Convolutional Neural Networks Based On Large-Margin Softmax Loss Function With Their Application To EEG Signal Classification

Electroencephalography(EEG)signal classification is a key part of Brain-Computer Interface(BCI)system,the accuracy of which directly determines the performance of BCI system.EEG signal is easy to be disturbed by other signals,so it is difficult to improve the classification performance with poor generalization ability.Therefore,it is a challenge to study how to pre-process EEG signals and how to learn effective and robust feature representations.Recently,deep neural networks have become the state-of-the-art technique for computer vision,speech recognition and natural language processing.Applying the deep learning method to classify EEG signal has become a hot research topic.In this thesis we design lightweight 1D and 2D Convolutional Neural Networks(CNNs),Recurrent Convolutional Neural Networks(RCNNs)and Max-out RCNNs(MRCNNs)with large-margin softmax loss functions for EEG signal classification.First,we extract the frequency domain features from the EEG signals.Then,the extracted frequency feature vectors and matrices are used to train the 1D and 2D CNNs.Our approach approximately achieves93.57% accuracy for 1D CNNs,91.97% for 2D CNNs with shorter training time,93.92%accuracy for 1D RCNNs and 91.97% accuracy for 2D RCNNs,94.06% accuracy for 1D MRCNNs and 92.97% accuracy for 2D RCNNson the 4 classes of cognitive load recognition task.The results indicate that the proposed method with few parameters gets good performance and is more efficient than other deep learning methods.We find that the largemargin softmax loss function works well for EEG signal recognition,and we also find that the 1D CNNs with 3 × 1 convolution kernel have good classification performance with few parameters and fast training rate to be more suitable for EEG signal classification than 2D CNNs for small EEG dataset.However,2D CNNs are suitable for large EEG dataset because of their fast training rate.RCNNs are more suitable for EEG signal classification than CNNs due to its recurrent unit,few parameters and high classification accuracy.We also find that the large-margin softmax loss functions are also suitable to classify EEG signal with high classification accuracy and can prevent overfitting.The MRCNNs have the best classification performance.