Subspace discovery with supervised learning for ssvep based brain computer interfaces

Official URL: https://risc01.sabanciuniv.edu/record=b2758680_(Table of contents)

With the advancements in the modern data streaming technologies and artificial intelligent systems, the amount of data becoming available to the applications do constantly increase everyday. Processing in large amounts, and also in high dimension, bring a great burden on both the system and algorithms for data-driven inference. To facilitate inference in high dimension, the presented thesis exploits that the data often lie on low-dimensional subspaces. We focus on the EEG signal processing for SSVEP (steady state visually evoked potentials)-based speller BCIs (brain computer interfaces), where the goal is character recognition (a multi-class classification problem) using the EEG SSVEP signal that is known to be dominated by a certain character-specific frequency component and its harmonics. Hence, the data of a character is in a linear subspace whereas separating, for instance, one class from the rest requires to separate a low-dimensional subspace (ambient is high) from a union of multiple low-dimensional subspaces. To this end, we propose a neural network that is tailored to the specific challenges of EEG SSVEP signal processing, i.e., handling data with high ambient but low intrinsic dimensionality as described. The multi-class classification problem was solved based on error correcting output codes. In our performance evaluations, after first testing the subspace detections of the proposed network on a synthetic dataset, we conducted extensive experiments with a publicly available benchmark SSVEP dataset and observed promising multiclass classification accuracy corresponding to an information transfer rate of 156 bits/min.