AI Governance Study

How machines learn: supervised, unsupervised, reinforcement and semi-supervised training

How a system was trained determines what data it needed, what could go wrong with that data, and how much human judgment went into it. Those are governance questions, and they all start with recognizing which of the four training approaches is in front of you.

Domain I · Understand what AI is and why it needs governance · about 8 min

Why this matters for the exam

How a system was trained determines what data it needed, what could go wrong with that data, and how much human judgment went into it. Those are governance questions, and they all start with recognizing which of the four training approaches is in front of you.

What you need to know

The last topic said that what makes something machine learning is improving with experience. This one explains how that training actually happens. A system does not inherently know how to solve a problem; it has to be trained on data, and there are four recognized ways to do it.

The four training approaches

The approaches differ in one main thing: what the training data looks like, and in particular whether it comes with labels — correct answers attached by people.

ApproachData usedHow it learnsStrengths and challengesSubtypes and examples
Supervised learningPre-labeled, classified datasetThe algorithm analyzes inputs together with their correct labels, compares its own output to the correct answer, and adjusts to reduce the error.Accurate when the labeled data is good. But labeling is labor-intensive, expensive, and can itself introduce bias.Regression predicts a continuous value (a car's price from mileage, year and features). Classification predicts a category (spam or not-spam).
Unsupervised learningRaw, unlabeled dataFinds patterns, structures and relationships on its own, with no predefined targets.Discovers hidden insights cheaply. But interpreting the results is subjective, and behavior is less predictable.Clustering groups similar data points (DNA samples). Association rule learning finds relationships ("customers who bought X also bought Y").
Reinforcement learningNo pre-labeled dataset; learns through interactionAn agent acts in an environment and receives rewards or penalties. Rewarded behavior is reinforced, and the system learns to maximize its total reward over time.Learns complex behaviors without supervision. But designing the right reward is hard, and there is a constant trade-off between exploring new actions and exploiting known ones.Robotics navigation; real-time ad bidding; the core learning mechanism of agentic AI.
Semi-supervised learningA small labeled set plus a large unlabeled setCombines supervised and unsupervised techniques to improve reliability while cutting labeling cost.Useful whenever building a large, fully labeled dataset is impractical.Image and speech analysis; web search ranking; large language models commonly rely on it.

Large language models (LLMs) are deep learning models trained on massive volumes of text, learning the patterns and relationships among characters, words and phrases. The course files their training under the semi-supervised umbrella. Strictly, LLM pretraining is self-supervised: the model generates its own labels from the raw text, so nobody has to label it by hand.

Reading a scenario: which approach is it?

The data description gives the answer, so check for labels first.

The scenario saysThe clueThe approach
"The model studies a large set of unlabeled customer records to detect hidden patterns."No labels, no predefined targetsUnsupervised
"The model is trained on ten thousand emails marked spam or not-spam."Labeled data with known answersSupervised
"The system learns to route deliveries by trying routes and being scored on speed."No pre-labeled dataset; feedback arrives as rewardsReinforcement
"A small set of tagged photos plus a huge untagged archive."A little labeled data, a lot of unlabeled dataSemi-supervised

Whatever else a scenario describes, the training data tells you which approach it is.

Common ML algorithms

Governance professionals are not expected to build these. Recognizing them is enough to hold an informed conversation with a technical team and gauge risk.

AlgorithmWhat it is forExample
Linear regressionNumeric prediction from continuous variables.Predicting disaster-damage costs from housing and infrastructure data.
Logistic regressionA probabilistic technique predicting whether something is likely to occur, typically a yes-or-no outcome.Predicting customer churn.
Decision treesSupervised learning that classifies or predicts by walking a tree of decisions.Spam detection.
Random forestsAn ensemble of decision trees; more accurate than a single tree and better with complex data.Recommendation engines.
Neural networksBrain-inspired architecture for highly complex tasks.Natural language processing, facial recognition.

Architectures worth recognizing

Transformer models learn context and meaning by tracking relationships across sequential data, like the words in a sentence, using a mechanism called self-attention. They process inputs in parallel rather than one at a time, which is what makes modern LLMs and multimodal models practical to train. The same architecture is used in protein and DNA sequencing.

Three other architectures each specialize in a kind of data:

Retrieval-augmented generation (RAG) is a technique layered on top of a generative model. It lets the model pull in external, authoritative information when producing a response, which improves accuracy and reduces hallucination risk without eliminating it.

Terms that sound like machine learning, but aren't

Three terms regularly cause confusion because they sound like machine-learning vocabulary while meaning something else.

TermWhat it actually means
Cognitive learningNot a recognized way of training a model. The recognized training approaches are the four covered in this topic.
Data miningAnalyzing datasets to discover previously unknown patterns and properties. It is related to machine learning, but the goal is discovery in existing data, not a system that improves with experience.
Greedy algorithmA general programming strategy: take the best available choice at each step, without considering the overall best outcome. Not an ML training approach at all.

Next up: the kinds of models this training produces — classic versus generative, proprietary versus open-source, small versus large, and how capable they are.

Remember

  • Check the data first: labeled data with known answers means supervised learning (regression for a value, classification for a category).
  • Unlabeled data with no predefined targets means unsupervised learning (clustering, association rules).
  • No pre-labeled dataset, learning from rewards and penalties in an environment, means reinforcement learning.
  • A small labeled set plus a large unlabeled set means semi-supervised learning — the usual answer for LLMs.
  • Transformers process sequences in parallel and underlie modern LLMs; RAG adds an external knowledge source to a generative model to reduce hallucination.
  • Cognitive learning, data mining and greedy algorithms are not training approaches.

Practise this topic

Domain I is free in the app, including its practice questions and flashcards, with progress tracking and no card details.

Start studying free

Previous: What AI is: from AI to ML to deep learning to generative and agentic AI
Next: Comparing AI models: four axes, expert systems, and capability levels
Back to the AIGP study guide.

AI Governance Study is an independent study aid. It does not represent a government entity: it is not affiliated with, endorsed by or authorised by any government, government agency or regulatory authority, and it does not provide government services or legal advice. Laws and frameworks are described in our own words — the official texts are listed at official sources. It is also not affiliated with, endorsed by, or sponsored by the IAPP. The AIGP name is used only to identify the exam this material helps you prepare for.