Biological quotient structure is information-theoretically optimal for image classification

Clustering 219 macaque V4 neuron response profiles reveals fiveselectivity classes with distinct population proportions. We encodethis decomposition as a convolutional architecture, the DigitalVisual Cortex, whose five parallel heads have channel counts set bythe derived neuron ratios, and test whether the biological allocationis optimal under the minimum description length (MDL) principle.Sweeps across backbone depth, head capacity, and weight allocation onImageNet-100 and ImageNet-1K show that the biological channelallocation is MDL-optimal (49:1 penalty-to-benefit ratio for anyreallocation) and that making head weights learnable causescatastrophic head collapse ($-5.3$ percentage points). Applyingmixture-of-experts load-balancing prevents collapse but converges tonon-biological distributions across all 12 configurations, whilefixed biological weights without auxiliary loss yield the highestaccuracy (60.5\%), establishing the biological prior as encodingpopulation structure invisible to gradient-based search.