IMPROVING THE CONVOLUTIONAL NEURAL NETWORK PERFORMANCE THROUGH TRANSFER LEARNING FOR BRAIN-MACHINE INTERFACE SYSTEMS

In the recent years, deep learning has been widely implemented for robotics applications. However, a main issue that remains to be solved especially for intelligent robot implementations is the limited number of training data. In this paper, we propose a transfer learning-based method to overcome this issue. To verify the performance of the proposed algorithm, we implemented the transfer learning in Convolution Neural Net-works (CNNs) that maps the human brain signals into motor movements and its impact on window size is studied. The focus of this work is to investigate the effect of trans-fer learning in CNNs for subjects performing similar Brain Machine Interface (BMI) tasks. The results are promising in terms of improving the recognition rate of CNNs. The trained CNNs are also implemented to map in real time the brain signals to the humanoid robot arm motion.