Evolving training sets for improved transfer learning in brain computer interfaces

Abstract

A new proof-of-concept method for optimising the performance of Brain Computer Interfaces (BCI) while minimising the quantity of required training data is introduced. This is achieved by using an evolutionary approach to rearrange the distribution of training instances, prior to the construction of an Ensemble Learning Generic Information (ELGI) model. The training data from a population was optimised to emphasise generality of the models derived from it, prior to a re-combination with participant-specific data via the ELGI approach, and training of classifiers. Evidence is given to support the adoption of this approach in the more difficult BCI conditions: smaller training sets, and those suffering from temporal drift. This paper serves as a case study to lay the groundwork for further exploration of this approach.