Influence of prior expectations on contour integration: psychophysics and modeling

Abstract

Contour integration is a process linking oriented andcolinearily aligned edge elements into coherent percepts.This process is thought to rely predominantly on feedfor-ward and recurrent integration mechanisms, making itspercepts very salient (pop-out) and independent of top-down processes. However, recent studies have shown thatattention can strongly modulate contour integration, sug-gesting the importance of the confluence of bottom-up(sensory input) and top-down (prior expectations) proc-esses for this task. To uncover neural substrates and mech-anisms underlying the influence of prior expectations oncontour integration, we combine psychophysical withmodeling studies. Participants had to carry out two exper-iments with identical visual stimuli but different behav-ioural tasks: a detection task (A) and a discrimination task(B). Stimuli consisted of vertical and horizontal ellipsesformed by aligned Gabor elements, being embedded in aGabor field with random orientations and positions. Eachvisual hemifield could contain one of a vertical, a hori-zontal or no ellipse. All combinations of these three basicconfigurations were possible, totalling to nine stimuluscategories for the two hemifields. In experiment A, partic-ipants had to give a yes response whenever one stimuluscontained at least one ellipse (contour); in experiment Bonly when a target was present (this target could be eithera horizontal or a vertical ellipse, in any hemifield of thestimulus).In the detection task, reaction times (RT) for horizontalellipses are ~70 ms shorter than for vertical ellipses. In thediscrimination task, RTs for targets are consistently shorterthan for distractors, even if the bias for horizontal ellipsesis taken into consideration. The presence of redundanttargets (e.g. two horizontal ellipses instead of only onehorizontal ellipse) also shortens RTs. Thus, the psycho-physical data clearly demonstrate a pronounced influenceof higher cognitive processes on contour integration.In our contour integration model [1], we explore thehypothesis that top-down influences directly modulatethe response characteristics of elementary feature detec-tors [2]. We first assume that the prevalence of horizontallines in natural images leads to an increased activation ofhorizontal feature detectors. In a discrimination task, weprovide a second bias either to horizontal or vertical fea-ture detectors. When presenting a stimulus from the ninepossible categories, any existing contour is integrated bythe model and leads to a higher activation of edge detec-tors stimulated by contour elements. The mean amplitudeof this activation is determined by the balance betweenthe two different biases. In conjunction with fixed thresh-olds for contour detection and discrimination, thesedynamics yield RTs that not only depend on the exactstimulus, but also on the top-down priming. The simu-lated RTs are in good qualitative agreement with our psy-chophysical experiments. We conclude that featureintegration may serve as a mechanism to convert priors onsingle features into preferences for global Gestalt proper-ties.