本文为美国内布拉斯加大学（作者：Jacob M. Williams）的硕士论文，共82页。

尽管在其他的空间和时间序列数据中取得了最先进的分类精度，深度学习很少用于脑电图（EEG）信号的分类。相反，大多数研究继续使用手动特征提取和传统分类器，这在很大程度上是由于每个实验的样本数较少、数据的高维性，以及难以找到合适的深度学习结构来分类脑电信号。在这篇论文中，我们比较了几种深度学习架构与传统的视觉诱发脑电信号分类技术。我们发现，使用长-短期记忆单元（LSTM）的深度学习结构优于传统方法，而小型卷积结构的性能与传统方法相当。我们还探讨了通过跨多个主题的训练和特定主题的改进来使用迁移学习，这种形式的迁移学习进一步提高了深度学习模型的分类精度。

Deep learning is seldom used in the classification ofelectroencephalography(EEG) signals, despite achieving state of the artclassification accuracies in other spatial and time series data.Instead, mostresearch has continued to use manual feature extraction followed by atraditional classifier, such as SVMs or logistic regression.This is largely dueto the low number of samples per experiment, high-dimensional nature of thedata, and the difficulty in finding appropriate deep learning architectures forclassification of EEG signals. In this thesis, several deep learningarchitectures are compared to traditional techniques for the classification ofvisually evoked EEG signals. We found that deep learning architectures usinglong short-term memory units (LSTMs) outperform traditional methods, whilesmall convolutional architectures performed comparably to traditional methods.We also explored the use of transfer learning by training across multiplesubjects and refining on a particular subject. This form of transfer learningfurther improved the classification accuracy of the deep learning models.

1.     引言
   

2.     相关工作与项目背景
   

3.     问题定义与研究方法
   

4.     数据集、具体实现与结果
   

5.     总结与展望
   

下载地址：

https://url92.ctfile.com/f/1850492-520379399-537bf9

（访问密码：3660）

更多精彩文章请关注公众号：