Nearest-neighbor class prototype prompt and simulated logits for continual learning

Yue Lu, Jie Tan, Shizhou Zhang, Yinghui Xing, Guoqiang Liang, Yanning Zhang

Abstract: Highlights•We propose a nearest-neighbor class prototype method for aligning prompts.•Prompts are incorporated into query function finetuning to improve performance.•We design a simulated logit method to effectively address inter-task confusion.•Our method achieves SOTA performance on benchmarks and excels in long-term learning.