A Transfer Learning Approach towards Zero-training BCI for EEG-Based Two Dimensional Cursor Control

Non-invasive Brain Computer Interface (BCI) has attained substantial achievements in a number of areas such as rehabilitation and neuroprosthetics. However, the costly and time consuming process of acquiring subject-specific data and model calibration limit the expansion of BCI utilization. Moreover, due to the large variability of inter-subject and inter-trial data patterns, we cannot expect a satisfactory result from classifier on a new subject. In this study, we introduced a novel offline BCI platform using imagined body kinematics paradigm for computer cursor control in which the issue of inter-subject and inter-session model variability were addressed. We present a method to transfer parameters extracted from a pool of labeled recorded data of individuals (source) and generate a tailored classifier for a new subject (target) with minimal calibration. Due to disparity of the source and the target in terms of parameter distribution and variations in electroencephalography (EEG) response, part of the source data contributed to learn the target model while there were other parts harmful to the learning phase. Consequently, a measure was introduced to partake only the suitable part of the source data. Despite the fact that the BCI paradigm we resorted to is historically more prone to suffer from subject variability, the results showed encouraging figures in terms of both predictability and saving the number of target-specific samples. The lagged samples of subject's EEG from the whole brain activity contributed to a regression model. A measure of likelihood between the observed cursor velocity and the corresponding predicted velocity has been introduced. This measure not only was responsible for pulling out a satisfactory set of subjects' trials in order to construct a pool of data, but also defined how source data would impact the target model. Figure 1 depicts the average likelihood of the subject-specific model and the corresponding transferred model for horizontal cursor velocity over all participants.