.png){kind=icon}
.png){kind=icon}
Return
PSM: Prompt specialization module for prompt-based continual learning
.png){kind=icon}
PRE
AI
DOI：10.1016/j.cviu.2026.104798.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• Targeted Resolution of Core Bottlenecks in Prompt-based Continual Learning. We propose class-level prompt specialization to address weak cross-task discrimination in prompt-based continual learning, alleviating classification space conflicts across tasks. • Modular Design for Efficient Plug-and-Play Adaptability.We design a plug-and-play Prompt Specialization Module (PSM) that can be flexibly integrated into various prompt-based continual learning frameworks without reconstruction. • Dual-Dimensional Enhancement of Anti-Forgetting Capability. PSM enhances anti-forgetting ability via dual mechanisms: discriminative feature extraction and class-specific prototype generation, ensuring both accuracy and global consistency. • Verified Universality and Superiority on Benchmark Datasets. Extensive experiments show PSM consistently improves baseline accuracy by 0.42%–2.53% with clearer decision boundaries and stronger generalization on class-incremental benchmarks.
Keywords:
Prompt-based continual learning
Class-level prompt specialization
Anti-forgetting capability
Plug-and-play adaptability
Modular design
Journal
IF:
3.5
Papers:
441
Citations:
7.3K