New hyperbolic geometry approach bridges the gap between brain signals and visual perception in AI models

Researchers propose HyFI (Hyperbolic Feature Interpolation), a framework that uses hyperbolic space to align neural brain activity with visual features from pre-trained vision models

The method addresses two key challenges: the modality gap between brain signals and image representations, and the entanglement of semantic and perceptual features in neural activity

Hyperbolic space is leveraged for its geometric properties that naturally account for differences in information density and representational capacity between modalities

This work advances brain-vision alignment research by moving beyond previous approaches that independently aligned neural activity with individual feature types