address the limitation of background suppression by introducing a data enrichment pipeline leveraging vision-language models to generate open-vocabulary pseudo-annotations for latent objects. This strategy establishes forward compatibility by reserving feature capacity for future classes during base training. Extensive evaluations on CIFAR-100, MiniImageNet, CUB-200-2011, MS-COCO, and DOTA demonstrate that our proposed frameworks achieve state-of-the-art accuracy across both domains, validating the effectiveness of efficient parameter isolation and semantic data enrichment for incremental learning.