Supplementary Figure 5: Comparing temporal extent of post-spike filters.
From: Encoding and decoding in parietal cortex during sensorimotor decision-making
Majority of neurons showed a long self-exciting tail in the post-spike filter (Supplementary Fig. 7). We quantify how much information the tail contributes to, and test if there is a trial length scale modulation of gain. The post-spike filter used throughout the paper consists of ten 1 ms bins to model fast response components, in addition to ten raised cosine bases stretched logarithmically over 250 ms; here we also consider a model with shorter post-spike filter which has 7 temporal bases over 100 ms, and a model with longer post-spike filter which has 25 temporal bases over 5 s, which is long enough to cover the entire trial. (a) Resulting fits for 3 example cells, comparing short/medium/long spans for post-spike filter basis functions. Inset shows 5 seconds scale. (b) Spike prediction accuracy comparison across 3 models averaged over 80 neurons. Errorbar indicates standard error. Our "original" model used throughout the paper explained 7.9% more information per spike compared to short filter model (p < 10−10), and the long filter model explained 0.7% more information (p < 0.001). Thus, longer post-spike effect is negligible, and there is no systematic trial time scale gain modulation.
