This paper presents a systematic application of machine learning techniques for classifying high-density EEG signals elicited by face and non-face stimuli. The two stimuli used here are derived from the vase-faces illusion and share the same defining contours, differing only slightly in stimulus space. This emphasizes activity differences related to high-level percepts rather than low-level attributes. This design decision results in a difficult classification task for the ensuing EEG signals. Traditionally, EEG analyses are done on the basis of signal processing techniques involving multiple instance averaging and then a manual examination to detect differentiating components. The present study constitutes an agnostic effort based on purely statistical estimates of three major classifiers: L1-norm logistic regression, group lasso and k Nearest Neighbors (kNN); kNN produced the worst results. L1 regression and group lasso show significantly better performance, while being abl e to identify distinct spatio-temporal signatures. Both L1 regression and group lasso assert the saliency of samples in 170ms, 250ms, 400ms and 600ms after stimulus onset, congruent with the previously reported ERP components associated with face perception. Similarly, spatial locations of salient markers point to the occipital and temporal brain regions, previously implicated in visual object perception. The overall approach presented here can provide a principled way of identifying EEG correlates of other perceptual/cognitive tasks.