A Unified Geometric Space for Topological Alignment Between Transformer-Based Models and Human Brain Networks
Researchers have developed a novel method to compare the organizational properties of transformer-based AI models by mapping their attention topologies to human brain networks. This approach allows for modality-agnostic and task-free analysis across vision, language, and multimodal systems. Their study of 151 models revealed a continuous arc-shaped distribution of topological alignment, with models focused on global abstraction aligning more with higher-order brain networks and local detail-focused models aligning with lower-level networks. Unexpected findings included reduced alignment in DINOv2 and a scaling inversion in distilled DeiT models, suggesting complex relationships between model architecture, training, and brain-like organization. AI
IMPACT Provides a new quantitative lens for comparing AI model architectures and their emergent organizational properties, potentially guiding future model development.