Transfer Learning for SSVEP Electroencephalography Based Brain–Computer Interfaces Using Learn++.NSE and Mutual Information

Matthew Sybeldon,Lukas Schmit,Murat Akcakaya

doi:10.3390/e19010041

Matthew Sybeldon, Lukas Schmit + Show 1 more

Open Access

https://doi.org/10.3390/e19010041

Copy DOI

Journal: Entropy	Publication Date: Jan 19, 2017
Citations: 7	License type: CC BY 4.0

Affiliation: University of Pittsburgh, UPMC Hillman Cancer Center

Abstract

Brain–Computer Interfaces (BCI) using Steady-State Visual Evoked Potentials (SSVEP) are sometimes used by injured patients seeking to use a computer. Canonical Correlation Analysis (CCA) is seen as state-of-the-art for SSVEP BCI systems. However, this assumes that the user has full control over their covert attention, which may not be the case. This introduces high calibration requirements when using other machine learning techniques. These may be circumvented by using transfer learning to utilize data from other participants. This paper proposes a combination of ensemble learning via Learn++ for Nonstationary Environments (Learn++.NSE)and similarity measures such as mutual information to identify ensembles of pre-existing data that result in higher classification. Results show that this approach performed worse than CCA in participants with typical SSVEP responses, but outperformed CCA in participants whose SSVEP responses violated CCA assumptions. This indicates that similarity measures and Learn++.NSE can introduce a transfer learning mechanism to bring SSVEP system accessibility to users unable to control their covert attention.

Highlights

Brain–Computer Interfaces (BCI) are an emerging input modality for disabled users seeking to communicate by computer [1,2]
Responses, but outperformed Canonical Correlation Analysis (CCA) in participants whose State Visual Evoked Potentials (SSVEP) responses violated CCA assumptions. This indicates that similarity measures and Learn++.NSE can introduce a transfer learning mechanism to bring SSVEP system accessibility to users unable to control their covert attention
We introduce an ensemble learning technique called Learn++.NSE combined with similarity measures such as mutual information or Mahalanobis distance for data set selection

Summary

Introduction

Brain–Computer Interfaces (BCI) are an emerging input modality for disabled users seeking to communicate by computer [1,2]. BCI systems use brain signals directly as input to infer user intent. These systems are useful for users capable of minimal movement who cannot rely on typical input modalities such as keyboards, mice, or joysticks. Typical BCI systems require calibration sessions to collect labelled training data. These data points are used to estimate the statistical distribution of the features calculated from the electroencephalography (EEG) signals.nonstationarities exist in the EEG signal. Common nonstationarity sources include artifacts, equipment changes, environmental variables, and user fatigue [1] This last issue is of particular interest since lengthy calibration can cause user fatigue and limit the application of the learned data.

Methods

Results

Conclusion