Fig. 2: All-TNNs mirror key features of the visual system’s topography.

a, Left, visualization of orientation selectivity across the first layer of an All-TNN (α = 10), an LCN and a CNN (network instances with random seed 1; see Supplementary Fig. 3 for all seeds). Right, average cluster size of orientation selectivity for all models (the data are averaged across all seeds; the error bars show standard deviation across seeds; n = 5). b, Top, entropy maps as a measure of local variability in feature selectivity for different models (seed 1; see Supplementary Fig. 9 for all seeds). Bottom, entropy decreases more with eccentricity as α increases in All-TNNs. Entropy remains relatively constant for CNNs and LCNs (the data are averaged across all seeds; the shaded regions indicate the 95% CIs). c, Training on ecoset with all images shifted towards the bottom right leads to a corresponding shift of the region with high feature variability in orientation selectivity, indicating that feature variability is linked to task-relevant information in the network input. d, Left, analysis of the last network layer selectivity for high-level visual categories (measured using d′ for tools, scenes and faces; seed 1; see Supplementary Fig. 13 for all seeds; n = 5). Right, average size of category-selective clusters (averaged over scenes, faces and tools) for All-TNNs and control models. All CNN and LCN layers are visualized using the same reshaping procedure as for the creation of All-TNN 2D sheets (Methods).