Why Does the Type of Learning Matter?
When people say "machine learning", they're actually talking about a family of techniques that train in very different ways. The type of learning you choose shapes everything: what data you need, how long training takes, and what kinds of problems you can solve. Getting this distinction right early will make every AI article, job listing, and research paper you read much easier to follow.
There are three core types: supervised, unsupervised, and reinforcement learning. Let's walk through each one with a concrete analogy, then we'll compare them side by side.
Supervised Learning: Learning from Labelled Examples
Supervised learning is the most widely used type. The idea is simple: you hand the algorithm a large set of examples where every input already has the correct answer attached. The algorithm studies those input–answer pairs until it can predict answers for examples it has never seen before.
The "labels" are the pre-written answers. Gathering them is often the expensive part: a human (or a reliable automated process) has to tag each example correctly. Once you have a labelled dataset, though, training is relatively straightforward.
Classification vs Regression
Supervised problems split into two flavours depending on the type of answer you want:
- Classification — the answer is a category. Is this email spam or not? Is this photo a cat, dog, or bird? Which digit (0–9) is handwritten in this image?
- Regression — the answer is a number on a continuous scale. What will this house sell for? How many units will we ship next quarter?
Everyday supervised learning examples include spam detection, house-price prediction, medical image classification, credit-risk scoring, and sentiment analysis of product reviews. If you'd like to train a classifier yourself right now, head to Lesson 2: How Machines Learn.
Unsupervised Learning: Finding Structure Without Labels
What happens when you have lots of data but no labels? That's where unsupervised learning comes in. The algorithm is given raw inputs and told to find whatever patterns or groupings exist — with no guidance on what the "right" answer is.
Because there are no labels to generate, unsupervised learning is often far cheaper to apply to large datasets. The trade-off is that the algorithm's output — say, three distinct clusters — may still require a human to interpret what each cluster means.
Common applications include:
- Customer segmentation — grouping shoppers by behaviour so marketers can personalise campaigns.
- Anomaly detection — flagging transactions that don't match any known pattern, which is how many fraud-detection systems work.
- Recommendation grouping — finding which products or songs tend to appear together, powering "you might also like" features.
Understanding the raw ingredients of your data is central to unsupervised work. Our guide on Data & Features (Lesson 3) covers how to think about what goes into any ML model.
Reinforcement Learning: Learning by Trial and Error
Reinforcement learning (RL) is the odd one out. There are no labelled examples and no fixed dataset at all. Instead, an agent takes actions inside an environment and receives a reward signal — positive when it does something good, negative when it does something harmful. Over thousands or millions of attempts, it learns which actions lead to the highest cumulative reward.
RL is behind some of the most dramatic AI achievements: DeepMind's AlphaGo defeated the world champion at Go; OpenAI Five beat professional Dota 2 players; robotic arms learn to grasp objects without explicit programming. It's also increasingly used in recommendation and ad-bidding systems where the agent's "action" is choosing what to show next and the "reward" is a click or a purchase.
The downside: RL typically needs a simulator or live environment to generate experience, and training can be extremely slow and computationally expensive compared with supervised or unsupervised approaches.
Side-by-Side Comparison
| Type | What it learns from | Everyday example | When to use it |
|---|---|---|---|
| Supervised | Labelled input–output pairs | Spam filter, house-price predictor | You have labelled data and a specific prediction target |
| Unsupervised | Unlabelled data — structure only | Customer segments, fraud detection | You want to explore data or lack labels |
| Reinforcement | Rewards and penalties from an environment | Game-playing AI, robotics | You can define a reward and simulate many interactions |
How Do I Choose the Right Type?
Ask yourself three questions:
- Do I have labelled data? If yes, and you have a clear prediction target (a category or a number), supervised learning is usually your first stop.
- Am I trying to explore or discover patterns? If you don't have labels and want to understand the structure of your data, try unsupervised methods first.
- Can I simulate an environment and define a reward? If your problem is about sequential decisions — where each action changes the situation — reinforcement learning may be the best fit, though it's also the hardest to set up correctly.
In practice, many real-world systems combine approaches. It's also worth knowing that semi-supervised learning sits in the middle: you label a small fraction of your data, then use unsupervised techniques to extend those labels across the rest — a useful middle ground when labelling everything is too expensive.
For a broader look at where ML fits relative to AI and deep learning, see our article AI vs Machine Learning vs Deep Learning.
Frequently Asked Questions
Which type of machine learning is most common?
Supervised learning is by far the most widely used in production systems today. Most business applications — fraud detection, recommendation engines, image recognition — rely on it because companies usually have enough historical data with known outcomes to create labelled datasets. Unsupervised techniques run a close second as an exploratory tool, while reinforcement learning, despite its headline-grabbing results, is still relatively niche outside gaming and robotics.
Is ChatGPT supervised or unsupervised?
Both, in different stages. Large language models like the ones powering ChatGPT are first pre-trained on massive amounts of unlabelled text using a technique called self-supervised learning (a variant of supervised learning where the labels are generated automatically — for example, predicting the next word). They're then fine-tuned using human feedback, which is a form of reinforcement learning called RLHF (Reinforcement Learning from Human Feedback). So the short answer is: all three types play a role at different points in training.
Do I need labelled data to get started with machine learning?
Not necessarily. If you have no labels, unsupervised methods let you start discovering structure immediately. Many open datasets also come pre-labelled, so you can practice supervised learning without building a dataset from scratch. For hands-on practice, Lesson 2 walks you through training a classifier on a ready-made dataset — no data collection required.