World's fastest brain-computer interface: Combining EEG2Code with deep learning.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Non-invasive Electroencephalograms (EEG) are widely used for brain-computer interfaces (BCI).
• Decoding sensory information from EEG signals presents significant challenges.
• Previous BCI approaches have limitations in information transfer rate and stimulus discrimination.

Purpose of the Study:

• To present a novel deep learning approach for decoding sensory information from non-invasively recorded EEG.
• To evaluate the performance of this approach in both passive and active BCI applications.
• To investigate the information capacity of EEG signals for BCI control.

Main Methods:

• A deep learning model was developed to decode sensory information from EEG.
• The method was tested in a passive BCI to predict visual stimulus properties.
• The method was also applied to an active BCI for asynchronous spelling control.

Main Results:

• An average information transfer rate (ITR) of 701 bit/min was achieved in the passive BCI, with a peak online ITR of 1237 bit/min.
• Up to 100% accuracy was obtained in discriminating 500,000 visual stimuli using only 2 seconds of EEG data.
• An average utility rate of 175 bit/min was achieved for asynchronous spelling BCI, translating to 35 error-free letters per minute.

Conclusions:

• The presented deep learning method significantly outperforms previous approaches, extracting over three times more information from EEG.
• EEG signals contain more information than currently utilized for BCI applications.
• A performance ceiling for non-invasive visual BCI control may have been reached due to the high information content in EEG.