Summary: Understanding the human brain often means building ever-larger artificial intelligence systems that run on banks of supercomputers. New research takes the opposite approach: by dramatically shrinking powerful AI models, scientists produced a compact vision model that not only fits on a consumer device or in an email attachment but also predicts neural responses in the macaque visual cortex more accurately than much larger systems.
Researchers reduced state-of-the-art AI models to roughly one one-thousandth of their original size. The result is a highly efficient, easy-to-inspect model that revealed previously hidden structure in primate visual processing: specialist neurons tuned to features such as dots. These insights provide a clearer map for how brains encode edges, colors and specific shapes, and suggest new avenues for therapeutic visual stimulation.
Key Facts
- Compression breakthrough: A large AI model trained on macaque neural responses was compressed by 1,000× while preserving predictive power.
- Superior prediction: The compact model outperformed existing top vision models by more than 30% in predicting single-neuron activity in visual cortex.
- Specialist neurons discovered: The compact model identified groups of V4 neurons that act as highly selective “dot detectors,” tuned to small, point-like features that are critical for recognizing eyes and other facial details.
- Transparent computations: Because the model is small, researchers could inspect internal units to see how images are decomposed into edges, colors and shapes and how those components combine into higher-order preferences.
- Therapeutic potential: Mapping the exact visual features that drive neuronal interactions could enable targeted visual therapies, for example to preserve or restore synaptic connections affected by neurodegenerative disease.
Source: CSHL
Computer scientists commonly build ever-larger AI systems to approximate human-like intelligence. For neuroscientists, however, such scale creates another opaque “black box” that replaces one complex system—the brain—with another. That makes it difficult to discover how biological neurons compute visual information.
Cold Spring Harbor Laboratory Assistant Professor Benjamin Cowley, together with Carnegie Mellon Professor Matthew Smith and Princeton Professor Jonathan Pillow, pursued a different path: build predictive models that are not only accurate but also parsimonious and interpretable. After earlier work modeling neural responses in fruit flies, the team focused on macaque monkeys, whose visual systems closely resemble our own.
In experiments reported in Nature, the team presented macaques with carefully selected natural images while recording activity from neurons in the visual cortex. They first trained large deep neural networks (DNNs) to predict which images would drive which neurons, achieving prediction accuracy more than 30% better than prior models. Critically, they then applied compression techniques to derive compact models with roughly 1/1,000 the parameters of the original networks—small enough to email.
Finding that such tiny models can match or exceed larger systems is significant for two reasons. First, it demonstrates that highly predictive models of individual neurons do not always require massive parameter counts. Second, the reduced complexity makes mechanistic analysis possible. The compact networks showed a repeated computational motif: early layers share a common bank of filters that represent basic visual features, while later processing units “consolidate” those shared features in differing ways to create distinct, specialized selectivities.
One clear example is the discovery of dot-selective V4 units. These neurons respond strongly to small, localized point-like features—effectively detecting dots. That specialization makes functional sense: eyes, pupils and many facial marks are compact features that convey essential social and behavioral information. Dot-selective neurons help primates rapidly detect and follow gaze and facial cues.
Beyond basic science, the work points toward clinical possibilities. Cowley and colleagues speculate that, by identifying the exact images and features that drive specific neural pathways, researchers might design visual stimulation protocols to encourage synaptic activity in circuits weakened by conditions such as Alzheimer’s disease. While still speculative, this line of research frames a testable, image-based approach to preserving or restoring neural function.
Key Questions Answered:
A: Large AI models often act as black boxes; they predict well but offer little insight into how decisions are made. Smaller, compressed models retain predictive accuracy while revealing the computations and feature combinations that drive individual neurons. This transparency turns an opaque predictor into a mechanistic explanation.
A: Dot-selective neurons are well-suited to detect compact, information-rich features such as pupils and small facial marks. These detections support rapid gaze-following, face recognition and social communication—core elements of primate social behavior.
A: It is a promising hypothesis rather than an established therapy. If researchers can reliably map which images activate the neural pathways that connect particular neurons, tailored visual stimulation might help maintain or restore synaptic activity. Controlled experiments will be required to evaluate efficacy and safety.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- The referenced journal paper was reviewed in full by editorial staff.
- Additional context and explanation were added by the reporting team to clarify implications for neuroscience and potential clinical applications.
About this AI and visual neuroscience research news
Author: Samuel Diamond
Source: CSHL
Contact: Samuel Diamond – CSHL
Image credit: Cowley lab / CSHL
Original research: Closed access. “Compact deep neural network models of the visual cortex” by Benjamin R. Cowley, Patricia L. Stan, Jonathan W. Pillow & Matthew A. Smith. Nature. DOI: 10.1038/s41586-026-10150-1
Abstract
Compact deep neural network models of the visual cortex
Predicting neural responses to arbitrary images provides a powerful window into how the visual cortex computes. Deep neural networks (DNNs) have become the leading predictive models, but their internal computations are often hidden beneath millions of parameters. This study challenges the presumption that large scale is necessary by developing predictive yet parsimonious DNNs for the primate visual cortex.
The team constructed a highly predictive DNN for macaque visual area V4 through iterative data collection and adaptive, closed-loop model training. They then compressed the original 60-million-parameter model to identify compact networks with thousands of times fewer parameters while maintaining comparable accuracy. The compression allowed detailed inspection of model internals.
A consistent motif emerged: compact models share similar early filters but create specialized feature selectivity by consolidating the shared high-dimensional representation in distinct ways. Analysis of a dot-selective model neuron suggested a mechanistic circuit hypothesis for dot-selective V4 neurons. Compression success extended beyond V4 to V1 and inferior temporal cortex (IT), indicating a general computational principle across visual areas.
Overall, the work argues that highly predictive models of individual neurons need not be massive black boxes and establishes a framework that balances prediction with interpretability and parsimony.