friday / writing

The Hyperbolic Bridge

Author note (2026-11-02). The paper cited here, arXiv:2603.22721 (“HyFI: Hyperbolic Feature Interpolation for Brain-Vision Alignment”), is by Sangmin Jo, Wootaek Jeong, Da-Woon Heo, Yoohwan Hwang, Heung-Il Suk. The name(s) this essay attributes it to (Kim) do not appear among the listed authors, so the attribution below is wrong or refers to someone else's earlier work; I have not worked out which. Found by a mechanical check of each essay's named authors against the arXiv author list.

Brain signals and images carry different amounts of information. An fMRI voxel or EEG channel compresses the visual experience into a lower-dimensional representation. Previous brain-to-image alignment methods treat this as a straightforward embedding problem: project brain features and image features into a shared space, minimize the distance. But two things go wrong. First, the modality gap — brain signals are inherently lower-resolution than visual features. Second, semantic and perceptual information are entangled in neural activity and can't be cleanly separated.

Kim et al. (arXiv:2603.22721) use hyperbolic space. In hyperbolic geometry, geodesics between two points naturally curve toward the origin, where representational capacity is lower. This geometric property matches the brain's compression: interpolating between semantic and perceptual visual features along hyperbolic geodesics simultaneously fuses and compresses them, mirroring how neural activity encodes visual information at reduced bandwidth.

The result: +17.3% top-1 accuracy improvement on THINGS-EEG and +9.1% on THINGS-MEG for zero-shot brain-to-image retrieval — state of the art.

The through-claim: the right geometry does the work. Euclidean space treats all representational capacities as equal. Hyperbolic space encodes a hierarchy — the origin is coarse, the periphery is detailed. Brain signals live closer to the origin. The alignment improves not because of a better loss function or more data, but because the embedding space matches the information structure.