Extended Data Fig. 4: VEIs generalized across different conditions.
From: Functional bipartite invariance in mouse primary visual cortex receptive fields
a, VEIs synthesized under different conditions remained highly specific to their target neurons. The confusion matrices showed the representational similarity between the original VEIs and VEIs generated from (1) different image initializations for VEIs synthesis, (2) different model initializations, (3) different image and model initializations, (4,5) synthesis using a single model from the ensemble, (6) a different diversity metric based on in silico population responses, (7,8) a different model architecture14 with two different model seeds, (9,10) a different synthesis method using implicit 1D and 2D periodic latent space neural representation model and contrastive regularization21. 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 correction). b, VEIs synthesized from different conditions were more similar to original VEIs than random neurons’ VEIs (two-sided Wilcoxon signed-rank test, W = 0, 0, 1, 1, 34, 0, 0, 0, 0, 426, and 357, with P = 1.2 × 10−17, 1.2 × 10−17, 1.3 × 10−17, 1.3 × 10−17, 3.5 × 10−17, 1.2 × 10−17, 1.2 × 10−17, 1.2 × 10−17, 1.2 × 10−17, 2.2 × 10−17, and 3.7 × 10−17, respectively, after BH correction). c, The mean Euclidean pairwise distances between VEIs generated using different methods were strongly correlated with those of the original VEIs (Pearson r = 0.89, 0.56, 0.60, 0.43, 0.54, 0.45, 0.18, 0.08, 0.62 and 0.66, with P = 1.5 × 10−34, 2.5 × 10−9, 2.2 × 10−10, 1.3 × 10−5, 1.4 × 10−8, 3.5 × 10−6, 0.07, 0.41, 1.4 × 10−11, and 1.2 × 10−13, respectively, two-sided t-test). Data were pooled over 97 neurons randomly sampled from eight mice.
