Definition
Plain language
A smooth, often curved surface that a set of related data points all sit on, even when each individual point is a long list of numbers.
As stated in the literature
A lower-dimensional surface embedded in a high-dimensional representation space on which meaningful data concentrates; two systems are said to share a manifold when their internal representations live in compatible geometry, as when a predicted output and its target occupy the same encoder feature space.
Why it matters: Recognizing that meaningful data sits on a lower-dimensional surface lets systems compare and connect their internal representations instead of drowning in raw numbers.
For example, photos of a face turning slowly from left to right form a smooth curved path through data space, even though each photo is millions of pixels.
Heard on the show
“Most of the head's apparent richness is collapsing onto a tiny manifold.”Episode 038 — How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial