Figure 3.

Time-adaptive STRF estimation from simulated responses to human speech. (A) Model cell with time-varying STRF whose inhibitory components gradually increase over time. Shown are averaged linear filters for non-overlapping parts of length 30 s. Responses were generated by filtering 300 s of speech spectrograms with the time-varying STRF and applying a static non-linearity to the filtered stimuli. (B) The temporal average of the ground-truth time-varying STRF. (C) STRF estimates for the different parts of data obtained using a GLM with zero-mean (top), adaptive (middle), and mixed (bottom) prior. For the adaptive and the mixed prior, the static STRF in (D) has been used during inference of STRF parameters. The GLM with zero-mean prior produces highly unreliable STRF estimates for the different parts. Both adaptive and mixed prior GLMs allow robust tracking of the time-varying linear filter. Numbers in the lower left corner of each filter estimate indicate correlation between estimated local STRF and mean STRF for each part. The GLM with mixed prior performs marginally better in some cases with strong deviation of the time-varying from the static STRF. (D) The static STRF estimated from the whole stimulus-response sequence. (E) Dependence of mixed prior performance on part length. Shown are mean and standard deviation of correlations between estimated local and average true STRF across 10 trials. As part length decreases, local STRF estimates become more unreliable as indicated by the noticeable increase in standard deviation. Part lengths of 15 s and longer allowed robust tracking of time-varying processing. (F) Comparison of performance of a GLM with adaptive and mixed prior for different human speech stimulus ensembles and different non-linearities. Across all conditions (N = 100), the GLM with mixed prior reveals a better reconstruction of the true time-varying STRF (Wilcoxon rank test).