You are viewing a public preview
Sign in to save your progress, do assignments, join discussions and get feedback.
After this unit you will: be able to distinguish labeled from unlabeled data, and name the three groups of machine learning algorithms.
By the end of this lesson you can:
Key idea: Whether your data has labels decides which kind of learning you can use. With labeled data (the correct answer is known) you can use supervised learning, the focus of this course; without labels you need unsupervised learning.
The lecture shows three items in yellow: a drawing of a cat, a drawing of a rabbit and a news headline. They are unlabeled data: raw information with no extra context about what it is. A computer looking at the cat picture only sees pixels; it does not know it is a cat.
In the middle stands an annotator, usually a person. The annotator looks at each item and attaches the correct answer: this is a cat, this is a rabbit, this headline is about sport. The same items, now tagged and shown in green, are labeled data.
| Unlabeled data | Labeled data | |
|---|---|---|
| What it contains | Raw inputs only | Inputs plus the correct answer (label) |
| Example | A photo with no description | The same photo tagged "cat" |
| Learning it allows | Unsupervised learning | Supervised learning |
The type of data you have decides the type of algorithm you can use. If you know the correct answer for each example, you can train a model by comparing its predictions with those answers. If you do not, the model must find structure on its own.
Labeling is often expensive and slow, because humans usually have to go through and tag each example by hand. In real projects, collecting and labeling data can take far more time and money than training the model. It is one of the biggest practical bottlenecks in machine learning.
The lecture shows three icons branching from "Machine learning algorithms":
| Family | Icon in the lecture | Data | How it learns | Example use |
|---|---|---|---|---|
| Supervised learning | A teacher at a blackboard | Labeled | Compares predictions with known correct answers | Predicting house prices, spam detection |
| Unsupervised learning | A colourful scatter of cluster nodes | Unlabeled | Finds hidden structure or patterns by itself | Grouping customers by purchasing behaviour |
| Reinforcement learning | A small robot and its environment | Rewards from interaction | An agent observes a state, takes an action, receives a reward or penalty | Game-playing AI, robot control |
Supervised learning matches the teacher analogy from lesson 1.1: the model learns by checking its work against the solution. It is the most widely used and most foundational family, and almost all of this course is about it.
Unsupervised learning has no answer key. The algorithm tries to discover groups or patterns without being told what they should be.
Reinforcement learning is quite different. An agent interacts with an environment: it sees the current state, chooses an action and receives a reward or penalty. Over time it learns which actions lead to better rewards.
A retail company has three ideas. Decide which family of algorithm each one needs.
Now suppose the company wants to sort 10,000 product photos into "good" and "damaged" but none of the photos has a tag. Before any supervised model can be trained, someone must act as the annotator and label at least part of the photos. Budget time and money for that step first.
Flashcards · 1.3 Labeled Data and the Three Groups of Algorithms
9 wordsRaw data with no information about what each example is or which category it belongs to.
Data where each example carries the correct answer, usually added by a human.
A person (or tool) who assigns the correct labels to raw data.
Learning from labeled data by comparing predictions with known correct answers.
Learning from unlabeled data by finding hidden structure or patterns without an answer key.
Learning by an agent that takes actions in an environment and receives rewards or penalties.
In reinforcement learning, the learner that observes a state and chooses actions.
The feedback signal an agent receives after an action, telling it how good the action was.
The practical problem that labeling data by hand often costs more time and money than training the model.