13. The expected amplitude of the average of N trials with randomly distributed phase is 1/ ͌ N of mean single-trial amplitude. On average, the single-subject ERPs were averages of 922 trials, so the expected mean ԽERPԽ/ԽEEGԽ amplitude ratio was 1/ ͌ 922 ϭ -29.65 dB. 14. S. Makeig, Electroencephalogr. Clin. Neurophysiol. 86, 283 (1993). 15. Time/frequency analyses used three-cycle Hanningwindowed sinusoidal wavelets at each frequency moved through the 3-s data epochs (from -1 s before to 2 s after stimulus onset) in 14-ms steps. Eventrelated spectral perturbation (ERSP) plots showed time/ frequency points at which mean log power, across the input epochs, was higher or lower than mean power during the 1-s prestimulus baseline period of the same epochs. The ERSP transform of the averaged ERP data ( Fig. 1C) was computed for each scalp channel across the 15 individual-subject ERPs. The ERSP transform of the single-trial epochs (Web. fig. 3B) was computed, for each subject and scalp channel, across a mean of 922 single-trial epochs. Bootstrap P Ͻ 0.02 significance levels were computed from distributions of ERSP values computed from surrogate data windows drawn at random from the same data epochs. 16. Also referred to as "phase-locking factor." [C. Tallon-Baudry, O. Bertrand, C. Delpuech, J. Pernier, J. Neurosci. 16, 4240 (1996)]. Window lengths, step size, and significance levels were the same as for the power analyses (15). 17. These results parallel the demonstration of Sayers et al. (2) that the averaged auditory ERP in their experiments was produced by a stimulus-induced partial phase resetting of EEG rhythms in single trials. 18. These data used a right mastoid reference. Similar relationships between alpha phase and subsequent ERP morphology have been reported by B. H. Jansen, M. E. Brandt, Electroencephalogr. Clin. Neurophysiol. 80, 241 (1991), and related results in (19-21). . 26. For each subject, ICA training data consisted of approximately 922 concatenated 52-point, 31-channel data epochs. Initial learning rate was 0.004; training was stopped when learning rate fell below 10 Ϫ6 . Initial block size was 128. Training required less than 30 min per subject on a PC workstation. Source and binary code for the enhanced version (24) of the infomax ICA algorithm (25) we used are available, together with a MATLAB toolbox for EEG time/frequency analysis and visualization, from www.sccn. ucsd.edu. 27. A relatively large ICA training data set (31 channels by 150,000 time points) was used for maximum stability assuming similar early visual processing in all conditions. Detailed comparisons of different stimulus and attention conditions including target trials will be pursued in future studies. 28. Clustering was based on normalized component activity spectra and scalp maps, and used a modified Mahalanobis distance metric [S. Enghoff, "Moving ICA and time-frequency analysis in event-related EEG studies of selective attention," INC-9902 (Institute for Neural Computation, La Jolla, CA, 1999)]. 29. BESA software (Megis So...