Marcelo Mattar

Compact neural networks that explain the visual cortex with 5,000× fewer parameters Deep neural networks are the dominant tool for modeling the brain's visual system—feed in an image, predict neural firing rates. Models like ResNet50 predict responses in visual area V4 (a mid-level region selective for shapes, curves, textures, and colors) with reasonable accuracy. But with tens of millions of parameters, understanding what computation they've actually learned is nearly impossible. The field has been stuck: large models predict well, but large models are opaque. Benjamin Cowley and coauthors challenge this directly. They first build a highly predictive hybrid model of macaque V4—combining a task-driven backbone with a data-driven ensemble trained in closed-loop experiments, where the model selects which images to show next based on its own uncertainty. This active learning strategy across 44 sessions and 78,000+ images yields a noise-corrected R² of 0.62, roughly 30% above the best standalone models. Then the key move: they compress this 60-million-parameter black box into compact models of ~10,000 parameters using knowledge distillation followed by channel-wise pruning. The resulting models, with ~150 filters across 5 layers, lose only ~5% in accuracy while being 5,000× smaller—and their filters can all be visualized in a single diagram. Crucially, these predictions are validated in vivo: synthesized and selected maximizing images drive real V4 neurons far above baseline, and subtle image perturbations shift responses exactly as predicted. The most striking insight is structural. Early layers share similar representations across neurons, acting as a common feature basis. Then, between layers 3 and 4, filter count drops sharply—a "consolidation step" where each neuron specializes uniquely. This is demonstrated concretely in a dot-detecting neuron: four excitatory corner filters and two inhibitory edge filters interact at consolidation to produce size-selective dot detection—a testable circuit hypothesis for real V4 neurons. Beyond V4, models of V1 (primary visual cortex, the earliest cortical stage of visual processing) and IT (inferior temporal cortex, a higher-order region encoding complex objects and faces) are also compressible—V1 needing just 5 shared filters, V4 ~10, IT ~60—tracking the known hierarchy of feature complexity along the ventral stream. The message: systematic compression is not just an engineering trick but a scientific tool that strips away the redundant and reveals essential computations, bridging prediction and understanding. Paper: nature.com/articles/s41586-0…