A Recurrent Network for Tilt Illusions Based on Retinotopic Coding

Abstract

We propose a computational model for the mechanism underlying the perception of tilt illusions. Several studies have suggested that cortical mechanisms based on orientation-dependent lateral connections account for tilt illusions such as the Zöllner and Delboeuf illusions [1970 Nature (London)228 37 – 39; 1989 Biological Cybernetics58 35 – 49]. In these studies, it has been proposed that two intersecting bars evoke strong activation of V1 cells tuned to the orientations of the bars, and that, because of lateral interactions, cells tuned to orientations several degrees off those of the bars are the most strongly activated of all orientation-tuned cells. Thus the peaks in the neural orientation profiles are shifted by several degrees compared with the actual orientations of the bars. We investigate the retinotopic distribution of V1 cell activities in simulations of a large-scale recurrent network based on V1 anatomical connectivity, and show that the orientation profile of cell activities is significantly different for different retinotopic positions. For most positions, a single peak appears in the orientation profile. Taking into account population coding of the cells, we show that the bars will be tilted in a stimulus-configuration-dependent manner. The network shows angle expansion at crossed (X-shaped), branched (Y-shaped), or bent (L-shaped) intersections if the angle between the two bars is acute, and contraction if the angle is obtuse. The maximum expansion appears between 15° and 30°. These results are in good agreement with psychophysical observations on tilt illusions, including the Zöllner, Delboeuf, and Poggendorff illusions.