Fig. 2: VEIs evoked strong and selective responses in target neurons while exhibiting population-decodable differences.
From: Functional bipartite invariance in mouse primary visual cortex receptive fields
a, Examples of MEI and VEIs for simulated simple and complex cells, and mouse V1 neurons. For each neuron, zero-crossing contours from individual VEIs (locations where the image intensity transitions from positive to negative values or vice versa) were overlaid. b, Diversity indices for 60 simulated complex cells (red), 60 simulated simple cells (blue) and 10,228 V1 neurons pooled from 14 mice (gray), including 500 tested in closed-loop experiments from eight mice (unfilled). Diversity index is defined as the normalized average pairwise Euclidean distance in pixel space across the VEIs. Diversity indices for noiseless simple cells (0, blue dashed) and complex cells (1, red dashed) were shown for reference. V1 neuron diversity indices differed from simulated simple and complex cells (P = 3.1 × 10−49 and 1.2 × 10−67, two-sided Welch’s t-test with 72.4 and 69.0 d.f., respectively). For closed-loop experiments, we randomly selected neurons with high diversity indices (Methods). Example neurons from a were indicated on the x axis with the corresponding colors. Diversity indices <−0.25 were clipped to −0.25 for visualization (0.09% of all V1 neurons). c, Response of an example neuron to its MEI and ten random VEIs. Both MEI and individual VEI were averaged across 20 repeats. Only two out of the ten VEIs elicited responses lower than 85% of the MEI response (one after Benjamini–Hochberg (BH) correction for multiple comparison). d, Comparison of mean responses to MEI and one random VEI per neuron. VEIs stimulated in vivo responses in target neurons close to the level predicted in silico relative to MEI (74 ± 4% versus 85%) (two-sided Wilcoxon signed-rank test, W = 4,902, P = 0.19) with only 274 of 1,490 VEIs (18.4%) showing responses lower than 85% of the corresponding MEI response (3.0% after BH correction) (P < 0.05, one-sided Welch’s t-test with 32.6 average d.f.). Data were pooled over 149 neurons from two mice. e-h, VEI responses were averaged across 20 different VEIs with each presented once. e, Both MEI and VEIs activated neurons with high specificity. Confusion matrices showed responses of each neuron to MEI (left) and VEIs (right) for 61 neurons in one mouse. Responses of each neuron were normalized, with each row scaled so the maximum response across all images equaled 1. Neurons’ responses to their own MEI and VEIs (along the diagonal) were larger than those to other MEIs and VEIs, respectively (two-sided permutation test, P < 10−4 for both cases). f, Predicted versus observed responses of one example neuron to its own MEI and VEIs and 79 other neurons’ MEI and VEIs. g, Our model exhibited high predictive accuracy for both MEI and VEI responses (Pearson correlation coefficient between predicted and observed neuronal responses r = 0.74 and 0.75, respectively). h, VEIs stimulated in vivo responses close to the level predicted in silico relative to MEI (75 ± 3% versus 85%) (two-sided Wilcoxon signed-rank test, W = 51,360, P = 4.9 × 10−4), with only 9.6% of all neurons showing different responses between VEIs and 85% of MEI (1.2% after BH corrections) (P < 0.05, two-sided Welch’s t-test with 34.06 average d.f.). g,h, Data were pooled over 500 neurons from eight mice. i, In vivo population responses in mouse V1 L2/3 discriminated between a randomly selected pair of VEIs for each neuron. VEI identity in individual trials was decoded using a logistic regression classifier (see Methods for details), with decoding accuracies across neurons (median 80%) exceeded chance level (50%, dashed; one-sample t-test, t = 28.0, P = 5.0 × 10−61). Data were pooled over 149 neurons from two mice.
