Supplementary Figure 4: Nonlinearity comparison with cross-validated log-likelihoods for exponential and linear rectification.
From: Encoding and decoding in parietal cortex during sensorimotor decision-making
Positive cross-validated log-likelihood difference implies exponential model better fits the data. We used Fano factor as a measure of neural variability, to test for a relation between variability and the quality of the fit with an exponential nonlinearity. There is a noticeable trend (correlation coefficient -0.44) indicating that underdispersed neurons (Fano factor less than 1) are better modeled with an exponential nonlinearity, while overdispersed neurons are often better modeled with linear rectification.
