They reported an accuracy of baseline FBCSP and ConvNets on the BCIC dataset around ~68%-70%, which is reminiscent of early WER for speech recognition (prior to

They reported an accuracy of baseline FBCSP and ConvNets on the BCIC dataset around ~68%-70%, which is reminiscent of early WER for speech recognition (prior to the advent of specialized speech architectures). However, when you consider the datasets (BCIC 2a and 2b) are only using 22 scalp electrodes (for 2a) and 3 (for 2b) at 250 Hz, that's pretty amazing. That's analogous to using something like 8-bit audio for speech recognition. Clearly there is lots of room for improvement here, both in neural net architectures, measurement technology, and preprocessing. One technique the authors didn't explore is source localization [1], which typically requires much higher density (somewhere on the order of 128-256 active electrodes), but offers much higher spatial resolution (up to ~5mm isotropic). Given that most of the EEG datasets are recorded on less than 32 channels, I wonder how much more signal can be detected from such low resolution methods, or if we're approaching the theoretical maximum channel capacity of 32 scalp electrodes. It's like Feynman's analogy of measuring the height of corks in a swimming pool to determine the physics of an object that hit its surface. The more corks you have, the more information about the source (location, velocity, trajectory, geometry) you can reconstruct. [1]: http://www.scholarpedia.org/article/Source_localization#Electroencephalography_.28EEG.29 http://www.scholarpedia.org/article/Source_localization#Elec...