基于自适应多视角深度神经网络的脑电识别
ADAPTIVE MULTI-VIEW DEEP NEURAL NETWORK FOR EEG RECOGNITION
王域枫 1冯伟 2杭文龙2
作者信息
- 1. 南京科技职业学院信息工程学院 江苏南京 210048
- 2. 南京工业大学计算机与信息工程学院 江苏南京 211816
- 折叠
摘要
由于已有深度学习方法没有从脑功能分离与整合机制角度出发构建脑电(Electroencephalogram,EEG)识别网络,导致识别精度不高,因而提出一种融合多视角学习与 自适应权重学习机制的自适应多视角深度学习模型.将脑电信号划分为不同脑区的多个局部视角,将整个大脑区域视作全局视角,构建能够反映脑功能分离与整合机制的多视角深度学习框架;利用注意力机制 自适应学习多个视角之间的重要程度.该模型不仅可以学习不同脑区EEG深度特征,而且可以 自适应地学习各个脑区权重分配.在公开及自采集EEG数据集上的实验结果均验证了该方法的有效性.
Abstract
Current deep learning methods do not take the mechanism of brain function separation and integration into consideration for EEG recognition,resulting the poor recognition accuracy.In view of this,we propose an adaptive multi-view deep learning model which combines multi-view learning and adaptive weights learning mechanism.By dividing the EEG signals into multiple local perspectives according to different brain regions,and regarding the entire brain area as global perspective,a multi-view deep learning framework that can reflect the mechanism of brain function separation and integration was constructed.The attention mechanism was used to adaptively learn the weights of multiple views.The proposed learning model not only could learn deep features of EEG signals in different brain regions,but also could adaptively learn the weights of multiple views.Experimental results on public EEG dataset and self-collected EEG dataset demonstrate the effectiveness of the proposed method.
关键词
脑电信号/深度学习/多视角学习/注意力机制Key words
Electroencephalogram/Deep learning/Multi-view learning/Attention mechanism引用本文复制引用
基金项目
国家自然科学基金项目(61802177)
江苏省高校自然科学研究面上项目(18KJB520020)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据