ISSN 1671-3710
CN 11-4766/R

心理科学进展 ›› 2023, Vol. 31 ›› Issue (2): 173-195.doi: 10.3724/SP.J.1042.2023.00173

• 研究方法 • 上一篇    下一篇


陈新文, 李鸿杰, 丁玉珑()   

  1. 华南师范大学脑认知与教育科学教育部重点实验室; 华南师范大学心理学院; 华南师范大学心理应用研究中心; 广东省心理健康与认知科学重点实验室, 广州 510631
  • 收稿日期:2021-12-16 出版日期:2023-02-15 发布日期:2022-11-10
  • 通讯作者: 丁玉珑
  • 基金资助:

Exploring the neural representation patterns in event-related EEG/MEG signals: The methods based on classification decoding and representation similarity analysis

CHEN Xinwen, LI Hongjie, DING Yulong()   

  1. Key Laboratory of Brain, Cognition and Education Sciences (South China Normal University), Ministry of Education; School of Psychology, South China Normal University; Center for Studies of Psychological Application, South China Normal University; Guangdong Key Laboratory of Mental Health and Cognitive Science, Guangzhou 510631, China
  • Received:2021-12-16 Online:2023-02-15 Published:2022-11-10
  • Contact: DING Yulong


探究不同心智活动下的神经表征差异, 是认知神经科学关注的核心问题之一。早期的脑电/脑磁分析方法主要关注组平均后的神经响应水平, 这要求在关注的时间进程上, 各个被试在相同刺激条件下事件相关电位/事件相关磁场的振幅大小和方向、以及地形图分布和极性均要有较高的一致性。近些年来, 研究者们将功能性磁共振成像研究中常用到的两种技术——机器学习中的分类算法(即基于分类的解码)和表征相似性分析——引入到了脑电/脑磁数据分析中。这两种新技术可以克服传统脑电/脑磁数据基于具体电压/磁感应强度波形平均分析的缺点, 具有在个体水平上探究神经表征编码的特点, 为人们探究大脑在不同时间进程上如何对特定的神经表征信息进行动态编码提供了新的思路。两种技术基于不同的方法学原理来抽提个体间一致的脑认知加工机制, 还为脑电/脑磁研究开展跨时域、跨任务、跨模态、跨群体比较不同认知过程中的表征差异提供了更多新颖的途径。我们首先通过与传统的脑电/脑磁分析方法进行比较, 系统性介绍了基于分类的解码和表征相似性分析的原理和操作流程, 之后对两种方法的应用场景进行了梳理, 并在最后对未来可供研究的方向提出了我们的见解。

关键词: 脑电/脑磁, 神经表征, 机器学习/基于分类的解码, 表征相似性分析


It is generally considered that the human brain will generate distinct neural representations corresponding to different mental processes. Exploring the differences of neural representations under various mental activities is one of the core issues in cognitive neuroscience. During recent decades, researchers have used different neuroimaging techniques to record brain activities involved in complex cognitive processes from the perspective of temporal or spatial measurement. Among these techniques, the non-invasive EEG/MEG with temporal resolution of millisecond has become a popular one to study the time courses of various cognitive activities. Due to the characteristics of EEG/MEG data (e.g., low S/N), in order to obtain relatively reliable results, traditional EEG/MEG studies mainly focused on the neural responses after group averaging, paying less attention to individual differences. Such method assumes that, for each subject, the amplitudes and directions of ERPs/ERMFs and their topographic maps in a specific time window of interest exhibit a consistent pattern under an experimental condition. In the case of poor consistency, the neural responses across subjects may cancel each other to a great extent after group averaging, which makes it difficult to get a reasonable interpretation.

In recent years, researchers have introduced two techniques commonly used in fMRI studies, classification algorithms in machine learning (i.e., classification-based decoding) and representation similarity analysis, into the EEG/MEG data analysis. These two new techniques can overcome the shortcomings of traditional EEG/MEG data analysis based on averaging of voltage/magnetic flux density waveforms by taking individual differences into account, which could be used to reveal the coding of neural representation at individual level and provide a new idea to explore how the brain encodes specific neural representations dynamically. In the study of ERPs/ERMFs, classification-based decoding and representation similarity analysis can be used to explore not only the neural mechanisms that show consistent patterns along time among individuals, but also those that are significantly different across individuals but keep stable for a given individual. Thus, these two techniques are able to reveal specific neural representation patterns and even identify "brain fingerprints" at individual levels. Based on different methodological theories, these two techniques provide novel ways for EEG/MEG studies to compare representational differences of cognitive processes across time windows, tasks, modalities, and groups. Firstly, we systematically introduced the principles and operational processes of classification-based decoding and representation similarity analysis, together with a comparison with those traditional analysis methods of EEG/MEG. Then, the EEG/MEG studies to date using these two techniques are reviewed. Finally, some possible future research directions with regard to these two techniques are proposed.

Key words: electroencephalography/magnetoencephalography, neural representation, machine learning/classification- based decoding, representation similarity analysis