Abstract
Statistical learning has been proposed as a mechanism enabling decomposition of complex sensory scenes. This type of learning occurs in multiple sensory domains and appears to happen without the viewer's explicit awareness. In vision, statistical learning may underlie the formation of object representations (Fiser and Aslin, 2001; Baker, Olson and Behrmann, 2004). While previous studies have demonstrated that humans can learn these regularities, little is known about the time course of this learning and the extent to which this mechanism is robust to changes in spatial configuration and the presence of distracting information. This study addressed these questions by means of an online measure of statistical learning. Participants viewed a scene containing multiple stimuli, one of which was a target (that indicated either a left or a right button press), one an associate, and the remaining were irrelevant distracters. Each target had a frequent associate (present on 80% trials) and an infrequent associate (present on 20% of trials). A given associate was paired equally often with left and right targets and carried no information about the response. Distracters were randomly selected and in no way aided observers in determining which target was present. Results indicated that statistical learning was rapid, occurring within 8 presentations of the target-associate pair. Furthermore learning occurred even though the spatial configuration of the target and associate varied from trial to trial and there were irrelevant distracters present. The ability of learning to take place despite configuration changes suggests that visual statistical learning is not just enabling the learning of specific spatial relations between stimuli, but rather that participants are picking up on a more abstract paired association between the target and its associate.