Fig. 4

Using subunits to predict ganglion cell responses. a Illustration of model structures for the linear-nonlinear (LN) model (left) and the subunit model (right). The subunit model takes each identified subunit as a spatial filter and then sums their outputs after rectification. Both models pass the obtained filter signal through a final output nonlinearity to predict the cell’s firing rate. b Raster plots of measured spikes and simulations of the fitted models for a sample cell under frozen spatiotemporal white noise. Spikes were simulated by a Poisson process that takes the models’ firing rates over 33-ms bins as input. Scale bar, 1 s. c Comparison of variance explained over the frozen-noise section for the different models. Each data point corresponds to a different ganglion cell (N = 28 cells from one retina). Insets show histograms of differences in variance explained. The variance explained is significantly higher for the subunit model compared to the LN model (p = 9 × 10⁻⁴; Wilcoxon signed-rank test) as well as compared to the shuffled subunit model (p = 1 × 10⁻⁵). d Sample image used to measure responses to briefly flashed natural images and corresponding measured spike responses from a sample ganglion cell for several stimulus repeats. The red ellipse shows the 3-sigma outline of the ganglion cell receptive field. The shaded region in the raster plot shows the window over which spikes were counted. e Average spike counts (black line) and standard deviations (grey region) of the sample cell for each of the 300 images. f Comparison of rank-order correlations between predicted and measured responses for the different models. Each data point corresponds to a different ganglion cell (N = 46 cells from six retinas). Insets show histograms of differences in rank-order correlation. Differences between the subunit model and the other two models are significant (p < 10⁻⁶ in both cases; Wilcoxon signed-rank test)