Figure 5. Robustness of RF filter estimation to higher-order correlations in the stimulus ensemble.

(A) Ground truth linear filter underlying the simulations. (B) Examples of sinusoid grating stimuli conveying higher-order correlations. The stimulus ensemble was composed of 80,000 grating stimuli with random orientation and spatial frequency. Second-order correlations were removed by a whitening transformation prior to simulation and analysis. (C) Filter estimates obtained with STA, MID and CbRF methods. Quadratic (p  = 2, upper row) and cubic (p  = 3, lower row) nonlinearities were used for LNP model simulations. Overestimation of the RF filter support visible in the STA result is a result of higher-order stimulus correlations, since the stimulus (second-order) auto-covariance matrix was diagonal by construction. Correlation of estimated with true RF filter indicated in lower right corner of each plot.