Behavioral experiments have found that infants and adults learn statistically defined patterns presented in auditory and visual input sequences in the same manner regardless of whether the input was linguistic (syllables) or nonlinguistic (tones and shapes). In order do determine the time course and neural processes involved in online word segmentation and statistical learning of visual sequence, we recorded event-related potentials (ERPs) while participants were exposed to continuous sequences with elements organized into shape-words randomly connected to each other. After viewing three 6.6min sessions of sequences, the participants performed a behavioral choice test. The participants were divided into two groups (high and low learners) based on their behavioral performance. The overall mean performance was 72.2%, indicating that the shape sequence was segmented and that the participants learned the shape-triplets statistically. Grand-averaged ERPs showed that triplet-onset (the initial shapes of shape-words) elicited larger N400 amplitudes than did middle and final shapes embedded in continuous streams during the early learning sessions of high learners, but no triplet-onset effect was found among low learners. The results suggested that the N400 effect indicated online segmentation of the visual sequence and the degree of statistical learning. Our results also imply that statistical learning represents a common learning device.