Broad intuition · Study map 16 of 85
Embeddings: useful coordinates for meaning
Suggested preparation: Optimizers: walking with noisy directions
Picture it
An embedding gives an item a learned address in a space. Training arranges those addresses so a useful relationship becomes easy to measure. Similarity is a property encouraged by an objective, not an automatic understanding of meaning.
What to understand
- Token, sentence, image, and item embeddings
- Contrastive learning, negative examples, transfer learning
Follow the references
1 · Course
Google · Recommendation systems ↗Use embeddings and matrix factorization to see how an objective shapes the space.
You understand it when…
Why should a document embedding model be evaluated on retrieval rather than only on pretty clusters?
This is a study map. Follow the reference for the full explanation; use the question to test your mental model. Browse the reference shelf.