Realistic neuron model boosts ANNs: more expressivity, less data

Since the 1950s, artificial neural networks have used the point neuron model—a simplistic abstraction of biological neurons. However, neuroscience has long shown this model fails to capture fundamental neural processes. In a new arXiv paper, researchers Raul Mohedano and colleagues replace the point neuron with a recently developed cortical cell model that better mimics real neurons. Remarkably, this substitution requires no additional parameters. Through theoretical analysis and experiments, they show that networks using the new neuron achieve higher expressivity, greater robustness, and faster learning. They also exhibit less memorization and need less training data, addressing common overfitting and data efficiency issues.

The work is a significant step toward biologically plausible AI without sacrificing performance. By simply updating the elementary unit of neural networks, the authors demonstrate gains across multiple domains, including computer vision and machine learning. The results suggest that decades-old architectural assumptions may be holding back progress. For practitioners, this offers a direct, parameter-free way to improve model efficiency and generalization. The paper is available on arXiv and has already sparked discussions in the AI community about integrating modern neuroscience insights into deep learning.