Skip to content
All articles

Supervised, Unsupervised, Reinforcement: ML Paradigms in Plain English

Dean Jain

Dean Jain

Senior Staff Software Engineer · Enterprise AI, Data & Cloud Architect

· 5 min read

Machine LearningFundamentalsData Science
---
config:
  theme: neutral
  fontSize: 16
---
flowchart TD
    ML["Machine Learning"]:::gov --> S["Supervised<br/>learns from labelled answers"]:::server
    ML --> U["Unsupervised<br/>finds structure, no labels"]:::obs
    ML --> R["Reinforcement<br/>learns from reward"]:::warn
    S --> S1["Classification"]:::good
    S --> S2["Regression"]:::good
    U --> U1["Clustering"]:::good
    U --> U2["Dimensionality reduction"]:::good
    R --> R1["Policy: state  action"]:::good
    classDef gov fill:#FFFFFF,stroke:#333333,stroke-width:1px,color:#111111
    classDef server fill:#F0F0F0,stroke:#333333,stroke-width:1px,color:#111111
    classDef obs fill:#DAE8FC,stroke:#333333,stroke-width:1px,color:#111111
    classDef warn fill:#DCDCDC,stroke:#333333,stroke-width:1px,color:#111111
    classDef good fill:#E8E8E8,stroke:#333333,stroke-width:1px,color:#111111

Figure 1: Three learning paradigms, distinguished by one thing what kind of feedback the machine gets.

“Machine learning” sounds like one thing. It’s really three, and they differ on a single question: what feedback does the machine get while it learns? Labelled answers? No answers at all? Or a reward signal it has to earn? Get that distinction and most of ML stops being a buzzword and starts being a set of tools you can actually choose between.

Summary

  • Supervised learning trains on examples with the answers attached (labels). It learns the mapping from input to output, then predicts on new inputs. Most real-world ML is this.
  • Unsupervised learning gets data with no answers and finds structure on its own groupings, compressions, anomalies.
  • Reinforcement learning has no fixed dataset at all an agent acts in an environment and learns from reward and penalty, discovering a strategy by trial and error.
  • The deciding question is your data, not your ambition. Do you have labels? Do you have an environment to interact with? That answers which paradigm you’re in.
  • They combine. Real systems mix them e.g. unsupervised pre-processing feeding a supervised model, or RL fine-tuned on top of supervised pretraining.

1. Supervised learning: learning from labelled answers

Supervised learning is the most common paradigm and the easiest to picture: you show the model thousands of examples where the correct answer is already known, and it learns the mapping from input (features/predictors) to output (the target). Show it enough labelled emails marked “spam” or “not spam,” and it learns to label the next one.

It splits cleanly into two jobs by the shape of the answer:

  • Classification the target is a category. Spam/not-spam, which-digit, fraud/legit, disease/healthy. The model is effectively drawing decision boundaries between classes.
  • Regression the target is a continuous number. House price, tomorrow’s demand, expected lifetime value. The model is fitting a curve through the data.

The catch that defines the whole paradigm: you need labels, and labels are expensive. Someone, or some process, had to mark every training example with the right answer. That labelling cost is the single biggest practical constraint on supervised learning and the reason the other two paradigms exist.

2. Unsupervised learning: finding structure with no answers

Now take the labels away. Unsupervised learning gets raw data with no target column and has to find structure on its own a self-directed search for patterns and relationships nobody pointed out in advance. Three things it’s good at:

  • Clustering grouping similar points together. Customer segmentation, grouping documents by topic, finding communities. K-means and hierarchical clustering are the workhorses.
  • Dimensionality reduction compressing many features into a few that retain the signal. PCA is the classic; it makes data easier to visualize, store, and feed to downstream models.
  • Anomaly detection flagging points that don’t fit the learned structure. Fraud, defects, intrusions the rare thing that differs sharply from the majority.

