Compact deep neural network models of the visual cortex

Nature

by Benjamin R. Cowley

February 26, 2026

AI-Generated Deep Dive Summary

A recent study challenges the notion that large deep neural networks (DNNs) are necessary for accurately predicting neural responses in the primate visual cortex. By building compact DNN models with significantly fewer parameters—reducing a 60 million-parameter model by a factor of 5,000 while maintaining comparable accuracy—the researchers revealed key insights into how these simplified models function. Their findings demonstrate that smaller, more efficient models can capture the essential computations of the visual cortex without sacrificing predictive power. The study employed an adaptive closed-loop approach to build and refine their models. Starting with a highly predictive DNN for macaque visual area V4, they compressed it to identify simpler yet effective models. This process not only reduced computational complexity but also allowed them to probe the inner workings of these compact networks. A notable discovery was that early layers in these models shared similar filters, while later layers specialized by "consolidating" this shared representation in unique ways. For instance, a dot-detecting neuron model revealed a specific computational mechanism aligning with known properties of V4 neurons. The researchers extended their findings to other visual areas, including V1 and IT (inferior temporal cortex). The successful compression across these regions suggested a general principle underlying the computations of the visual cortex. This breakthrough challenges the assumption that larger models are inherently better for predicting neural activity and opens new avenues for understanding how vision is processed in the brain. The study’s significance lies in its potential to advance both neuroscience and artificial intelligence. By simplifying DNNs, researchers can better interpret their inner workings, making them more accessible for scientific inquiry. These compact models could also pave the way for more efficient AI systems that mimic biological visual processing while requiring less computational power. The findings underscore the importance of balancing model complexity with explanatory power in understanding neural computations. Overall, this research provides a novel framework for modeling the visual cortex, demonstrating that parsimony and predictive accuracy can coexist. This approach not only deepens our understanding of how vision is processed in the brain but also offers practical insights for developing more efficient AI systems inspired by biological principles.

Verticals

scienceresearch

Originally published on Nature on 2/26/2026