Learning techniques and methods
The named techniques and model types that show up as one-line scenario answers - including federated learning (data never leaves the site), transfer learning, fine-tuning, and the discriminative-vs-generative split.
These are the named techniques and model types the exam drops into one-line scenarios. Several are easy to confuse, so anchor each to its distinguishing move.
- Active learning → the model flags the most informative unlabelled examples and asks a human to label them, cutting labelling cost.
- Adaptive learning → systems that adjust behaviour or content in response to new data and user interactions over time (personalised tutoring).
- Transfer learning model → reuses knowledge from one task as the starting point for a related task, saving data and compute.
- Federated learning → local sites train a shared model on their own data and only the results aggregate centrally; the data never leaves the site.
- Fine-tuning → further training a pre-trained model on a smaller, task- or domain-specific dataset to specialise it.
- Classification model vs Clustering → classification predicts a discrete category (supervised); clustering groups by similarity with no labels (unsupervised).
- Decision tree / Random forest / Bootstrap aggregating (bagging) → a single interpretable tree; an ensemble of trees on random subsets; and the resampling-with-replacement method that reduces variance.
- Greedy algorithms → locally optimal choice at each step, fast but not guaranteed globally optimal.
A Discriminative model learns the boundary between classes to classify - it distinguishes rather than creates. Generative AI (next cluster) creates new content. This is one of the eight planted confusion pairs.
Federated learning (privacy: data stays put) vs transfer learning (efficiency: reuse a prior model) vs active learning (labelling: ask a human about the hard cases) vs fine-tuning (specialise a pre-trained model). Each answers a different problem.