Semi-Supervised Learning for Auditory Event-Related Potential-Based Brain-Computer Interface
Semi-Supervised Learning for Auditory Event-Related Potential-Based Brain-Computer Interface
复制标题
基于听觉事件相关电位的脑机接口半监督学习
DOI：
10.1109/access.2021.3067337
复制
发表时间：
2021-01-01
期刊：
影响因子：
3.9
通讯作者：
Mitsukura, Yasue
中科院分区：
文献类型：
--
作者：
Ogino, Mikito;Kanoga, Suguru;Mitsukura, Yasue
脑机接口（BCI）是一种分析神经活动并传递翻译命令以执行动作的通信工具。近年来，基于视觉事件相关电位（ERP）的脑机接口和运动意象脑机接口作为一种能够适应不同受试者和试验模式差异的有效技术，受到了人们的关注。SSL技术的应用也有望提高基于听觉erp的bci的性能。然而，没有确凿的证据支持SSL技术对基于听觉erp的脑机接口的积极影响。如果这种积极效果能够得到验证，将对BCI社区有所帮助。在这项研究中，我们使用以下机器学习算法评估了SSL技术对两个公共听觉BCI数据集（amuse和pass2d）的影响：逐步线性判别分析、收缩线性判别分析、时空判别分析和最小二乘支持向量机。这些骨干分类器首先使用标记数据进行训练，然后在每次测试数据中使用未标记数据进行增量更新。尽管数据集的一些数据受到了负面影响，但在所有情况下，大多数数据都明显得到了SSL的改善。随着每增加一个未标记数据，总体精度呈对数增长。这项研究支持SSL技术的积极作用，并鼓励未来的研究人员将其应用于基于听觉erp的脑机接口。
A brain-computer interface (BCI) is a communication tool that analyzes neural activity and relays the translated commands to carry out actions. In recent years, semi-supervised learning (SSL) has attracted attention for visual event-related potential (ERP)-based BCIs and motor-imagery BCIs as an effective technique that can adapt to the variations in patterns among subjects and trials. The applications of the SSL techniques are expected to improve the performance of auditory ERP-based BCIs as well. However, there is no conclusive evidence supporting the positive effect of SSL techniques on auditory ERP-based BCIs. If the positive effect could be verified, it will be helpful for the BCI community. In this study, we assessed the effects of SSL techniques on two public auditory BCI datasets-AMUSE and PASS2D-using the following machine learning algorithms: step-wise linear discriminant analysis, shrinkage linear discriminant analysis, spatial temporal discriminant analysis, and least-squares support vector machine. These backbone classifiers were firstly trained by labeled data and incrementally updated by unlabeled data in every trial of testing data based on SSL approach. Although a few data of the datasets were negatively affected, most data were apparently improved by SSL in all cases. The overall accuracy was logarithmically increased with every additional unlabeled data. This study supports the positive effect of SSL techniques and encourages future researchers to apply them to auditory ERP-based BCIs.