Figure 1. The network architecture. The four nonlinear models presented in this study share the same connection scheme. A, Visual input drives the LGN, which provides excitatory input to IN and EX populations. The EX units receive inhibitory drive from the IN population and drive the DS units. B, Populations of ON- and OFF-center LGN units provide input to the V1 simple cell populations, with connections determined by a Gabor template (Eq. 4). The ON and OFF units marked here (white and black dots, respectively) provide input to the EX cell circled in white in D. C, An approximation of the RF corresponding to the synaptic map in B is shown (red box) along with RFs for templates having three other spatial phases at the same spatial (x,y) location in the map. The RFs are rendered by replacing each synaptic input with the center Gaussian of the LGN DOG filter (Eq. 6). D, The EX units are arranged around pinwheels in an orientation map. There are four layers of units, with each layer varying in spatial phase (0, 90, 180, and 270°). E, The IN inputs to a single EX unit are shown in color, coded to match the preferred orientation of the individual inputs (inset left). The weight of each IN input depends on how well its RF is anti-correlated with the RF of the postsynaptic EX unit. F, For efficiency, the DS population has only four units, all of which prefer vertical orientation (being centered on the orientation map) and motion to the right (0°). The DS unit circled in red receives an input from the EX unit circled in D. G, Averaged direction tuning curves at 100% stimulus contrast for DS units (purple trace, n = 4) and for the EX units (green trace, n = 16) and IN units (blue trace, n = 16) that had preferred orientation 0° (red region of orientation map). The left and right panels correspond to the presynaptic delay model and linear model, respectively.