2 · Classical machine learning

5. Unsupervised Learning & Representations

Clustering, dimensionality reduction and embeddings as the bridge to deep learning.

9 min read · 3 MCQs

Clustering

k-means partitions points around centroids and needs k up front; DBSCAN finds density-connected clusters and labels outliers; hierarchical clustering yields a dendrogram you can cut at any granularity.

Dimensionality reduction

PCA projects data onto the directions of greatest variance, compressing features while retaining structure. UMAP and t-SNE are non-linear methods used for visualising high-dimensional data — useful for inspection, not as model inputs.

Embeddings

An embedding maps a discrete item — a word, a user, a product — to a dense vector where distance means similarity. Embeddings are the shared currency of modern AI: search, recommendation and retrieval-augmented generation all run on nearest-neighbour lookups in embedding space.

Chapter quiz

3 questions · pass mark 75%
  1. 1. Which clustering method identifies outliers natively?

  2. 2. PCA keeps the directions of…

  3. 3. In embedding space, similarity is expressed as…

Answer every question to submit. Progress for ai-05 is saved in this browser.