Improving Generalization Based on l1-Norm Regularization for EEG-Based Motor Imagery Classification

Yuwei Zhao,Peng Zhang,Hongji Sun,Yushu Chen,Ang Ke,Yi Zhang,Changyong Wang,Jin Zhou,Yao Han,Jiayun Chen,Jiuqi Han

doi:10.3389/fnins.2018.00272

Abstract

Multichannel electroencephalography (EEG) is widely used in typical brain-computer interface (BCI) systems. In general, a number of parameters are essential for a EEG classification algorithm due to redundant features involved in EEG signals. However, the generalization of the EEG method is often adversely affected by the model complexity, considerably coherent with its number of undetermined parameters, further leading to heavy overfitting. To decrease the complexity and improve the generalization of EEG method, we present a novel l1-norm-based approach to combine the decision value obtained from each EEG channel directly. By extracting the information from different channels on independent frequency bands (FB) with l1-norm regularization, the method proposed fits the training data with much less parameters compared to common spatial pattern (CSP) methods in order to reduce overfitting. Moreover, an effective and efficient solution to minimize the optimization object is proposed. The experimental results on dataset IVa of BCI competition III and dataset I of BCI competition IV show that, the proposed method contributes to high classification accuracy and increases generalization performance for the classification of MI EEG. As the training set ratio decreases from 80 to 20%, the average classification accuracy on the two datasets changes from 85.86 and 86.13% to 84.81 and 76.59%, respectively. The classification performance and generalization of the proposed method contribute to the practical application of MI based BCI systems.

Highlights

Noninvasive brain-computer interface (BCI) based on electroencephalography (EEG) has attracted an increasing interest in recent decades owing to its significant potential in practical applications (Wolpaw et al, 2002; Nicolas-Alonso and Gomez-Gil, 2012)
For the sake of mitigating generalization error caused by overfitting, we introduce a sparse l1-norm regularization to solve the optimal weights of channels during combination of each channel’s decision value, in which the sparse optimal weights are solved by minimizing the least square error between the predicted labels and the real labels
Results on Different Sizes of Training Sets To further verify the generalization of the proposed COL, we plotted the mean classification accuracy with the ratio of Subject b f

Summary

Introduction

Noninvasive brain-computer interface (BCI) based on electroencephalography (EEG) has attracted an increasing interest in recent decades owing to its significant potential in practical applications (Wolpaw et al, 2002; Nicolas-Alonso and Gomez-Gil, 2012). As a mirror of the total brain activity in specific regions, multichannel EEG signals interact with each other intrinsically. This interaction is believed to originate from the fundamental mechanism of the information processing within the brain, such as the distributed and co-related function of different cerebral cortex (Baillet et al, 2001). A specific brain activity is typically mirrored by more than one site on the scalp, leading to considerably redundant information involved in multichannel EEG signals. Informative EEG features such as task relevant and event-related potentials are likely mixed with blurred features and submerged into the raw data owing to the artifacts and merging effects of the conductive scalp and skull (Pfurtscheller et al, 2006). Due to the insufficient EEG data for classifier training, the complexity of classification algorithms may increase with redundant features involved in EEG signals, adversely affecting their generalization

Methods

Results

Discussion

Conclusion