Multi-electrode recordings and high-resolution stimuli were used to record from populations of identified RGCs in isolated primate retina while stimulating individual cones. Responses were fitted with a model consisting of two linear-nonlinear stages. The first stage consists of subunits that linearly combine signals from groups of cones followed by a nonlinearity. The second stage is a weighted sum of subunit responses followed by a final output nonlinearity. The assignment of cones to subunits was inferred using a greedy search for assignments that maximized response likelihood. Estimates of weights at both stages, as well as a smooth parameterization of the subunit nonlinearity, were obtained using block coordinate ascent on likelihood.

Fitted subunits for ON and OFF midget RGCs revealed varying degrees of rectification. Subunits typically included 1-3 cones, and convergence varied with eccentricity as predicted from anatomical data. The improvement in explained variance of RGC responses was typically 10-20% over a standard linear-nonlinear model for white noise stimuli, but much larger for noise segments that maximally differentiated the models. Additional validation was performed with repeated white noise, sinusoidal gratings, and targeted stimulation of selected pairs of cones. The results provide a picture of nonlinear signaling and circuitry in RGC populations at cellular resolution.