基于迁移学习的脑机接口特征提取和预测方法研究

项目名称： 基于迁移学习的脑机接口特征提取和预测方法研究

项目编号： No.61304140

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 吕俊

作者单位： 广东工业大学

项目金额： 23万元

中文摘要： 脑机接口技术是当前国际上神经工程、信号处理等领域的前沿研究热点。脑信号的特征提取和预测是脑机接口技术的关键,然而小样本,即训练样本少，是脑机接口应用中常见的问题,容易导致特征提取和预测模型的过拟合,目前尚没有得到很好地解决。近年来兴起的迁移学习方法是解决脑机接口中小样本问题的可能的有效途径,但是在理论和实践方面都还存在许多问题。因此, 我们拟研究基于迁移学习的脑机接口特征提取和预测方法, 主要内容如下: (1) 针对脑信号分布不稳定的特点, 设计基于迁移风险度量的迁移学习目标函数;(2) 依据该目标函数, 提出同时基于样本和特征的可迁移信息抽取方法; (3) 针对脑信号噪声大的特点, 研究抑制噪声的机制, 以增强迁移学习算法在小样本情况下的鲁棒性; (4) 对迁移学习进行理论分析,研究迁移学习可改进的条件、泛化错误的上界等。

中文关键词： 脑机接口；特征提取；预测；迁移学习；

英文摘要： Brain-computer interface (BCI) technology is a hot spot of the present international frontier research in the fields of neural engineering, signal processing,etc.. The key of BCI technology is the feature extraction and prediction of brain signals. However, small sample size (SSS), namly few training sample, is a common problem for BCI applications which could easily lead to the overfitting of feature extraction and prediction model, and has not been well solved. Transfer learning (TL) method developed in recent years is a possible efficient way to solve the SSS problem of BCI, but many issues still exist in both aspects of theory and practice. Therefore, we plan to study the feature extraction and prediction method of BCI based on TL, the main contents are listed as follows: (1) According to the instability of the distributions of brain signals, we design the objective function of TL based on the measure of transfer risk. (2) On the basis of the objective function, we propose the method to extract transferable information simultaneously based on samples and features. (3) According to the big noise in brain signals, we study the noise suppressing mechanism to enhance the robustness of TL algorithm in the case of SSS. (4) We perform the theoretical analysis on TL， investagate the improvable conditions, the upper

英文关键词： brain-computer interface；feature extraction；prediction；transfer learning；