The trade is the mirror image of supervised learning: no labels needed, but no ground truth to check against. You can’t compute a clean accuracy number when nobody told you the right answer. Evaluating unsupervised results takes judgement and domain knowledge which is exactly why it’s often used to explore and to prepare data for a supervised step, rather than as the final answer on its own.

3. Reinforcement learning: learning from reward

Reinforcement learning throws out the dataset entirely. Instead of examples, there’s an agent acting in an environment, and it learns by consequence: take an action, observe the new state, collect a reward or penalty, repeat. Over many trials it converges on a policy a strategy mapping situations to the actions that earn the most reward over time.

---
config:
  theme: neutral
  fontSize: 16
---
flowchart LR
    A["Agent"]:::server -->|"action"|E["Environment"]:::obs
    E -->|"new state"|A
    E -->|"reward / penalty"|A
    classDef server fill:#FFFFFF,stroke:#333333,stroke-width:1px,color:#111111
    classDef obs fill:#F0F0F0,stroke:#333333,stroke-width:1px,color:#111111

Figure 2: The RL loop act, observe the new state, collect reward, and refine the policy.

This is how you train a system to play a game, balance a robot, route traffic, or manage a portfolio anywhere the “right answer” isn’t a fixed label but a long-run outcome you can only learn by trying. The defining difficulty is the exploration–exploitation trade-off: exploit the actions you know pay off, or explore new ones that might pay off more? And rewards are often delayed the move that won the game happened twenty moves earlier so the agent must learn to credit actions for outcomes that arrive much later. That’s what makes RL powerful and notoriously hard to get stable.

The three at a glance

SupervisedUnsupervisedReinforcement
FeedbackLabelled answersNoneReward / penalty
LearnsInput → output mappingHidden structureA policy (state → action)
Typical tasksClassification, regressionClustering, dim. reduction, anomaliesControl, games, sequential decisions
Data you needLabelled datasetUnlabelled datasetAn environment to interact with
Main painLabels are expensiveNo ground truth to score againstUnstable; delayed, sparse rewards

Figure 3: One table to keep straight which paradigm fits which situation.

How to choose

You don’t pick a paradigm by ambition you pick it by what feedback your problem can actually provide. Walk it as a quick decision:

  • Do you have labelled examples of the answer? → Supervised. (Is the answer a category or a number? That picks classification vs regression.)
  • No labels, and you want to discover structure, segments, or outliers? → Unsupervised.
  • No fixed dataset, but you have an environment to act in and a reward you can define? → Reinforcement.
---
config:
  theme: neutral
  fontSize: 16
---
flowchart TD
    Q1{"Have labelled<br/>answers?"}:::gov -->|"yes"|S["Supervised"]:::server
    Q1 -->|"no"|Q2{"Want to discover<br/>structure?"}:::gov
    Q2 -->|"yes"|U["Unsupervised"]:::obs
    Q2 -->|"no"|Q3{"Have an environment<br/>+ a reward?"}:::gov
    Q3 -->|"yes"|R["Reinforcement"]:::warn
    classDef gov fill:#FFFFFF,stroke:#333333,stroke-width:1px,color:#111111
    classDef server fill:#F0F0F0,stroke:#333333,stroke-width:1px,color:#111111
    classDef obs fill:#DAE8FC,stroke:#333333,stroke-width:1px,color:#111111
    classDef warn fill:#DCDCDC,stroke:#333333,stroke-width:1px,color:#111111

Figure 4: The choice as a decision tree your data's feedback picks the paradigm, not your ambition.

And don’t assume it’s one or the other. Production systems routinely stack them: unsupervised clustering or PCA to clean and compress data before a supervised model; reinforcement learning layered on top of supervised pretraining to fine-tune behavior. The paradigms are a vocabulary for thinking about feedback, not rival teams.

Why it matters: “should we use ML here?” is the wrong first question. The useful one is “what feedback can this problem give a model?” labels, structure, or reward. Answer that and you’ve already chosen your paradigm, scoped your data work, and known your main risk before writing a line of code.

References