Extended Data Fig. 4: Extended visualization of neuronal dynamics analysis during evolution.
From: Neuronal tuning aligns dynamically with object and texture manifolds across the visual hierarchy
A. Schematics of the temporal attribution analysis, activation difference between PSTHs of the first and max activation block is normalized and attributed to each time bin. B. Difference in temporal attribution strength between DeePSim- and BigGAN evolution threads computed in time bins ranging from 5 ms to 50 ms, threads with significant activation increase were included, non-paired. Error bands represent mean ± SEM across evolution threads. Independent t-test (two-sided) were applied separately at each time bin, with false discovery rate (FDR) correction applied to the raw p values across all time bins to control for multiple comparison. Statistical significance was defined at FDR corrected p < 0.05. Dots on top row indicate significant comparisons per FDR, dots on lower row indicate significance per original p < 0.05; Red dots indicate time bins where BigGAN > DeePSim, and blue dots indicate DeePSim > BigGAN. Exact P values for significant bins ranged from p = 0.0037 to p = 0.042 (the full sets of p values can be reproduced via released code and data). Sample size (number of successful evolution threads, p<0.01) was DeePSim: V1, N = 10; V4, N = 37; IT, N = 67; BigGAN: V1, N = 0; V4, N = 20; IT, N = 65. C. Normalized evolution trajectories for specific time windows using time bins of 10, 20, 25, and 50 ms (similar to Fig. 5c), same normalizer was applied for each paired evolution across the two threads and across block and time windows. The line and shaded error bar indicates mean ± SEM across threads. At each block, activations from DeePSim and BigGAN evolution trajectories were compared using a paired t test (two-sided). Raw p values from each block underwent False Discovery Rate (FDR) correction, and statistical significance was determined at corrected p < 0.05, and significant blocks are indicated by colored dots (the full sets of p values can be reproduced via released code and data). Sample size was N=24 neuronal sites in V4, N=61 in IT, for all analyses.
