一种基于预训练模型的类增量学习近似重放方法分析
Analysis of a Class Incremental Learning Approximate Replay Method Based on Pre-trained Model
尹为民1
作者信息
摘要
阐述通过对预训练模型表征的近似采样,快速修正上一轮增量学习训练中灾难性遗忘导致的参数偏差.通过实验数据的验证和评估,与传统方法相比,该方法在性能表现与训练速度上具有较大优势.
Abstract
This paper expounds that by approximating the sampling of pre-trained model representations,which rapidly corrects the parameter deviations caused by catastrophic forgetting during the previous round of incremental learning.Experimental validation and evaluation demonstrate that,compared to traditional methods,this approach offers significant advantages in both performance and training speed.
关键词
预训练模型/增量学习/灾难性遗忘Key words
pre-trained model/incremental learning/catastrophic forgetting引用本文复制引用
出版年
2024
NETL
NSTL
万方数据