Fig 10.
A. Impact of model output nonlinearity on prediction accuracy. Groups of bars compares mean prediction accuracy of LN (orange) and local STP models (blue) with different output nonlinearities using the vocalization-modulated noise data. In both the LN and STP architectures, the double exponential sigmoid shows better performance than a model with no output nonlinearity (linear), linear rectification (relu), and a logistic sigmoid (***p<10−5; NS: p>0.05 sign test). B. Comparison of initialization method and parameterization on LN (orange) and local STP model (blue) performance. Full models used non-parameterized temporal filter functions, and DO indicates model in which the temporal filter is constrained to be a damped oscillator. Single fits started from a single initial condition, and random fits stated a 10 different initial conditions, selecting the best-performing model on the estimation data. Initialization and parameterization had little impact on LN model performance, but both random initialization and DO parameterization improved performance for the local STP model (**p<10−4; ***p<10−5; NS: p>0.05 sign test).