Yesterday we looked at the relationship between AI, machine learning, and deep learning. Today we will learn the three major learning types into which machine learning is divided based on "how it learns."
These three appear frequently on the exam, and the key differences are "whether there are answers (labels)" and "how the learning signal is received." We will also cover basic task concepts such as classification, regression, and clustering.
| Learning type | Answers (labels) | Key question | Representative tasks |
|---|---|---|---|
| Supervised learning | Present | "How do we learn to predict when the answers are known?" | Classification, regression |
| Unsupervised learning | Absent | "How do we find hidden structure without answers?" | Clustering, dimensionality reduction |
| Reinforcement learning | Replaced by rewards | "How do we learn the best actions through trial and error?" | Control, games, robotics |
Supervised learning trains on data where inputs and answers (labels) are given together. Think of it as a "teacher teaching while providing the correct answers."
For example, if you train a model on dog/cat photos each labeled "dog" or "cat," the model learns to predict which one a new photo shows.
Supervised learning is broadly divided into two types.
Used when the answer is one of a fixed set of categories (kinds). The output is categorical, like "is it this or that?"
Used when the answer is a continuous numeric value. The output is a quantity, like "how much?"
Click a choice to reveal the answer and explanation.
Question 1
Which approach trains a model on input data given together with answers (labels), so that the model learns to predict those answers?
Question 2
You want to take inputs such as a house's floor area and number of rooms and predict its "expected sale price (a numeric amount)." Which task is most appropriate?
Question 3
You want to automatically group customers with similar tendencies from customer data that has no answer labels at all. Which task is most suitable?
Question 4
Which statement best describes how reinforcement learning works?
Question 5
What is the most essential difference between classification (supervised learning) and clustering (unsupervised learning)?
💡 Related theory: The easiest criterion for distinguishing classification from regression is "is the output a category or a number?" "Spam/legitimate" is classification; "price prediction" is regression.
Unsupervised learning finds hidden structure or patterns on its own in data without answers (labels). The approach is "organize the data without an answer key."
The task of automatically grouping similar data together. It differs from classification in that there are no predefined answer groups.
For example, if you say "divide the customers into 3 groups," unsupervised learning groups similar people together without ever being told in advance who belongs to which group.
When data has too many features (variables), this task reduces the number of variables while preserving as much important information as possible. It simplifies the data and makes it easier to visualize.
💡 Related theory: Classification (supervised) and clustering (unsupervised) are easy to confuse because "both divide things into groups." The decisive difference is the presence or absence of answer labels. Classification assigns to groups with known answers; clustering groups similar items together without answers.
Reinforcement learning learns through trial and error and rewards. Instead of being given answers directly, the agent receives a reward when an action's outcome is good and a penalty when it is bad, gradually learning better behavior. It is similar to "training a dog by giving treats when it does well."
The core components are as follows.
| Term | Meaning | Analogy |
|---|---|---|
| Agent | The entity that learns and acts | Game player |
| Environment | The world the agent interacts with | Game screen |
| Action | The choice the agent makes | Pressing a button |
| Reward | A score for the action (good/bad) | Gaining/losing points |
Representative examples include game AI (Go, chess), robot control, and autonomous driving control.
💡 Related theory: Reinforcement learning has no prepared answer data. Instead, the agent learns an action strategy that maximizes the rewards it receives while interacting with the environment. This is fundamentally different from supervised/unsupervised learning.
Even with the same "customer data," the learning type differs depending on the objective.
| Objective | Learning type | Reason |
|---|---|---|
| Predicting whether a customer will churn (yes/no) | Supervised learning (classification) | Learns from past churn outcomes (answers) |
| Predicting this customer's spending next month | Supervised learning (regression) | Predicts a numeric value |
| Grouping similar customers together | Unsupervised learning (clustering) | Groups by similarity without answers |
| Learning the optimal recommendation order per customer via rewards | Reinforcement learning | Learns a strategy through trial and error and rewards |