The visual system can quickly recognize faces, objects and scenes, yet the exact way single neurons react to images has been hard to study.
A recent study published in Nature shows that much simpler computer models can now capture these neuron responses with high accuracy. Scientists led by Carnegie Mellon University started with a very large model designed to predict activity in the visual cortex, the part of the brain that handles sight. This original model contained millions of parameters. Because it was so large, the model was nearly as difficult to understand as the brain itself. Using machine learning techniques, the scientists compressed the model. The new versions were thousands of times smaller but remained highly accurate.
Insights from simpler models
This approach shows that massive, complicated networks are not needed to explain what individual neurons do. The scientists note that smaller and interpretable models, meaning models whose inner steps can be clearly understood, now give scientists real intuition about vision. They can form ideas that can be tested in experiments.
Even with the dramatic size reduction, the models still captured subtle differences in how neurons react to similar images. This suggests the brain relies on specific computational patterns that can be described in straightforward ways. The simplified models revealed how individual neurons detect key features, such as the eyes in a face or the dots in a pattern.
The findings also point to practical uses beyond basic science. Modern computer vision systems for tasks like face recognition on phones or guiding self-driving cars often struggle in ways humans do not. Insights from these compact models could make such artificial intelligence systems more reliable in everyday situations.
Researchers are now extending the models to handle moving images such as videos. This step could explain how the visual system tracks motion, notices changing patterns and focuses on important details in the world around us. The work combines experimental neuroscience with computational methods and offers clearer rules for how the brain interprets what we see.