Skip to content

Prototype Embedding Map

Visualization artifact — instantiates Prototype-Centered Category Modeling

Projects examples into a spatial map so the category's center, its multiple sub-clusters, and how membership thins toward the edges become visible at a glance.

A Prototype Embedding Map places every example as a point in a low-dimensional space — via an embedding plus dimensionality reduction — so that similar things sit close together and the shape of the category becomes something you can look at. Its distinctive move is that it neither scores, ranks, nor classifies; it renders. Where a table of grades or a rule flattens a category into a list or a line, the map lets a human see structure directly: where the center is, whether the category has one prototype or several, and how density thins from a dense core out to a diffuse fringe. It turns "this category feels lumpy and has a soft edge" into a picture you can point at.

Example

A museum curator is assembling a teaching collection for "Impressionism." She embeds a few thousand painting images with a vision model and projects them to two dimensions. The map is not one blob but several dense knots — a Monet-like landscape cluster, a Degas-like figures cluster — joined by a sparse bridge of hybrid works, and shading off into a diffuse fringe that blends toward Post-Impressionism. Seeing this, she recognizes "Impressionism" as a genuinely multi-prototype category rather than one with a single exemplar. She picks a canonical painting from each dense knot for the wall text, and marks the diffuse fringe "review individually" rather than forcing each borderline canvas to a yes-or-no. The map changed the category from a definition she was defending into a structure she could teach.

How it works

  • Embed, reduce, plot. Represent each example as a vector, project it to two or three dimensions with a method like t-SNE or UMAP, and render it; nearness on the plot encodes similarity.
  • Read the density, not just the dots. Dense knots are prototypes; the densest core is the category's center; thinning density outward is graded membership fading toward the fringe; empty gaps are boundaries.
  • Overlay to interrogate. Colour points by label, source, date — or by a demographic attribute — to check whether the clusters track the intended category or something illegitimate.

Tuning parameters

  • Embedding model / feature space — what "similar" means. The entire map is downstream of this choice, and a different embedding is a different category.
  • Reduction method and hyperparameters — t-SNE perplexity or UMAP neighbor count; these change apparent cluster tightness and separation, and can manufacture or dissolve clusters that are not really there.
  • Colouring and overlays — which attribute is painted onto the points, turning the same map into a structure lens, a drift lens, or a bias lens.
  • Granularity / zoom — whether you read global structure or drill into one local neighborhood.
  • Static versus interactive — a printed snapshot versus an explorable view where hovering a point reveals the underlying exemplar.

When it helps, and when it misleads

Its strength is that it reveals multi-prototype and radial structure that flat rules and tables flatten away; it makes the center visible so it can be taught and audited, and density gives an intuitive read of graded membership. Overlays can expose that a category is secretly tracking an illegitimate attribute. Its central hazard is that dimensionality reduction lies convincingly: cluster sizes, the distances between clusters, and even whether clusters exist at all can be artifacts of the projection's settings rather than facts about the data.[1] The classic misuse is reading a clean two-dimensional separation as proof the category is clean — or, worse, tuning the projection until the clusters you hoped to find appear. The discipline that keeps it honest is to vary the hyperparameters and keep only structure that is stable across them, to treat the map as a hypothesis generator rather than evidence, and to validate any cluster against held-out labels before acting on it.

How it implements the components

Prototype Embedding Map fills the components a spatial visualization can make legible:

  • multi_prototype_structure — dense knots on the map reveal whether the category has a single prototype or several sub-centers, and where they sit relative to each other.
  • radial_extension_map — density thinning outward from each center is the radial fade from prototype to fringe, made visible as distance from the core.
  • category_center_visibility_layer — its core purpose: rendering the category's center(s) as a place a human can literally point to.

It does not produce the numeric graded scale or the ranked anchors — that is the Typicality Rating Exercise — and it does not classify or route live cases, which is the Nearest-Neighbor or Exemplar Classifier operating in the same space.

Notes

The map and the Nearest-Neighbor or Exemplar Classifier often run on the same embedding; when they do, the map is the human-readable face of the very space the machine computes in, which keeps what people see aligned with how cases are actually classified. Run them on different embeddings and that alignment silently breaks.

References

[1] t-SNE and UMAP are nonlinear dimensionality-reduction methods that preserve local neighborhoods but distort global geometry. Apparent cluster sizes and between-cluster distances are therefore not reliable, and clusters can appear or vanish as the perplexity or neighbor-count setting changes — a well-known reason to read such maps as suggestive, not conclusive. ↩