Emotion Classification from Electroencephalogram Signals Using a Cascade of Convolutional and Block-Based Residual Recurrent Neural Networks

To determine the quality of experience for users of technological devices, we must consider the human influential factors, which encompass the emotional state. Hence, we propose a model to estimate user emotions from Electroencephalogram (EEG) signals. The model is a cascade of deep learning networks consisting of a pre-trained convolutional neural network which extracts spatial relations and residual block(s) of recurrent neural network which learn the temporal relations of multi-channel EEG signals and uses shortcuts across the neural layers for a more effective training of the deep network. We adopted the DEAP dataset to train and evaluate our model. To confirm that the proposed work is user-independent, we ensure that the data in the test set corresponds to subjects that are not included in the training set. We explored several input sets to determine the one that performs best on the DEAP dataset. We implemented existing popular state-of-the-art methods and compared with the proposed model. The results indicate that the proposed model consistently outperforms the previous models achieving 0.61 and 0.63 accuracy on the validation and 0.65 and 0.68 accuracy on the test dataset for valence and arousal respectively.