Shrunken AI Models Reveal How the Brain’s Visual Neurons Actually Work

Somewhere in the visual cortex of a macaque monkey, a single neuron fires every time a small dot appears in the right location. Not a circle, not a line — a dot, specifically, at a specific size. For decades, neuroscientists could describe this selectivity without really explaining it. Now, for the first time, they can watch it happen in a model small enough to read like a circuit diagram.

The breakthrough came from an unlikely direction: aggressive compression. A team at Carnegie Mellon University and Princeton began with a deep neural network containing 60 million parameters — the kind of sprawling, accurate-but-opaque model that modern neuroscience has increasingly relied on to predict how visual neurons respond to images.

Using machine learning techniques called knowledge distillation and pruning — where a larger “teacher” model trains a smaller “student,” then redundant filters are stripped away iteratively — the researchers squeezed each neuron’s model down to roughly 10,000 parameters. That is 5,000 times smaller, yet the compact models predicted neural responses with accuracy that still outperformed every standard task-driven architecture tested. The key insight was that most of those 60 million parameters were simply redundant.

The results suggest that individual neurons are doing something far more tractable than the field assumed. Individual V4 neurons, it turns out, can be captured by fewer than 200 filters.

What those filters reveal is striking. Early layers across all the compact models looked nearly identical — a shared vocabulary of edges, curves and colour contrasts. Then, at a sharp transition between layers three and four, something the team calls the “consolidation step,” each model suddenly diverged. The later layers were unique to each neuron, the point at which a generic visual processor becomes a dot detector, a curve detector, or a texture specialist.

Working through the dot-detecting model filter by filter, the team found that dot-size selectivity emerges from just six components. Four detect corner-like curvatures at the edges of a dot; two inhibit responses to large continuous edges. A small dot activates all four excitatory filters simultaneously and triggers minimal inhibition. A large dot spreads the excitatory activity too far and trips the inhibitory ones. The response collapses.

“This work shows that we don’t need massive, complicated networks to understand what individual neurons are doing,” said Matthew Smith, professor of biomedical engineering at Carnegie Mellon. “By making the models smaller and interpretable, we can actually gain intuition about how the visual system works and develop hypotheses that can be tested in the lab.” That testability matters: the corner-and-edge mechanism is now a specific, concrete prediction about which V1 and V2 circuits feed V4 dot detectors — something anatomical tracing and targeted stimulation experiments can directly probe.

The approach generalised across the visual hierarchy. Models of V1 neurons compressed furthest, requiring as few as five shared filters. V4 needed around ten. Neurons in inferior temporal cortex — the brain’s object-recognition territory — needed sixty, consistent with the greater complexity of what they encode.

The models weren’t just theoretically compact. When the team presented macaques with images that the compact models predicted would maximally drive specific neurons, those cells fired above 98 per cent of responses to randomly chosen pictures.

“By working together across institutions and disciplines, we were able to build models that are not only predictive, but also interpretable and meaningful,” Smith said. The work also points toward a practical limit: compression works for individual neurons and small populations, but whether it scales to the full diversity of millions of cortical cells remains an open question.

What the study changes is the expectation. The visual cortex is not necessarily an impenetrable black box requiring impenetrable black-box models to describe it. A dot detector, at least, fits in a diagram. If enough neurons do, the Hubel-and-Wiesel tradition of explaining vision through simple, testable circuits may have a longer future than the deep learning era suggested.

Study link: https://www.nature.com/articles/s41586-026-10150-1