LAUSR

Abstract Since the deep learning classifier has the capability to hierarchically abstract the useful information from the physiological signals, it receives more attention in human emotion recognition in recent studies. Considering the structure of the deep network is required to be independently determined for multiple physiological modalities, we propose an ensemble deep learning framework by integrating multiple stacked autoencoder with parsimonious structure (M-SAE) to reduce the model complexity and improve the recognition accuracy. In M-SAE framework, the physiological feature abstractions from the deep hidden neurons of each signal modality are separately extracted via a group of member SAEs. The structural hyper-parameters are identified by minimizing the loss of the data distribution similarity across the original features and activation potentials in the hidden layer. The performance comparison on DEAP database validates the competence of the M-SAE against several classical emotion classifiers.