We used a novel application of principal components analysis (spatiotemporal PCA) to decompose the event-related brain potentials (ERPs) obtained with a dense electrode array, with the purpose of elucidating the late ERP components elicited by deviant stimuli under “attend” and “ignore” conditions. First, a “spatial” PCA was performed to identify a set of scalp distributions (spatial factors or “virtual electrodes”) that accounted for the spatial variance in the data set. The data were expressed as spatial factor scores or “virtual ERPs” measured at each of the virtual electrodes. These virtual ERPs were submitted to a “temporal” PCA, yielding a set of temporal factors or “virtual epochs.” Statistical analyses of the temporal factor scores found that (1) attended deviant stimuli elicited the P300 and Novelty P3 components, the latter being largest for highly salient nontargets; (2) “ignored” deviants elicited a small Novelty P3, and depending on the primary task, a small P300; and (3) the classical Slow Wave consisted of separate frontal-negative and posterior-positive components.