Thursday 27 February 2025
Neural networks and human brains may seem like vastly different systems, but they both encode information in complex patterns. A new study has found a way to analyze these patterns by using a technique called Ollivier-Ricci curvature to compare the structures of neural representations across humans and artificial intelligence.
Researchers used this method to examine how well a deep learning model aligned with human judgments about face similarity. They created a custom version of VGG-Face, a popular facial recognition algorithm, that was trained on human similarity judgments. By analyzing the geometry of the data, they found that the aligned model’s representations were more similar to those of humans than an unaligned version.
The study used a technique called Ricci flow, which is inspired by the way heat flows through a material. In this case, it was applied to the neural networks to identify community structures within the data. These communities corresponded to clusters of similar faces, and the aligned model’s representations were more cohesive within these groups.
This work has implications for understanding how artificial intelligence can learn from human behavior. By analyzing the geometry of neural representations, researchers may be able to better understand how AI systems make decisions and improve their performance on complex tasks.
The study also highlights the importance of considering the intrinsic geometry of high-dimensional data when analyzing neural networks. This is a critical issue in machine learning, as many algorithms rely on simplifying assumptions about the structure of this data.
In the future, this research could be used to develop more advanced AI systems that are better aligned with human cognition. By understanding how humans and machines encode information in complex patterns, researchers may be able to create more effective and efficient AI systems.
Cite this article: “Unraveling the Geometry of Neural Networks and Human Brains”, The Science Archive, 2025.
Neural Networks, Human Brains, Artificial Intelligence, Ollivier-Ricci Curvature, Deep Learning Model, Facial Recognition Algorithm, Ricci Flow, Community Structures, High-Dimensional Data, Machine Learning.
