We simulated a population of correlated, noisy MT neurons with 60 preferred speeds that uniformly tiled the log space between 0.5 and 512 deg/s and 60 preferred directions evenly distributed between −180 and 180 degrees. Each model unit took on a scalar mean response Rmean determined by the sum of the baseline activity R0 and the product of the direction and speed tuning curves: equation(Equation 17) Rmean(θ,s)=R0+ge−0.5(log(s/ps)σs)2e−0.5(θ−pθσθ)2 Here, s and θ are stimulus speed and direction, and ps and pθ are preferred speed and direction. We set amplitude g = 4, R0= 1, bandwidth of the speed tuning σs = 1.5, and bandwidth of the direction tuning σθ = 40. The angle θ-pθ ranged from −180 to +180. The

MEK inhibitor magnitude of the MT-pursuit correlations depended strongly

on the value of g, and we selected the value 4 to reflect our expectation of a mean response of 4 spikes in the 40 ms intervals used to analyze our data for MT neurons with preferred speed and direction near the parameters of target motion. We followed the methods of Shadlen et al. (1996) to add to each neuron’s mean response correlated noise drawn from a normal distribution with the variance scaled to the mean response. The expected correlation structure rij between neurons i and Lumacaftor mouse j was: equation(Equation 18) rij=rmaxe−(log(psi/psj)τs)2e−(pθi−pθjτθ)2 We set the peak correlation rmax = 0.18, and the widths of the correlation structure for speed and direction τs = 1.35 and τθ = PR-171 solubility dmso 45. These values are slightly different from those suggested by our prior report of neuron-neuron correlations in MT ( Huang and Lisberger, 2009). The values were chosen so that the amplitude of the best

model’s MT-pursuit correlations matched those in the data and the neuron-neuron correlations in the model MT populations provided good matches to the data from Huang and Lisberger (2009) for analysis intervals of durations 150 and 300 ms. Given that noise correlations are similar in those two windows, and MT responses are highly correlated across time ( Osborne et al., 2004), we see no reason to think that the noise correlations would be very different in the analysis window of 40 ms used here. We also do not think our conclusions are affected by our assumption that higher-order correlations in the MT population are small and would play little role in the structure of MT-pursuit correlations. We do realize that the exact parameter values for the neuron-neuron correlations are underconstrained by available data, and we take this uncertainty into account in interpreting our results. We thank Mehrdad Jazayeri for insightful comments on an earlier version of the manuscript and Mark Churchland for clarity on issues of population decoding. We are grateful to Lisberger lab members, as well as Jonathan Pillow, Peter Latham, Surya Ganguli, and Valerio Mante for valuable discussions.