Diagram of the randomly connected network. An image is first convolved with each random group of filters in the first layer (the number of filters in a group is fixed but which filters belong to a group are random), the outputs of which are rectified by the sigmoid and the absolute functions. Then, in the second layer, the rectified outputs within each random group are pooled to produce a single feature map, which is subjected to the inhibition simulated by the convolution with the DoG function. Feature maps from different random groups are combined to produce a final saliency map as output.
Qi et al. BMC Neuroscience 2012 13:145 doi:10.1186/1471-2202-13-145