embedding
/ em-BED-ing /
An embedding is a way of turning something a computer can't do math on — a word, a product, a user, an image — into a list of numbers that it can. Crucially, it is not a random code: the numbers are arranged so that things with similar meaning land close together in space, and unrelated things land far apart. Picture a giant map where every word is a pin: "king" and "queen" sit near each other, "banana" sits across the room, and the very direction from "man" to "woman" turns out to match the direction from "king" to "queen." Meaning becomes geometry.
Concretely, an embedding is a vector — a short list of numbers, perhaps a few hundred of them — assigned to each item. These numbers are not hand-written; the network learns them while training on a real task, nudging similar items together because they behave similarly in the data. Words used in similar contexts drift close; products bought by the same people cluster; songs with the same vibe gather. The same word in three languages may even land in nearby spots once trained together. An embedding is the practical form that learned, hierarchical-representation takes for discrete things.
Why this matters: once meaning lives in geometry, many tasks become easy. "Find me similar items" is just "find nearby points." A recommender suggests neighbors; a search engine matches by closeness rather than exact words. Embeddings are the quiet engine under search, recommendations, and the input layer of every language model. The honest caveats: an embedding is only as fair and accurate as the data that shaped it — it cheerfully absorbs the biases in that data — and the individual numbers in the vector almost never carry a tidy human meaning. The structure lives in the relationships between vectors, not in any single coordinate.
In a well-trained word embedding, the arithmetic king − man + woman lands closest to the vector for queen. The directions in the space have quietly captured a relationship — royalty held fixed while gender is swapped.
When meaning becomes geometry, analogies turn into arithmetic.
An embedding's power lives in the relationships between vectors (who is near whom), not in any single number. And because it is learned from data, an embedding faithfully inherits that data's biases — a known source of unfair behavior in real systems.