Enhancing Motor imagery Decoding via Transfer Learning

Motor imagery (MI) is arguably one of the most common brain–computer interface (BCI) paradigms. The decoding process, in many cases, involves the use of small amounts of data gathered over a period. The decoding performance might therefore be limited, due to the size of available data. Also, the non-stationarity of signals across sessions and subjects can pose a challenge to effective decoding. To solve these challenges, transfer learning is proposed as the suitable approach, which could yield optimal performance even with small amounts of data and handle the non-stationarity of signals with adaptation. It has been applied across domains and tasks where only small amounts of data are available and where signal distribution changes are more rapid. Transfer learning (TL) aids the decoding process by utilizing previous knowledge learnt in the decoding process to enhance future decoding. In this study, we apply the concept of transfer learning in the classifier space, yielding improvements in the decoding process up to 3%. We investigate its effect across two main scenarios: within- and across-subject motor imagery decoding. Within each of these scenarios, we consider how to optimally transfer useful knowledge, across sessions, to aid the decoding process, taking into consideration factors such as the update mode and data sources to be used for transfer across subjects. To our knowledge, this is one of the few works making such considerations in the application of transfer learning in electroencephalography (EEG)-based MI. From our results, we conclude that transfer learning is useful in motor imagery experiments, where small amounts of data are available, and can be used in mitigating the effect of non-stationarity across sessions and subjects