Motor imagery classification via stacking-based Takagi–Sugeno–Kang fuzzy classifier ensemble

Abstract

Cyborg intelligence is committed to combining artificial intelligence with biological intelligence and human body technology through the tight integration of machines and biological organisms, so as to improve the natural capabilities of human beings. Brain–computer integration as a typical and effective way to explore cyborg intelligence has been widely applied in many scenarios. In cyborg systems regarding brain–computer integration, task planning in the artificial intelligent unit aims to identify intention of biological unit, which highly relies on a classifier. An urgent challenge for task planning is to effectively deal with the uncertainties associated with the complexity and variability of brain dynamics, which is reflected in the non-stationary nature of brain signals. This poses a severe problem for existing models to the intention identification in task planning. Additionally, model interpretability is also very important in task planning for several reasons such as human–machine communication, trust, etc. In this study, we focus on EEG-based motor imagery and develop a stacking-based ensemble fuzzy inference framework for task planning. In this framework, the zero-order Takagi–Sugeno–Kang (0-TSK) fuzzy system is taken as the basic component and the output of one component is augmented into the input feature space as new input for another component. This framework has two main merits: (i) its promising performance is guaranteed by the general stacking principle asserting that the manifold structure of the input feature space can be opened by the outputs generated from this space. (ii) its interpretability is guaranteed from three levels. Additionally, from a probabilistic statistical point of view, we theoretically demonstrate that as long as the basic learning component is a weak classifier, the general stacking principle can guarantee promising performance. Experimental results conducted on motor imagery data demonstrate the promising performance and high interpretability of the proposed framework.