A revised Inception-ResNet model-based transfer learning for cross-subject decoding of fNIRS-BCI

An extended calibration procedure is required to collect sufficient data for establishing a stable and reliable subject-specific classifier before the user can use a brain-computer interface (BCI) system based on functional near-infrared spectroscopy (fNIRS). In addition, individual differences and data collection conditions can lead to low generalization performance of the subject-specific classifier cross-subject. To address the above dilemma and improve the versatility of fNIRS-BCI system applications, we propose a revised Inception-ResNet (rIRN) model-based transfer learning (TL) to improve the cross-subject decoding accuracy of mental tasks. The TL-rIRN is a deep transfer learning model that combines an elaborated rIRN model for fNIRS signal classification with model-based transfer learning. The rIRN model is pre-trained using source domain fNIRS data from multi-subjects to extract the common features of neural activation among subjects. The pre-trained model is then fine-tuned using a small amount of calibration data from the target subject to obtain a fine-tuned transfer model that can accurately classify the new data from the target subject. The fNIRS data of eight participants are collected for the purpose of distinguishing between mental arithmetic and mental singing tasks. The leave-one-subject-out cross-validation method is used to evaluate the cross-subject decoding performance of TL-rIRN. The results show that TL-rIRN improves cross-subject decoding accuracy and effectively reduces model training time, calibration time, and computational resources.