MM1041 Chap.6 Three Types of Machine Learning
Three Types of Machine Learning
One distinction, three labels
Both the first lecture and the e-learning lesson divide machine learning into three types, using the same labels. Supervised learning is trained with targets, unsupervised learning is trained without targets, and reinforcement learning is trained with mistakes.
The whole chapter rests on one question you can ask of any system: what does its training signal contain?
Supervised learning
Supervised models learn from examples that people have labelled. The lesson's images carry labels such as cat, bird or car; the model learns features such as wheels, eyes and ears, and predicts the category of a new test image. That is classification, with a discrete label.
Regression uses a continuous label: the lesson trains on people described by gender, height and age, each labelled with weight, and predicts the weight of a new person. Object detection for autonomous driving is the showcase application.
Unsupervised learning
Unsupervised models learn from data with no labels.
They can group similar images, such as dogs with dogs and pandas with pandas, without being told what either animal is, and they can reduce dimension by storing a compact representation that keeps the key information.
The lesson's demonstration turns a video of a horse into a realistic video of a zebra with no labelled examples.
Reinforcement learning
A reinforcement learning agent senses the state of its environment, takes an action and receives a new state with a reward or a penalty. It adjusts its strategy to maximise reward.
The examples are computer games and a robot that learns to navigate a changing room without hitting obstacles.
Training data, testing data and overfitting
The first lecture adds the data side. Data is called the new oil: expensive, only valuable once converted into something useful for decisions, produced in greater quantity as operations grow, and less valuable the more people hold it.
Training data trains an algorithm; testing data tests it. An algorithm that fits its training data but not the real situation is overfitting, which the lecturer compares with doing every past paper without understanding the subject.
What this chapter covers
- 01
Trained with targets, without targets, with mistakes
- 02
Classification with categorical labels
- 03
Regression with continuous labels
- 04
Clustering and dimension reduction
- 05
Agent, environment, reward and penalty
- 06
Training data against testing data
- 07
Overfitting and the past-papers analogy
Worked example · free
Choose a learning type for three bank problems
- 2Classify the ATM project.
- 2Classify the spending-groups project.
- 2Classify the chatbot project.
Key terms
- Classification
- A supervised task that predicts a discrete category, such as spam or not spam.
- Regression
- A supervised task that predicts a continuous number, such as a weight or a sales figure.
- Clustering
- An unsupervised task that groups similar data points without predefined labels.
- Dimension Reduction
- Representing data with fewer features while keeping the information that matters.
- Training Data
- Data used to train an algorithm.
- Testing Data
- Separate data used to check how well a trained algorithm performs on examples it has not seen.
- Reward Signal
- Feedback that tells a reinforcement learning agent how good its last action was.
Three Types of Machine Learning FAQ
What is the easiest way to tell supervised from unsupervised learning?
Check whether the training examples carry the answer. If each example comes with a label that people supplied, such as a category or a number, the learning is supervised. If the data arrives with no labels and the model has to find structure by itself, it is unsupervised. The course summarises this as trained with targets against trained without targets.
Why is reinforcement learning described as trained with mistakes?
Because there is no labelled right answer for each step. The agent tries actions, and the environment returns a reward when things go well and a penalty, which the lesson calls a mistake, when they do not. Over many attempts the agent shifts toward actions that earn reward, as a game program or a navigating robot does.
How can a model that scores well still be a bad model?
If it was scored on its own training data. A model can fit the examples it learned from almost perfectly and still fail on new cases, which is overfitting. The protection is to keep separate testing data and judge the model only on examples it did not see, just as practising the same past papers does not prove you understand the subject.
Which type of learning does a chatbot like ChatGPT use?
The first lecture sets this as an in-class pairing question for discussion. A defensible reply is that a large chatbot combines types: it learns patterns from vast amounts of unlabelled text, is refined with examples of good replies, and is further tuned with human feedback that acts like a reward. Say which part of the system you mean.
Exam move
Build a two-column flash card set: a business problem on the front, the learning type and task on the back. Aim for twelve cards spread across the three types, including at least two regression cases. Then rehearse the overfitting explanation in one sentence with the past-papers analogy.
Working through Three Types of Machine Learning in MM1041? Sia is AskSia’s AI Information Technology tutor — ask any MM1041 Three Types of Machine Learning question and get a clear, step-by-step explanation grounded in how MM1041 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
