Fig. 1. Predictive coding and natural scene statistics.

a, Schematic of the linear model of a receptive ganglion cell encoding noisy photoreceptor outputs. b, Average stimulus power in the mouse FOV in the UV range (natural image data, courtesy of H. Asari¹⁸). Orange dashed line denotes the simulated horizon. Orange frame illustrates the size of the model RF. c, Stimulus power in the UV range (left; red line) and example noise power level (left; gray line) as a function of elevation in the visual field. Increasing stimulus power increases the SNR (right). d, Vertical SNR asymmetry in the UV range as a function of elevation in the visual field (left). Change in SNR asymmetry is due to asymmetric power in the stimulus at the horizon line (right). e, Predictive coding RFs optimal for different levels of SNR. RFs were smoothed with a 2 × 2-pixel window for display purposes. f, Relative surround strength (top) and center size (bottom) of optimal predictive coding RFs increase and decrease respectively, with increasing photoreceptor SNR. Purple, green and orange lines correspond to the UV, green and joint spectra, respectively. g, Predictive coding RFs optimal for different levels of vertical SNR asymmetry. RFs were smoothed with a 2×2-pixel window for display purposes. h, Surround asymmetry of optimal predictive coding RFs increases with increasing vertical SNR asymmetry of photoreceptor output. Line colors analogous to f.