Extended Data Fig. 10: Method comparison with spike-triggered clustering (STClus).

a, Receptive-field mosaics from a single peripheral marmoset retina recording. Sample cells are marked with red outlines. b, Summary of cell type responses to contrast-reversing gratings. c, Spatial filter from white noise of a sample cell for each cell type and the corresponding nonlinear subunit layout obtained by the subunit grid method. Darker pixels in spatial filters denote larger (positive) values. d, Nonlinear subunits obtained by STClus³⁰ applied to white-noise responses. The selected number of subunits maximized the likelihood of a validation set. e, The log-likelihood for different numbers of subunits for all cells of the same type. Error bars are 95% confidence intervals. f, Number of subunits that maximized the validation likelihood for each cell. Black bars are medians over cells belonging to the same type. g, We compared the prediction performance of the two models using the natural movie. STClus subunit outputs were exponentiated and summed to obtain generator signals. Summation weights were determined by the STClus fitting procedure using the white-noise data. Generator signals were then related to spiking responses by fitting a model-specific output nonlinearity³⁰, using the non-repeated part of the natural movie (as for the subunit grid model). h, Model performance comparison between the two nonlinear subunit methods.