Figure 4. Robustness of RF filter estimation to stimulus distribution asymmetries, obtained with LNP model simulations and asymmetric white noise stimuli.

(A) Ground truth linear filter underlying the simulations. (B) Stimulus amplitude distribution with long tail towards positive values, created by drawing samples from a Gaussian white noise distribution and subsequent expansion (compression) of positive (negative) amplitudes, respectively. The stimulus auto-covariance matrix remains diagonal, simplifying the linear estimator to STA without covariance correction. 50 samples of the temporal stimulus sequence shown in the inlet. (C) Estimates of the linear filter obtained using STA, MID and CbRF approach. While the latter two methods reconstruct the true linear filter faithfully, the STA-based estimate shows a non-symmetric scaling of the positive and negative deflection. Linear filter amplitudes rescaled to arbitrary units (a.u.) in the interval for visualization.