Macaque parasol RGCs were recorded using 512­electrode arrays ex vivo. Responses were fitted with a model in which stimulus intensities are combined linearly to produce subunit activations, which are then exponentiated and summed to obtain the firing rate. For a given number of subunits, the likelihood of the model parameters is maximized by alternating between estimating subunit membership by soft­clustering the ensemble of stimuli preceding spikes, and estimating the linear weights associated with each subunit.

Fitting the model to simulated data, using fewer fitted subunits than the simulated number of bipolar cells, produced inferred subunits that subsumed several simulated bipolar cell RFs completely. Hence, a complete characterization may require more data. Subunits estimated from RGC data were spatially localized and non­overlapping, as expected from bipolar inputs of a single type, even though these properties were not assumed in the model. Finally, a "null" stimulus was presented, orthogonal to the linear RF estimated from spike­triggered averaging. A linear model predicts no response to this stimulus, but RGCs exhibited clear temporal structure. The subunit model captured some of this structure (R² ≈ 0.3).

A novel technique for estimating subunits of primate RGCs shows promising results in recovering putative bipolar cell spatial structure.