Figure 2. Optimal nonlinear codes depend on the amount of input noise. Information transmitted by a pair of neurons is plotted as a function of average noise in the two pathways and the difference in thresholds between neurons. Both neurons have the same amount of input noise in (A) and finite difference in noise in (B). All curves are plotted for a fixed total spike rate of the two neurons across a range of inputs, described by a Gaussian function. Threshold values and noise are given in units of the input standard deviation. In (A) there is a critical amount of noise below which neuronal specialization into two classes occurs. The two maxima are equivalent because both pathways have the same amount of noise. In (B) finite noise differences adds a slope to the picture in (A). Breaking the symmetry between neurons, it becomes optimal to associate positive thresholds differences with positive noise differences. In other words, neuronal pathway with lower input noise should optimally have a threshold closer to the mean of the input distribution.