Model estimation and validation procedures.

<p><b>A.</b> Data for each neuron were split into an estimation data set, used to fit model parameters, and a validation data set, used only for testing prediction accuracy. <b>B.</b> Models were defined by a sequence of functions mapping stimulus to predicted response. This study focused on alternative parameterizations of the spectro-temporal filter module, and other modules were kept fixed for most comparisons. <b>C.</b> Spectrogram of example 3-second ferret vocalization used for testing prediction accuracy (top). Raster response of one neuron to 20 repetitions (middle) and PSTH binned at 100 Hz (bottom, blue). Predicted PSTH response is overlaid in red. <b>D.</b> The STRF is typically implemented as a multichannel finite impulse response (FIR) filter, requiring one parameter for each frequency and time lag. In this heat map, red areas indicate stimulus frequencies and time lags associated with increased neuronal spike rate and blue areas with decreased rate. <b>E.</b> Factorized STRF approximation is generated by the outer product of spectral- and temporal filter matrices, reducing the total parameter count. <b>F.</b> Parameterized models generate factorized matrices from parametric tuning curves, further reducing parameter count. Example Gaussian spectral and P3Z1 temporal parameterizations are shown. Factorized and parameterized models also permit the insertion of nonlinear modules, such as a filter mimicking short-term plasticity (STP), after projection onto the spectral channels.