Lesson 18 showed that a network is a composed function that maps inputs to outputs. An embedding is a particular, very useful kind of output: instead of a class label or a number, the network maps each item — a word, a product, an image, a user — to a point in a high-dimensional vector space (say, 128 or 768 numbers). Two items that end up as nearby points are called “similar” in that space, and that single move — turning things into points so you can measure distance between them — underlies recommendation systems, search, clustering, and most of what “semantic search” means in practice.
The trap is treating “close in embedding space” as if it meant “similar” in some single, objective, universal sense. It doesn’t. An embedding space is shaped entirely by what the training objective rewarded. Train a model to predict which products get bought in the same order, and it will place products close together exactly when they tend to co-occur in orders — regardless of whether they “mean” anything alike. Train a different model to predict which products share similar text descriptions, and closeness now tracks vocabulary overlap in the copy — regardless of purchase behavior.
The scenario: a retailer trains Model A on co-purchase data and Model B on product-description text, over the same catalog. In Model A’s space, “charcoal” and “lighter fluid” land close together. In Model B’s space, they land far apart. Neither model is broken.
Your task: pick the explanation that correctly accounts for both placements being right.