Fig. 3: VEIs evoked stronger responses than controls and generalized across different synthesis conditions.
From: Functional bipartite invariance in mouse primary visual cortex receptive fields
a, MEI, VEIs (top), synthesized controls (middle) and natural controls (bottom) for one example neuron. Synthesized controls were generated by perturbing MEI in random directions, while natural controls were selected by searching through random natural patches. For each neuron, both controls were restricted to be closer to the MEI than all the VEIs as measured by Euclidean distance in pixel space. b, Synthesized controls failed to stimulate their target neurons in vivo compared to VEIs (55 ± 2% of VEI activation, two-sided Wilcoxon signed-rank test, W = 3,258, P = 2.3 × 10−41), with 36.8% neurons showing lower responses to synthesized controls compared to VEIs (20.8% after BH corrections; P < 0.05, two-sided Welch’s t-test with 30.4 average d.f.). c, Natural controls failed to stimulate their target neurons in vivo compared to VEIs (63 ± 3% of VEI activation, two-sided Wilcoxon signed-rank test, W = 6,442, P = 9.4 × 10−31), with 31.5% neurons showing lower responses to natural controls compared to VEIs (16.0% after BH corrections; P < 0.05, two-sided Welch’s t-test with 31.1 average d.f.). b,c Response to each stimulus type was averaged over 20 different images with single repeat. Data were pooled from 318 neurons across five mice. d, MEI and VEIs for the same neuron in a, synthesized under various conditions: (1) different image initialization, (2) different model initialization, (3) different diversity metric, (4) different synthesis method21, and (5) different model architecture14. e, VEIs synthesized under different conditions maintained high specificity to their target neuron. Confusion matrices showed in silico representational similarity between original VEIs and VEIs from different image initialization (left) or VEIs from a different synthesis method21 (right) (for other conditions, see Extended Data Fig. 4a). Each entry represents the mean pairwise cosine similarity between two sets of VEIs (Methods). Representational similarity between original VEIs and VEIs synthesized from different conditions for the same neurons (diagonal) was larger than cross-neuron similarity (off-diagonal) (two-sided permutation test, P < 10−4 for all conditions after BH corrections). f, VEIs synthesized under different conditions closely resembled the original VEIs. The original VEIs were more similar to VEIs generated from various modifications in d than random neurons’ VEIs generated using the original method (two-sided Wilcoxon signed-rank test, W = 0, 0, 1, 0, 0, 426, P = 1.2 × 10−17, 1.2 × 10−17, 1.3 × 10−17, 1.2 × 10−17, 1.2 × 10−17, and 2.2 × 10−17, respectively, after BH correction). Box plots show center line (median); box bounds (25th to 75th percentiles, IQR); whiskers extend to the most extreme data points within 1.5 × IQR of the quartiles; caps mark whisker ends; points beyond whiskers are plotted as outliers. e,f, Data were pooled from 97 neurons randomly sampled across eight mice.
