Embedding

An embedding is a list of numbers, typically a few hundred to a few thousand values, that a model produces to represent an input. The numbers themselves are not human-readable, but their geometry is meaningful: inputs with similar content are mapped to nearby vectors, and unrelated inputs land far apart. That single property is what makes embeddings so useful, because once similarity is a distance, it can be computed at scale.

Almost any data can be embedded, including images, text, audio, video, point clouds, and robot trajectories, which is why embeddings have become the common currency of modern AI systems.

An embedding model, often a neural network trained with contrastive or self-supervised objectives, transforms each input into a fixed-length vector. Distances between vectors, such as cosine similarity, quantify how alike two inputs are. Collections of embeddings can then be indexed for fast search, projected to 2D for visualization, or clustered to reveal a dataset's structure.

Embeddings are the foundation of semantic search, recommendation, near-duplicate detection, anomaly detection, and data curation. For visual and physical AI teams, they turn overwhelming datasets into something navigable: instead of scrolling through millions of samples, you can search by meaning, group by similarity, and surface the rare cases that matter.

Embeddings are the modern, learned form of features. The terms are often used interchangeably, though embedding usually implies a dense vector from a neural network.

Typically between a few hundred and a few thousand dimensions, depending on the model. Longer is not automatically better.

Image embeddings, Embedding model, Embedding space, Feature extraction, Embedding and semantic search.

Learn: How image embeddings transform computer vision · Multimodal embeddings for physical AI

Try it: Using image embeddings in FiftyOne · FiftyOne Brain