Optimized visual stimuli for BCI with hessenberg decomposition based extreme learning machine

Abstract

AbstractBackground and ObjectivesBrain-computer interfaces (BCIs) aim to provide neuroscientific communication platform for human-beings, in particular locked-in patients. In most cases event-related potentials (ERPs), averaged voltage responses to a specific target stimuli over time, have key roles in designing BCIs. With this reason, for the last several decades BCI researchers heavily have focused on signal processing methods to improve quality of ERPs. However, designing visual stimulus with considering their physical properties with regard to rapid and also reliable machine learning algorithms for BCIs remain relatively unexplored. Addressing the issues explained above, in summary the main contributions of this study are as follows: (1) optimizing visual stimulus in terms of size, color and background and, (2) to enhance learning capacity of conventional extreme learning machine (ELM) using advanced linear algebra techniques.MethodsTwo different sized (small and big), three different colored (blue, red and colorful) images with four different backgrounds (white, black and concentric) for each of them were designed and utilized as single object paradigm. Hessenberg decomposition method was proposed for learning process and compared with conventional ELM and multi-layer perceptron in terms of training duration and performance measures.ResultsPerformance measures of small colorful images with orange-concentric background were statistically higher than those of others. Visual stimulus with white background led to relatively higher performance measures than those with black background. Blue colored images had much more impact on improvement of P300 waves than red colored ones had. Hessenberg decomposition method provided 1.5 times shortened training duration than conventional ELM, in addition with comparable performance measures.ConclusionsHerein, a visual stimuli model based on improving quality of ERP responses and machine learning algorithm relies on hessenberg decomposition method are introduced with demonstration of their advantages in the context of BCI. Methods and findings described in this study may pave the way for widespread applications, particularly in clinical health-informatics.