Transfer Learning with Large-Scale Data in Brain-Computer Interfaces

Abstract

Human variability in electroencephalogram (EEG) poses significant challenges for developing practical real-world applications of brain-computer interfaces (BCIs). The intuitive solution of collecting sufficient user-specific training/calibration data can be very labor-intensive and time-consuming, hindering the practicability of BCIs. To address this problem, transfer learning (TL), which leverages existing data from other sessions or subjects, has recently been adopted by the BCI community to build a BCI for a new user with limited calibration data. However, current TL approaches still require training/calibration data from each of conditions, which might be difficult or expensive to obtain. This study proposed a novel TL framework that could nearly eliminate requirement of subject-specific calibration data by leveraging large-scale data from other subjects. The efficacy of this method was validated in a passive BCI that was designed to detect neurocognitive lapses during driving. With the help of large-scale data, the proposed TL approach outperformed the within-subject approach while considerably reducing the amount of calibration data required for each individual (similar to 1.5 min of data from each individual as opposed to a 90 min pilot session used in a standard within-subject approach). This demonstration might considerably facilitate the real-world applications of BCIs.