📝 SummarXiv

Abstract

Class-Continual Learning (CCL) enables models to continuously learn new class knowledge while retaining previous classes, facilitating adaptation and evolution in dynamic, real-world environments. Traditional CCL methods primarily rely on visual features, which limits their effectiveness in complex, multimodal scenarios. In contrast, Vision-Language Models (VLMs) show promising potential for enhancing CCL by leveraging pre-trained knowledge and fusing multi-modal semantic cues such as text and vision. However, existing approaches struggle to mitigate catastrophic forgetting while preserving the generalization strengths of VLMs across diverse modalities. To address these challenges, we propose CalFuse, a framework for feature Calibration enhanced parameter Fusion, which enhances dynamic knowledge fusion. CalFuse introduces a dynamic feature calibration mechanism that iteratively adjusts the contribution of original visual features to the final class decision, thereby preserving the model's intrinsic generalization capability across modalities. Simultaneously, a parameter fusion strategy effectively fuses newly acquired knowledge with prior task parameters, maintaining a balance between acquiring new class representations and preserving old knowledge. Experimental results on popular benchmarks (e.g., CIFAR100 and ImageNet100) validate the superiority of the proposed method.