Fig 4.

A. Scatter plot compares noise-corrected prediction correlation between the linear-nonlinear (LN) model and local short-term plasticity (STP) model for each A1 neuron. Black points indicate the 56/187 neurons for which the local STP model performed significantly better than the LN model (p<0.05, jackknifed t-test). B. Mean performance (noise-corrected correlation coefficient between predicted and actual PSTH) for each model across the set of A1 neurons. The global STP model showed improved performance over the LN and local rectification (relu) model. The local STP model showed a further improvement over the global STP model (*p < 0.01, **p < 10⁻⁴, ***p < 10⁻⁶, Wilcoxon sign test, n = 187/200 neurons with above-chance prediction correlation for any model). The best performing model, the local STP model, reweighted the two input envelopes into five spectral channels, each of which underwent independent STP prior to linear temporal filtering and a static nonlinearity. C. Pareto plot compares model complexity (number of free parameters) versus average prediction correlation for model architectures with and without STP, with and without parameterization of the temporal filter (full vs. DO) and for variable numbers of reweighted spectral channels (rank). Models with STP showed (purple, blue) consistently better performance than models without STP (orange, red) for all levels of complexity.