Fig 3. Examples comparisons of actual versus predicted PSTHs for two A1 neurons.

A. Segment of the stimulus spectrogram from the validation dataset used for testing prediction accuracy. B. Actual PSTH (gray) overlaid with the PSTHs predicted by the 1Dx2-CNN (blue), 2D CNN (light green), and pop-LN (purple) models. The two CNN models produced similar predictions that correlated well with the actual response (1Dx2-CNN, r = 0.812, 2D CNN: r = 0.800). The pop-LN model failed to predict many of the temporal features in the actual PSTH (r = 0.590, arrows at 3.5, 4.0 sec). C. Actual and predicted PSTH for a second neuron, plotted as in B. Again, both CNN models had higher prediction correlation. Here their accuracy is evident in the observation that they accounted for large transient responses better than the LN model (1Dx2-CNN, r = 0.729, 2D CNN: r = 0.748, pop-LN: r = 0.535; arrows at 0.25, 1.75 sec).