Figure 2.

RF estimation from simulated responses using the different priors in a linear-Gaussian model. A strongly correlated Gaussian noise sequence (N = 1000 samples) was filtered with a temporal linear RF (red solid line). Spikes were generated by from the filtered stimulus by an inhomogeneous Poisson process resulting in 124 spikes. (A) The RF estimate obtained using a sparse (zero-mean) Gaussian prior on RF parameters. The estimate is very noisy due to the small number of observations. (B) Suppose we have estimated a static RF for the same cell, and the RF shows a different scaling and a slightly different latency (gray line). The adaptive prior biases RF coefficients toward the static RF. The resulting RF estimate is very smooth and partially adapts to the rescaled amplitudes but does not account for temporal shift and asymmetric amplitude scaling. (C) The mixed prior uses an additional hyper parameter that allows to control the trade-off between zero-mean and adaptive priors. The resulting estimator allows to account for the temporal shift at the expense of a slight increase in noise. Numbers in the upper left corner indicate correlations between estimated and true RF.