An embedding is a fixed-length numeric vector produced by a neural network (encoder) that maps discrete inputs — tokens, images, or structured records — into a continuous metric space. The mapping is trained to preserve relational structure: inputs sharing semantic or functional similarity are placed near each other in vector distance. Persistence is through saved model weights and serialized vector files (e.g. .npy, .bin) distributed with the model. The embedding space is defined by its dimensionality (d), the distance metric (typically cosine or L2), and the training objective (contrastive, autoencoder, or next-token prediction). The encoder network applies learned linear and non-linear transformations to input token IDs or pixel arrays, producing a d-dimensional vector that serves as the input for downstream models.\n\n[formal: embeddere | substrate: mind | horizon: hours | explicit: yes | epoch: 0.01]
Accepted ontology entry
embedding
An embedding is a fixed-length numeric vector produced by a neural network (encoder) that maps discrete inputs — tokens, images, or structured records — into a continuous metric space. The mapping is trained to preserve relational structur…
Definition
Why it is in scope
A human-made continuous vector representation of discrete data (words, images, entities) in a high-dimensional space, designed so that semantic similarity is reflected as geometric proximity. Built to persist through training pipelines and deployed in retrieval, classification, and generation systems.
Names and aliases
- embeddingen · CANONICAL
Relations from this entry
- cmrng8t8a02cyd1nlcuj11wc0INSTANCE_OF →
An embedding is a specific kind of representation — a numerical/vector encoding that captures semantic relationships between entities. A competent speaker would call an embedding 'a representation'. Law 9: specific→general.
- cmrg0scos00ef2a1nklfvbk7xSERVES →
Embeddings are built and maintained for the sake of machine learning — they provide dense vector representations that ML models use to learn from data. The designed purpose is to further ML's operation.
- cmrg0pc1x00e82a1nppcnwovjDERIVED_FROM →
DERIVED_FROM (Law 7, which-came-first test): neural networks existed first (McCulloch-Pitts, 1943; backpropagation era 1986). Dense vector embeddings as a semantic representation technique emerged from neural network research — Hinton's distributed representations (1986) and word2vec (2013) built directly on neural network architectures. The concept of representing discrete items as continuous vectors is a product of neural network training dynamics.
- cmsle0laa06w4nobp7rttnbliDERIVED_FROM →
Representation learning (the broader field of learning useful feature representations from data) predates and feeds into embedding techniques. The which-came-first test (Law 7): representation learning emerged in the 1980s-90s with autoencoders and unsupervised feature learning; embedding techniques (word2vec, etc.) are a specialized application of that broader tradition. Y existed first and fed into X.
Relations to this entry
- cmsmfvta001bu1q13jgi20vv8← DEPENDS_ON
Latent space operates through embeddings — vectors in a continuous space. Remove embeddings and latent space has no mechanism to exist. The vectors/embeddings are the substance of the latent space, making this a present-tense dependency (Law 8).
Record identity
- Created
- Aug 3, 2026, 1:32 AM UTC
- Content hash
- 850b2956fb355a708f0340b7287560badc22df89b751a7f3b20bccaf57944f9b