Fig. 4.

Evaluation of bias-correction procedures. To evaluate the accuracy of our signal modulation measures, a population of 150 simulated neurons with known amounts of signal modulation were created from measured responses in IT and PRH. A: fractional bias, calculated as the ratio of the total bias (summed across all signal modulations) divided by the total signal (summed across all signal modulations), plotted for the uncorrected simulated population (black), the closed-form bias correction (red), and the bootstrap bias correction (cyan) as a function of the number of Poisson trials collected in each simulated experiment when spikes were counted in 50-ms bins centered 125 ms after stimulus onset. Shown are the averages over 100 simulated experiments. Plot on right shows an enlargement of the boxed region indicated on left. B: a histogram of the average (across the 150 simulated neurons) fractional error (total error/total signal) remaining after the closed-form correction for each of 100 simulated experiments with 20 Poisson trials to show that the fractional bias measured per experiment is always near 0. C, left: histogram of the fractional bias remaining for 1 representative simulated neuron to show that fractional error measured per neuron can be large. Right: the “ground truth” response matrix for this neuron plotted along with 1 example matrix measured from a simulated experiment that produced an extreme fractional error. D: results of the same analysis presented in A, but performed from responses counted in 2-ms windows. As in A, plot on right shows an enlargement of the boxed region indicated on left.