cuijiancorbin Goto Github PK

a-compact-and-interpretable-convolutional-neural-network-for-single-channel-eeg

In this project, we propose a CNN model to classify single-channel EEG for driver drowsiness detection. We use the Class Activation Map (CAM) method for visualization. Results show that the model not only has a high accuracy but also learns biologically explainable features, e.g., Alpha spindles and Theta burst, as evidence for the drowsy state.

benchmarking-eeg-based-cross-dataset-driver-drowsiness-recognition-with-deep-transfer-learning

In this project, we implemented the EDJAN model for EEG-based cross-dataset driver drowsiness recognition. The proposed model achieved mean accuracies of 83.68% and 76.90% on the cross-dataset transfer task.

eeg-based-cross-dataset-driver-drowsiness-recognition-with-an-entropy-optimization-network

eeg-based-cross-subject-driver-drowsiness-recognition-with-an-interpretable-cnn

Existing work in the field of BCI treats deep learning models as black-box classifiers. In this project, we develop a novel model named "InterpretableCNN" that allows sample wise analysis of important features for classification. The model not only achieves SOTA classification accuracy of EEG signals but also reveals meaningful features from EEG.

subject-independent-drowsiness-recognition-from-single-channel-eeg-with-an-interpretable-cnn-lstm

In this project, we propose a CNN-LSTM model to classify single-channel EEG for driver drowsiness detection. We designed a visualization technique by taking advantage of the hidden states output by the LSTM layer. Results show that the model not only has a high accuracy but also learns biologically explainable features, e.g., Alpha spindles, as evidence for the drowsy state.

towards-best-practice-of-interpreting-deep-learning-models-for-eeg-based-bci

In this project, we implemented 7 interpretation techniques on two benchmark deep learning models "EEGNet" and "InterpretableCNN" for EEG-based BCI. The methods include: gradient×input, DeepLIFT, integrated gradient, layer-wise relevance propagation (LRP), saliency map, deconvolution, and guided backpropagation