Fig. 5: Correlated activity results from fixations that evoke nonlinear responses.
From: Nonlinear receptive fields evoke redundant retinal coding of natural scenes
a, Top, sample image trajectory for the presentation of a single image. Middle, corresponding model predictions for two neighbouring OFF parasol cells (receptive fields in inset) for the DoG LN and the subunit grid models. Shaded areas mark two consecutive fixations, the first classified as differentiating (the two model predictions diverge) and the second as non-differentiating (model predictions align). Bottom, responses of the two cells during the same period. b, Comparison of average model performances (R2) for the subunit grid model and the DoG LN model during the top 20% maximally differentiating fixations. c, Contributions of linear and nonlinear fixations to the total pairwise correlations for the natural video (mean with 95% confidence intervals). d, Relationship between receptive-field nonlinearity, calculated as the ratio of subunit grid over DoG LN model performance during maximally differentiating fixations, and overall pairwise response decorrelation, calculated as the difference between stimulus and response correlations relative to the stimulus correlations. For cells with zero DoG LN performance, the ratio was set to the maximum value measured across cells of the same type. * denotes significant Spearman correlation (p = 0.037, two-sided permutation test). For b and d, error bars are median ± 95% robust confidence interval and number of cell pairs are ON parasol n = 269, OFF parasol n = 176, ON midget n = 355, OFF midget n = 494, transient-OΝα n = 63, transient-OFFα n = 315, sustained-ONα n = 2040, sustained-OFFα n = 889 from three marmoset and four mouse retina pieces (eight pieces for response decorrelation with n values given in Extended Data Fig. 5).
