Fig. 5: Effects of coupling.
a Differences in reconstruction quality between the coupled model (LNBRC) and uncoupled model (LNBR) for flashed natural images. Mean differences for four preparations: 0.023, 0.024, 0.037, and 0.023 (p-values 3.7 × 10−184, 1.2 × 10−211, 8.6 × 10−122, and 8.5 × 10−111, respectively, Wilcoxon one-sided ranked sign test). For panels (a-d), the box marks the median and the inter-quartile range (IQR), while the whiskers extend to 1.5 times the IQR. Outliers are marked with a +. b Same as (a), for reconstruction with fixational drift using the joint approach. Mean differences for three preparations: 0.019, 0.010, and 0.039 (p-values 3.2 × 10−218, 4.2 × 10−62, and 2.5 × 10−307, respectively). The blue and green boxes correspond to the same experimental preparations as the blue and green boxes in (a). c Differences in reconstruction quality between the unshuffled and shuffled trials for flashed image reconstructions, using LNBRCs fitted to unshuffled data. Mean differences: 7.4 × 10−3, 6.6 × 10−3, 5.1 × 10−3, and 3.4 × 10−3 (p-values 4.5 × 10−23, 5.6 × 10−12, 3.1 × 10−20, and 3.4 × 10−12, respectively), substantially smaller than those in (a). (d) Same as (c), for reconstruction with fixational drift using the joint approach. Mean differences (left-to-right): 5.9 × 10−5, 9.5 × 10−4, and 7.6 × 10−3 (p-values 0.45, 0.017, and 9.1 × 10−16), substantially smaller than those in (b). e Example homotypic (same cell type) nearest-neighbor spike train cross correlograms. Panels e–h are computed using the blue experimental preparation in panels (a–d) using repeat presentations of jittered natural image stimuli. Cross-correlograms for the data are shown in black, and repeat-shuffled data in red. Simulated cross-correlograms for the LNBRC (coupled) models and for the LNBR (uncoupled) models are shown in red and green, respectively. f Fraction of PSTH variance explained by the coupled and uncoupled models. g Cross-correlogram peak height comparison between model simulations and data. While the LNBRCs sometimes overestimated the correlations in the data, the LNBRs systematically underestimated them. h Cross-correlogram peak height comparison between repeat data and shuffled repeat data, for the same preparation as (f, g), with the jittered stimulus. Except for the OFF parasol cells, peak heights were similar, indicating that noise correlations were only weakly present.
