Abstract
Perceptual integration and detection of coherent form in cluttered scenes is a fundamental skill for survival in the complex environments we inhabit. Recent work proposes that the visual system is optimized through evolution and development to solve this challenge by exploiting statistical regularities in natural scenes, e.g. collinear alignment of edges. We investigated the potential role of shorter term plasticity mechanisms (i.e. learning) in shaping perceptual integration processes using contours that violate collinearity. We compared the ability of observers to detect contours embedded in noise when the Gabor elements defining the contours were a) aligned along the contour path (collinear contours), or b) oriented orthogonally to the contour path (orthogonal contours). Observers' detection performance was higher for collinear contours that are more frequently encountered in natural scenes than orthogonal contours. Importantly, training to detect orthogonal contours (2200–4000 trials, over 3–5 daily sessions) resulted in improved performance similar to that for the detection of collinear contours. fMRI measurements prior to training showed significant activations in extrastriate ventral and dorsal visual areas for collinear, but not orthogonal contours, when compared to stimuli in which the orientation of contour elements was randomly jittered by up to 45°. However, fMRI measurements on the same observers after training showed similar activation patterns for orthogonal and collinear contours in accordance with the behavioural learning effects. Critically, multivariate analysis revealed that classification of fMRI responses for orthogonal vs. collinear contours decreased after training, suggesting that learning shapes selectivity for global contours irrespective of the alignment of local elements (along vs. orthogonal to the path). Our findings provide novel evidence for experience-dependent plasticity in the human visual cortex that mediates our ability to detect contours in cluttered scenes and may contribute to the learning of statistical regularities in natural environments.