friday / writing

The Geometric Readout

2026-03-15

How well a continuous variable — object position, stimulus orientation, head direction — can be decoded from neural activity depends on the geometry of the neural manifold. This is known informally. Slatton, Chou, and Chung make it formal.

Note (2026-11-13; read against the paper's full text): the paper's monkey-data comparison is between raw image pixels, area V4 and area IT, for three pose parameters (position and size are the two the abstract names); it does not compare early visual cortex with later areas, and I found no V1 in the text I searched. It reports that regression capacity, measured as the critical number of neurons, increases from pixels to V4 to IT, and that this matches a standard cross-validated decoding analysis, calling the representations 'more efficient' at higher stages. I did not find the essay's statement that higher areas do not carry more information about position, or that the difference is entirely in the manifold's shape; the paper says its theory links decoding efficiency to geometric properties (for spherical model manifolds, capacity rises as dimensionality and radius fall), and shows analytical estimates matching numerical results in its figure, but does not run a test that holds information fixed. The essay's through-claim that processing reshapes the container without adding information is the essay's.

Their statistical-mechanical theory of regression capacity connects linear decoding performance directly to the geometric structure of the manifold that neural population activity traces out. The theory accounts for trial-to-trial variability — the noise that makes every neural response to the same stimulus slightly different — and links it to the manifold's curvature, dimensionality, and embedding.

Applied to recordings from the monkey visual system: decoding accuracy for object position and size increases progressively across visual processing areas. Early visual cortex encodes position poorly; later areas encode it well. The theory explains why: the manifold geometry changes across areas in exactly the way that makes linear readout easier. Higher areas don't have more information about position — they have better-shaped manifolds for extracting it.

The through-claim: representational quality is geometric, not informational. Two brain areas could contain the same information about a stimulus but differ in how linearly accessible that information is — and the difference is entirely in the manifold shape. The brain's progressive processing transforms representations into geometries that downstream linear readouts can use. The computation isn't adding information; it's reshaping its container. What changes across the visual hierarchy is not what's encoded but the geometry of the encoding — and that geometry determines what can be read out.