Reducing Calibration Time Using Novel Hybrid Transfer-Learning for P300-Based BCI Applications

P300 is an event-related potential determined by the changes in natural neurons activity, which occurs mainly as a response to the infrequent stimuli. Considering that the positive potential can be monitored by non-invasive methods such as electroencephalogram, and that the 'oddball' paradigm elicits deliberately this response, P300 can be used in brain-computer interfaces (BCI). P300-based BCI applications suffer from the subject dependency problem, which is one crucial issue in the real-time implementation, requiring time-consuming calibration and a large number of training samples for learning the model. In this paper, a new approach based on transfer learning to overcome these problems is proposed, where the fine-tuning ability of a deep neural network for high-level feature extraction is being used. Euclidean space data alignment approach is adopted to make feature extraction data give similar distributions. Finally, transferred features are applied to a single-layer discriminative restricted Boltzmann machine for P300 detection. We have used a state-of-the-art dataset (BCI Competition III dataset II) for evaluating the proposed method. The results show that the proposed technique offers robust performance using a small number of training samples compared to the previous studies.