The effects of data compression on the decoding of neuronal activity during handwriting with a comparison between linear and nonlinear models

Abstract: Brain computer interfaces (BCI) play a key role in enabling individuals who are suffering from the dreaded locked-in syndrome or severe forms of paralysis to communicate with the world around them. A state of the art BCI was developed by Willet et al. which allows users to achieve writing speeds of up to 90 characters per minute, using only their brain. This research aims to investigate how well the data provided in that study can be compressed and subsequently classified using different techniques in order to speed up computation time. Several linear and nonlinear methods were investigated in order to provide an understanding of the linear and nonlinear tendencies of the brain. By extrapolating from the results, it was possible to conclude that the linear dimensionality reduction method PCA was in most cases significantly better at compressing brain activity data compared to its nonlinear counterpart KPCA. Further, for highly compressed brain activity data, the nonlinear machine learning model KSVM was significantly better at classifying letters compared to its linear counterpart LR. Next, it was possible to conclude that increasing bin sizes of electrodes was an effective way of compressing the data, as long as it was subsequently compressed using PCA and not KPCA. Finally, it was also possible to conclude that only including the most important electrodes was an effective approach to reducing the size of the original data set.