TopoNets: High performing vision and language models with brain-like topography.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Biological neurons exhibit functional organization, with nearby cells sharing similar roles.
• Current AI models lack this inherent spatial organization, often sacrificing performance for topographic structure.

Purpose of the Study:

• To develop a method for creating spatially organized topographic representations in AI models.
• To achieve brain-like organization in AI without compromising task performance.

Main Methods:

• Introduction of TopoLoss, a new loss function designed to promote spatial topography in AI.
• Integration of TopoLoss into leading AI architectures including ResNet, ViT, GPT-Neo, and NanoGPT, creating TopoNets.
• Validation across vision and language models.

Main Results:

• TopoNets demonstrate superior performance among supervised topographic models.
• TopoNets exhibit brain-like properties: localized feature processing, reduced dimensionality, and enhanced efficiency.
• TopoNets successfully predict neural responses and replicate topographic signatures in the brain's visual and language cortices.

Conclusions:

• TopoLoss provides a robust and adaptable framework for integrating topography into AI models.
• TopoNets represent a significant advancement in creating high-performing AI that emulates human brain computational strategies.
• This work bridges the gap between biological and artificial systems by creating AI with brain-like topographic organization.