Abstract
A common model for the function of auditory cortical neurons is the linear-nonlinear spectro-temporal receptive field (LN STRF). This model casts the neural spike rate at each moment as a linear weighted sum of the preceding sound spectrogram, followed by nonlinear rectification. While the LN model can account for many aspects of auditory coding, it fails to account for long-lasting effects of sensory context on sound-evoked activity. Two models have expanded on the LN STRF to account for these contextual effects, using short-term plasticity (STP) or contrast-dependent gain control (GC). Both models improve performance over the LN model, but they have never been compared directly. Thus, it is unclear whether they account for distinct processes or describe the same phenomenon in different ways. To address this question, we recorded activity of primary auditory cortical neurons in awake ferrets during presentation of natural sound stimuli. We fit models incorporating one nonlinear mechanism (GC or STP) or both (GC+STP) on this single dataset. We compared model performance according to prediction accuracy on a held-out dataset not used for fitting. The STP model performed significantly better than the GC model, but the GC+STP model performed significantly better than either individual model. We also quantified equivalence between the STP and GC models by calculating the partial correlation between their predictions, relative to the LN model. We found only a modest degree of equivalence between them. We observed similar results for a smaller dataset collected in clean and noisy acoustic contexts. Together, the improved performance of the combined model and weak equivalence between STP and GC models suggest that they describe distinct processes. Models incorporating both mechanisms are necessary to fully describe auditory cortical coding.