MP-Prompt:multi-granularity prompt for continual learning
Renzheng Li, Bo Song · IET conference proceedings. · 2025
Continual learning aims to enable model to learn new task sequentially without forgetting previously learned task knowledge. In this paper, we propose a novel method called multi-granularity prompt for continual learning (MP-Prompt) with vision transformer. MP-Prompt leverages the global information of the task sequence, task-specific knowledge, and instance-level details to generate multi-granularity prompt. This prompt effectively guide the pre-trained model to learn new tasks while preserving the information acquired from previous tasks. However, existing approaches that focus mainly on local information and fail to fully exploit knowledge at different levels of the task sequence. Therefore, MP-Prompt provide a more comprehensive framework, effectively improving the model ability to handle sequential task. Experiments on benchmark datasets Cifar 100 and ImageNet-R demonstrate that MP-Prompt outperforms current state-of-the-art methods. The results show that MP-Prompt effectively learns new tasks without relying on a rehearsal buffer, significantly mitigating catastrophic forgetting. This approach enhances both the model’s plasticity and stability.