Avoiding subject-specific model selection via highway networks in EEG signals

Deep neural network is a promising tool in EEG signals processing, while it commonly faces difficulties in model selection given biomedical data. Since biosignals from different subjects can be highly various, a specific neural network structure may over-fit with some subjects. On the other hand, subject-specific parameter selection is usually expensive and impractical. In this study, we propose to use highway networks to avoid subject-specific model selection in EEG signals. In virtue of adaptive structure adjustment in highway networks, our method can adjust the ability of network feature expression to prevent overfitting for different subjects. Experiments on Freiburg epilepsy dataset show that, our method improves F1-score by about 12% compared with classical CNN models. On SEED emotion classification dataset, the improvement of accuracy is about 0.6%. On BCI IV-2A motor imagery dataset, our method improves accuracy by about 1%, and achieves the state-of-the-art performance.