Figure 4.

Correlation-based coding by integrators and coincidence detectors. A, We measured output correlation ρ in response to different combinations of mean μ and input correlation c (top). Curves on 2-D plots correspond to horizontal cross-sections, at different μ, through 3-D plots (bottom; compare Fig. 1C). Tight clustering of ρ–c curves despite differences in μ is conducive to good correlation-based coding by coincidence detectors (right), which contrasts to the broad distribution of those curves for integrators (left). Equivalent plots for the coincidence detector based on prediction by Equation 29 (inset) shows that the first-order prediction clearly fails to account for output correlation among coincidence detectors. Noticeably, ρ is smaller among coincidence detectors than among integrators. This stems from differences in the CCG, typical examples of which are shown in B based on measurement from pairs of each cell type with the same mean firing rate of 12 Hz; c = 0.3. These measured CCGs confirmed CCG shapes predicted in Figure 2D. Small values of ρ do not necessarily compromise synchrony transfer and correlation-based coding, whereas stimulus-dependent variability in S does (see text). In parts A and B, σ² = 100 pA² for the integrator and σ² = 1000 pA² for the coincidence detector.