Simple model for encoding natural images by retinal ganglion cells with nonlinear spatial integration
Fig 6
Analysis of the optimal spatial scale for calculating the local spatial contrast.
A) Sample image overlaid with the 3-sigma outline of a sample cell’s receptive field (red curve). Below: Local stimuli after smoothing with a 2D Gaussian filter with increasing spatial scale from 15 to 195 μm and pixel-wise weighting by the sample cell’s spatial receptive field. The first image is without smoothing. B) Relation between linear signal of the SC model and measured spike count, using the same Imean values, but LSC values derived from the differently smoothed images, displayed here for the original image and for the images with spatial scales of smoothing of 60 and 195 μm, respectively. The orange lines show the fitted nonlinearities, and the R2 values denote the corresponding model performance. C) Prediction improvement as a function of the level of smoothing for the sample cell. The optimal spatial scale is defined as the spatial scale at which R2 reaches its maximum (as determined by the 2nd-order polynomial fit around the maximal data point; green line). D) Prediction improvement, normalized by the prediction improvement with no image smoothing, as a function of the level of smoothing for all cells, shown separately for the four cell classes of Fig 4. The data from the sample cell is shown in black. E) Distributions of optimal spatial scales. F) Distributions of optimal scales, normalized by each cell’s receptive field size.