Fig. 10: Figure-ground separation.
From: Modeling attention and binding in the brain through bidirectional recurrent gating
a Schematic of the network architecture used for the multi-attribute classification task, highlighting both the feature and attention pathways. Colored circles indicate the locations of neurons analyzed in this study. b Top: Receptive field of neurons from the feature pathway (F-neurons), indicated by the dashed red box. Bottom: Two stimuli used for analysis; stimulus (i), a colored object over a blank background; and stimulus (ii), a colored background, matching the object color in stimulus (i), containing an uncolored stencil object. c From the receptive field’s viewpoint, stimulus (i) rotated by θ° makes it identical to stimulus (ii) rotated by θ + 180°. The network successfully generates accurate attention maps across all stimuli variations, achieving 91% attention accuracy. d Polar plot showing the activity of an example F-neuron as a function of stimulus rotation. Dashed boxes outside the plot illustrate receptive field content for stimulus (i) (red) and stimulus (ii) (blue). e Histogram illustrating the distribution of phase differences of preferred orientation for F-neurons between the two stimuli (n = 119 neurons). f Polar plot showing the activity of an example A-neuron as a function of stimulus rotation, with receptive field content indicated in the same way. Because the attention pathway receives input from the bottleneck, A-neurons have receptive fields spanning the entire input scene. g Histogram showing preferred orientation phase differences of A-neurons between the two stimuli (n = 112 neurons). h Average activity of A-neurons at their preferred versus non-preferred orientations, shown for both stimulus types.
