脑电时空特征的深度学习方法研究

脑电信号处理在基于脑电的脑科学、临床和脑机接口研究中具有广泛应用。本项目旨在已有研究工作基础上，系统深入地研究单被试和多被试脑电时空特征的深度学习方法。本项目针对脑电信号的时空相关性、非平稳性、高维度等特点，设计深度神经网络和学习算法对其进行非线性时空处理，提取有效特征。具体展开以下研究：1）构建时空受限玻尔兹曼机(ST-RBM)作为深度置信网络(DBN)的结构单元；2）基于逐层初始化思想开发学习算法对DBN参数进行估计；3）使用正则化和非参数贝叶斯推断确定深度网络规模；4）构建时空深度玻尔兹曼机(ST-DBM)网络融合多被试的脑电信号；5）基于变分贝叶斯和马尔科夫链-蒙特卡洛方法开发学习算法对ST-DBM参数进行估计。本项目的研究将丰富脑电信号分析和解码的理论，为相关应用提供算法支持。

Electroencephalogram (EEG) signal processing has broad applications in the EEG-based neuroscience, clinical, as well as brain-computer interface research. On the basis of our previous research work, this project is aimed at studying methods for the deep learning of single- and multi-subject EEG spatio-temporal features in an in-depth and systematic manner. To extract effective EEG features, the project designs deep neural networks and learning algorithms for the nonlinear spatio-temporal processing of EEG signals, according to their characteristics such as spatio-temporal correlation, non-stationarity, and high dimensionality. Specifically, the following lines of research will be conducted: 1) Constructing spatio-temporal restricted Boltzmann machines (ST-RBM) as building blocks of deep belief networks (DBN); 2) Developing learning algorithms to estimate the parameters of DBN based on layer-by-layer pre-training; 3) Employing regularization and non-parametric Bayesian inference to determine the size of DBN; 4) Building spatio-temporal deep Boltzman machine (ST-DBM) for the information fusion of EEG signals from multiple subjects; 5) Developing learning algorithms to estimate the parameters of ST-DBM based on variational Bayesian and Markov chain Monte Carlo (MCMC) approaches. The research of this project is expected to enrich the theories of EEG signal analysis and decoding, and afford algorithmic tools for relevant applications