EEG-Based Motor Imagery Classification with Deep Multi-Task Learning
Autor: | Danli Wang, Nan Zheng, Kang Yue, Zuo-Jun Max Shen, Yaguang Song |
---|---|
Rok vydání: | 2019 |
Předmět: |
medicine.diagnostic_test
Computer science business.industry Deep learning Multi-task learning Pattern recognition 02 engineering and technology Electroencephalography Convolutional neural network 03 medical and health sciences ComputingMethodologies_PATTERNRECOGNITION 0302 clinical medicine Motor imagery 0202 electrical engineering electronic engineering information engineering medicine Domain knowledge 020201 artificial intelligence & image processing Artificial intelligence business Classifier (UML) 030217 neurology & neurosurgery |
Zdroj: | IJCNN |
Popis: | In the past decade, Electroencephalogram (EEG) has been applied in many fields, such as Motor Imagery (MI) and Emotion Recognition. Traditionally, for classification tasks based on EEG, researchers would extract features from raw signals manually which is often time consuming and requires adequate domain knowledge. Besides that, features manually extracted and selected may not generalize well due to the limitation of human. Convolutional Neural Networks (CNNs) plays an important role in the wave of deep learning and achieve amazing results in many areas. One of the most attractive features of deep learning for EEG-based tasks is the end-to-end learning. Features are learned from raw signals automatically and the feature extractor and classifier are optimized simultaneously. There are some researchers applying deep learning methods to EEG analysis and achieving promising performances. However, supervised deep learning methods often require large-scale annotated dataset, which is almost impossible to acquire in EEG-based tasks. This problem limits the further improvements of deep learning models for classification based on EEG. In this paper, we propose a novel deep learning method DMTL-BCI based on Multi-Task Learning framework for EEG-based classification tasks. The proposed model consists of three modules, the representation module, the reconstruction module and the classification module. Our model is proposed to improve the classification performance with limited EEG data. Experimental results on benchmark dataset, BCI Competition IV dataset 2a, show that our proposed method outperforms the state-of-the-art method by 3.0%, which demonstrates the effectiveness of our model. |
Databáze: | OpenAIRE |
Externí odkaz: |