基于多样性样本回放的类增量学习方法
Class-incremental learning method based on diversity sample replay
李虓 1郭辉1
作者信息
- 1. 宁夏大学信息工程学院,宁夏银川 750021
- 折叠
摘要
针对类增量学习中的灾难性遗忘问题,提出一种基于多样性样本回放的类增量学习方法.为使类增量学习能够时序、快速地输入任务数据流,在因果卷积网络模型中引入自注意力机制;通过样本分类的不确定性和数据增强抽取多样性回放样本,提升新任务渐进式学习的效果并防止模型对旧任务的遗忘;采用一阶段生成伪特征与真实特征加权对比进行剪枝去除冗余的网络参数.实验结果表明,提出算法的精度优于经典增量学习算法,可有效缓解灾难性遗忘问题并提升增量学习性能.
Abstract
To solve the catastrophic forgetting in class-incremental learning,a class-incremental learning method based on diversity sample replay was proposed.To enable class-incremental learning to rapidly input task data streams in time sequence,a self-attention mechanism was introduced into the causal convolutional network model.The diverse replay samples were extracted through the uncertainty of sample classification and data enhancement,improving the effects of progressive learning of new tasks and preventing the model from forgetting old tasks.A pruning algorithm was used to remove redundant network parameters through a one-stage generated pseudo feature weighted comparison with the real feature.Experimental results show that the accuracy of the proposed algorithm is better than that of the classical incremental learning algorithm,which can effectively alleviate the problem of catastrophic forgetting and improve the performance of incremental learning.
关键词
增量学习/灾难性遗忘/因果关系/卷积网络/网络剪枝/注意力机制/生成对抗网络/数据增强Key words
incremental learning/catastrophic forgetting/causality/convolution network/network pruning/attention mecha-nism/GAN/data augmentation引用本文复制引用
基金项目
宁夏回族自治区自然科学基金(2021AAC03117)
出版年
2024
NETL
NSTL
维普
万方数据