Fig. 2
From: Proscription supports robust perceptual integration by suppression in human visual cortex
Schematic illustration of the proscriptive integration model of cue integration. a Cartoon of the network’s detectors. The input layer of the model consists of unimodal detectors for slant defined by disparity or texture. These are represented along the ordinate and abscissa. The next layer of the model is the combination units, which are tuned to different combinations of unimodal inputs. b Overview of the model’s architecture. A slanted 3D Voronoi stimulus (illustrated for red–cyan anaglyph glasses) activates separate unimodal units tuned to disparity or texture. The response of unimodal units is pooled by combination units. The activity of combination units is then read out according to a cosinusoidal weighting function to produce a likelihood function of slant angle. c Bias and reliability as a function of cue conflict, generated by the proscriptive integration model using experimental data from Girshick and Banks7. d Example simulation of the model when presented with conflicting cues. Combination unit activity is equal to the sum of unimodal (texture/disparity) responses. This response is multiplied by a cosine weight matrix, representing readout weights to the output layer. The final estimate is produced by summing diagonally
