An exploratory study of transfer learning frameworks in the context of few available shots of neurophysiological signals

Neurophysiological signals are relevant to human motions and can be used for exploring underlying neural mechanisms of human locomotion. Deep learning has been popular in the field of human-machine interaction in recent years due to its superior capacity in decoding of neurophysiological signals, such as electroencephalography (EEG) and electromyography (EMG). However, deep learning usually requires a large amount of data to be driven. In practice, it is always difficult to collect required data due to insufficient amount, poor quality, and missing data labels. In this study, we proposed two transfer learning frameworks to improve the decoding performance of deep learning in the context of only few available shots. EEG and EMG data were collected in an exoskeleton-aided walking experiment with four conditions. The results showed that classification accuracy was improved by up to 10.02% after using the proposed transfer learning frameworks. In addition, we compared different transfer learning strategies and found that convolution neural network based (CNN-based) transfer learning for EEG and EMG signals requires transferring only in convolution layers, while avoiding transferring in the fully connected (FC) layer. Based on the results, the proposed transfer framework constitutes practical tool to improve the decoding performance of neurophysiological signal for deep learning models in the context of few available shots, and our study provides a high reference value for the selection of transfer learning strategies.