Zirui Chen¹, Michael Bonner¹; ¹Johns Hopkins University

What are the axes along which visual systems learn and represent information? Studies using deep neural networks have debated whether sensory representations are best understood by focusing on individual neurons or latent dimensions of population activity. We approach this debate from the perspective of learning dynamics: which representational basis reveals the structure of how features develop during training? We trained ResNet-50 models on ImageNet and tracked the developmental trajectories of individual neurons versus principal components (PCs). We defined PCs from the fully trained network and projected activations from all training epochs onto this fixed PC basis to assess when each component emerges. We found clear hierarchical learning dynamics: lower-rank PCs with more variance are almost fully developed after just one epoch, while higher-rank PCs converge progressively throughout training, with some not stabilizing until the final epochs. This rank-dependent trajectory confirms theoretical predictions and reveals a natural learning process where the most important features emerge first. In contrast, individual neurons show no such differentiation in their developmental trajectories. There is no clear distinction between early-developed versus late-developed neurons. Critically, the hierarchical PC learning pattern holds robustly even when we incorporate spatial topographic constraints that induce distinct functional clustering in the fully trained network. These findings demonstrate that PCs, not individual neurons, constitute the natural basis for understanding visual feature learning and representation, suggesting that population-level latent dimensions should be the primary unit of analysis for interpreting neural codes.