Fig 1. Network diagrams and input statistics.

(A) The general network architecture characterized by an overcomplete representation with N input neurons, x_i (i = 1,…,N) and M ≥ N output neurons, s_i (i = 1,…,M). The input is linearly transformed by the feedforward connection matrix,W, and then nonlinearly processed by the recurrent dynamics at the output layer. The interactions among the output neurons are denoted by the matrix K. (B) A toy network model of a visual hypercolumn containing 2 input neurons and M output neurons. The feedforward connections are preset to unit vectors spanning all angles at equal intervals. The inputs are points on the plane with uniformly distributed angles and normally distributed distances from the origin. The distance from the origin represents the input contrast. (C) An ecological network model in which the inputs are natural images. The feedforward connections are Gabor filters with orientations equally spaced between 0° and 180°.