Module 6 — The three learning regimes
Facing a problem, the first structuring decision is not choosing an algorithm: it is identifying the learning regime, which follows directly from the shape of your data. This module reviews the three regimes with their requirements, documented use cases and pitfalls, then presents the hybrid regimes powering the most recent systems.
Supervised: you know the right answers
The dominant regime in business. You have labeled examples — each input comes with the expected output — and the model learns the mapping.
Two sub-families depending on the nature of the output. Classification predicts a category: fraudulent or legitimate, churns or stays, urgent, normal or informational. Regression predicts a quantity: a price, a delay, a consumption. The distinction drives the metrics (module 7) and sometimes the algorithm; the same business question can often be phrased both ways, and that phrasing choice is already a design decision.
The hidden requirement: the cost of labels. Labels come either from the activity itself (loan repayments are observed, sales are recorded) or from an annotation project with its budget, guidelines, inter-annotator disagreements and quality control. First question for any supervised project: where do the labels come from, how much do they cost, and are they reliable?
The time-lag trap. Observed labels describe the past. If the phenomenon drifts — behaviors, prices, frauds evolve — the model ages. Hence production monitoring, covered in module 8.
Unsupervised: you are looking for an unknown structure
No labels; the model looks for regularities nobody defined in advance.
Clustering groups similar examples: customer segments that emerge from actual behavior rather than marketing assumptions, failure families, groups of related documents.
Dimensionality reduction condenses hundreds of variables into a few axes carrying most of the variation: visualization, faster processing, denoising.
Anomaly detection models normal behavior and flags deviations — precious when incidents are too rare and too diverse to label: intrusions, equipment failures, atypical transactions.
The structural difficulty: no ground truth. Five customer segments are neither right nor wrong. The only criteria are interpretability, stability, and usefulness for decisions. Evaluation is an instrumented business judgment, not an automatic metric. Budget the interpretation time: that is where unsupervised projects succeed or fail.
Reinforcement: learning by acting
No labels and no fixed dataset. An agent acts in an environment, receives delayed rewards, and adjusts its strategy to maximize cumulative gain. This is the regime of the spectacular successes — the game of go, complex video games, robotic control in simulation.
Its rarity in business comes from three requirements seldom met together. You need a reliable simulator or an environment where mistakes cost nothing — a game replays for free, a pricing policy that learns by failing drives away real customers. You need a well-defined reward — and a badly designed reward is optimized to the letter, perverse effects included. You finally need a gigantic volume of trials.
Its discreet but real presence: ad-bidding optimization, long-horizon recommendation, data-center energy management — and the alignment of language models (RLHF), where the "simulator" is replaced by human preferences. A full course in this track is devoted to it.
Hybrid regimes: the best of both worlds
The boundaries have blurred, and that is excellent news for labeling costs.
Semi-supervised combines a small labeled set with a mass of raw data: train a first model on the few labels, use it to pre-annotate the rest, correct, iterate. Standard whenever annotation is expensive.
Self-supervised manufactures supervision from the data itself: mask a word and have it predicted, hide a patch of image and have it reconstructed. No human annotation, terabytes of free "right answers". It is the training regime of large language and vision models — the discovery that made foundation models possible.
The classic mistake is picking an attractive technique and then looking for a problem to justify it. The healthy approach starts from the inventory: what data exists, in what shape, with what labels? The regime — and often the algorithm — follows almost mechanically.
Key takeaways
- Supervised: labels available, input → output mapping; the critical question is the origin and cost of the labels.
- Unsupervised: unknown structure to discover; evaluation is a usefulness judgment, not a metric — budget the interpretation.
- Reinforcement: agent, environment, delayed reward; requires a simulator and a well-designed reward — rare in business, central to LLM alignment.
- Self-supervised: supervision manufactured from the data itself — the engine of foundation models.
Next module: evaluation metrics — accuracy, precision, recall, and the honest reading of the numbers you are shown.