Skip to content
CrowLingo

Methods · Sub-page

Latent space 101.

A practical guide to representations, distances and the limits of a two-dimensional map.

Methods · Latent space 101

A latent space is a vector representation learned by a model. Researchers can inspect relationships with a two-dimensional projection, but projection can distort distances. CrowLingo’s atlas uses seeded illustrative coordinates, not measured embeddings. No similarity search currently runs on this corpus.

Audio narration is unavailable while its source claims are corrected.

An embedding is a list of numbers

An is a learned numerical representation of an input. Its size and the audio preparation it expects depend on the model, version and selected output. BirdNET's documentation lists 1,024 dimensions for V2.4 and 1,280 for its V3.0 preview; those are examples, not a universal size or a specification for every crow recording.

The space they live in

A contains these representations. Researchers can compare vectors with a distance measure and investigate which recordings appear similar. A nearby pair does not by itself establish the same caller, call type or behavioral meaning. Those interpretations need separate labels, observations and task-specific evaluation.

Why project to two dimensions?

is a dimensionality reduction method often used to display high-dimensional data in a two-dimensional map. A projection changes the representation: its distances and apparent groups should not be treated as direct measurements of biological categories.

The UMAP authors also describe reduction as preprocessing for clustering. Analyses may use original or reduced representations; disclose which one you use, the distance measure and the settings, then evaluate whether the result supports the intended task.

Settings change what you see

n_neighbors affects the balance between local and broader structure. min_dist affects how tightly points may pack in the projected space, and the metric determines how input distances are calculated. There is no single setting that establishes the meaning of a crow call. Check how conclusions change across justified settings instead of selecting only the most convincing-looking map.

Reading CrowLingo's Atlas

CrowLingo's current Atlas uses seeded illustrative coordinates, not model embeddings or a UMAP result. Its groups are not categories discovered by clustering. Listen to the recordings and inspect their source descriptions, but do not read distance as a measured acoustic relationship. Model-backed similarity search and interactive UMAP settings are not current features or scheduled releases.

Sources and scope

Documentation checked November 11, 2026. These explanations do not report a measured CrowLingo embedding experiment.

Frequently asked

What people ask about this.

What is a latent space?

A latent space contains learned representations of inputs. Its dimensions and the interpretation of distances depend on the model, output and distance measure. UMAP can project representations for inspection, but nearby points do not establish a shared caller or meaning. CrowLingo uses illustrative coordinates rather than measured embeddings.

Should an analysis use original embeddings or a projection?

Either may be used, depending on the analysis. UMAP changes the representation and its authors also describe use as preprocessing for clustering. Identify the representation, metric and parameters, and evaluate the intended task. A visually convincing two-dimensional group is not proof of a biological category.