2.4. EEG acquisition and preprocessing

EEG was recorded by an Active 2 system (BioSemi, the Netherlands) using 64 sintered Ag/AgCl electrodes, at a sampling rate of 256 Hz with an online low-pass filter of 51 Hz to prevent aliasing of high frequencies. Additional electrodes were placed as follows: two on the mastoid processes, two horizontal EOG channels positioned at the outer canthi of the left and right eyes (HEOGL and HEOGR, respectively), two vertical EOG channels, one below (infraorbital, VEOGI) and one above (supraorbital, VEOGS) the right eye, and a channel on the tip of the nose. Electrodes were referenced to the average of the entire electrode set, excluding the EOG channels. Offline processing included a band-pass filter of 0.3–20 Hz, and computing bipolar vertical EOG (VEOG) and horizontal EOG channel (HEOG) channels as the difference between VEOGS and VEOGI for VEOG, and the difference between HEOGL and HEOGR for HEOG. The recorded data was segmented to 900 ms segments starting at each image onset. Therefore, each single trial recording yielded a data matrix of 900 ms over 64 channels. The data matrices were downsampled to 64 Hz to reduce computational time, and each dimension of the matrix was normalized to zero mean and variance. The resulting matrix consisted of 64 rows of EEG channels and 64 columns of time samples. We removed the DC baseline from each channel separately. Large artifacts, e.g., blinks, were removed by rejecting trials in which the VEOG bipolar channel exceeded ±100 μV. The images of the stimuli were downsampled to 50 × 50 pixels before using them as input to the network.