Parametric UMAP
Algorithms AIMech:ParametricUMAP DRAFT
Aliases: Parametric Uniform Manifold Approximation and Projection
A parametric dimensionality-reduction algorithm that trains a neural mapping from input points to embedding coordinates using sampled attractive and repulsive losses derived from a UMAP neighborhood graph.
Identity Notes & Concept Scope
The core distinction from AIMech:UMAP is optimization of a shared neural input-to-coordinate mapping, rather than one free coordinate vector per training point. A parametric t-SNE objective and a generic autoencoder are different constructions; the UMAP/autoencoder and supervised hybrids are optional variants. All 442 pinned AIO labels/scoped synonyms and 114 preceding native identities were compared, together with hidden/ignored-inclusive searches of both checkouts; source citations and pending frontier mentions were distinguished from existing concept identities.
Claims & Grounded Evidence
2 Claims
Mechanism
#method-scope
Construct locally scaled UMAP graph affinities, then optimize neural-network weights so the network outputs serve as embedding coordinates. Positive graph edges and sampled negative pairs supply a minibatch binary cross-entropy-style training loss; the learned mapping can embed new inputs by a forward pass. Reconstruction, classification or global-distance regularizers are optional extensions rather than necessary parts of the base algorithm.
Limitation
#parametric-sampling-scope
The sampling analysis distinguishes Parametric UMAP’s within-batch negative samples and gives a batch-size-dependent effective objective with degree-product repulsive weights. This differs from its nonparametric symmetrized objective and from nominal all-pairs cross-entropy; the result concerns the analyzed procedure, not every implementation.
⚠️ Curated Limitations & Negative Results
- The graph metric, neighborhood structure, network capacity, optimization and sampling affect the learned representation; a reusable mapping does not ensure reliable extrapolation to new distributions.
- Minibatch negative sampling changes the effective loss relative to the nominal all-pairs fuzzy cross-entropy. The analyzed parametric loss has batch-size and degree-product factors, distinct from the nonparametric symmetrized loss.
- A classifier regularized by graph structure requires a graph relevant to its labels; the paper does not establish universal semi-supervised improvement.
- The specified paper and sampling analysis were inspected; current software, all supplementary results and universal topology preservation were not certified.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:DimensionalityReduction
Dimensionality Reduction
|
related_match | A specific algorithm for constructing lower-dimensional representations, associated with the pinned dimensionality-reduction task. The target inherits UnsupervisedLearning under MachineLearningTask; related_match avoids treating the algorithm as identical to or a subclass of that task. Shared embedding components do not make this concept identical to PCA, MDS or t-SNE. |
Evidence Sources & Primary Literature
2 Sources| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Tim Sainburg, Leland McInnes, Timothy Q Gentner
|
primary paper
full_text
|
ID: arXiv:2009.12981
sha256:1df3222d6b8d…
|
sampling-analysis
|
Sebastian Damrich, Fred A. Hamprecht
|
primary paper
full_text
|
ID: https://proceedings.neurips.cc/paper/2021/hash/2de5d16682c3c35007e4e92982f1a2ba-Abstract.html
sha256:131723625e0d…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T05:21:17Z
Updated: 2026-10-11T05:21:17Z
Source File:
data/curated/ParametricUMAP.yaml