摘要
使用脑电进行情绪识别已经有了广泛的研究,但由于脑电的低信噪比、不平稳性以及受试者情绪表达方式的不同,不同受试者甚至单个受试者的脑电图情绪特征都会存在差异性,导致脑电样本在特征空间分布不均匀,容易出现模型泛化性能差的问题。为解决这一问题,该文提出了一种结合提升算法(boost)和梯度下降法(gradient descent)的双策略训练方法交替更新脑电情绪识别模型,梯度下降法在模型推理过程中更新网络参数,使损失最小化,提升算法用于更新脑电样本权重。在DEAP数据集上的实验结果表明,该方法在效价、唤醒和优势度3个维度上的准确率分别为71.25%、71.48%和71.80%,且在跨被试数据集下通过数据驱动的方式有效调整了脑电样本特征的分布,使其分布更均匀,从而提高了情绪识别模型的泛化性能。
There has been extensive research on emotion recognition using EEG,but due to the low signal-to-noise ratio of EEG,the non-stationarity of EEG and the different ways of emotional expressions among subjects,the emotional features of EEG collected from different subjects and even single subject may have variability.Thereby,the feature distribution of EEG samples is non-uniform and prone to poor model generalization performance.To solve this problem,a bi-strategy training method that combines the boost algorithm and gradient descent is proposed in this paper to alternately update the EEG-based emotion recognition model.Gradient descent method updates the network parameters to minimize the loss during model inference,and boosting algorithm is used to update the EEG sample weights.The experimental results on the DEAP dataset showed that the method in this paper achieved accuracies of 71.25%,71.48%and 71.80%on the valence,arousal,and dominance dimensions respectively.And the distribution of EEG sample features was effectively adjusted by a data-driven approach under the cross-subject dataset to make them more evenly distributed,thus improving the generalization performance of the emotion recognition model.
作者
贾巧妹
胡景钊
郑佳宾
王晨
张丽丽
赵晨宇
吴东亚
冯筠
JIA Qiaomei;HU Jingzhao;ZHENG Jiabin;WANG Chen;ZHANG Lili;ZHAO Chenyu;WU Dongya;FENG Jun(School of Information Science and Technology, Northwest University, Xi′an 710127, China;State-Province Joint Engineering and Research Center of Advanced Networking andIntelligent Information Services, Xi′an 710127, China)
出处
《西北大学学报（自然科学版）》
CAS
CSCD
北大核心
2022年第4期552-559,共8页
Journal of Northwest University（Natural Science Edition）
基金
国家重点研发计划项目(2020YFC1523300)。
关键词
情绪识别
脑电
双策略训练
非均匀
特征分布
emotion recognition
EEG
bi-strategy training
non-uniform
feature distribution
