ML types: Supervised, Unsupervised and Reinforcement

Summary

The video explains four types of machine learning through an analogy about how people learn and an example of organizing books.

Machine learning teaches a machine to learn from data—information and examples—much as people learn from experience. The key distinction is whether learning uses correct-answer labels, feedback from actions, or unlabeled data.

1. Supervised learning

• The learner is given labeled examples: data paired with the correct answer.
• In the book example, a manager identifies some books as fiction or non-fiction.
• The learner uses those examples to classify books it has not seen before.
• The goal is to produce correct answers for new data. This resembles learning under a teacher’s guidance.

2. Reinforcement learning

• The learner is given a goal, but not necessarily the correct action or answer for each situation.
• It acts, observes the consequences, and adjusts its behavior to achieve the goal.
• In the book example, different ways of arranging books lead to different sales. Sales provide feedback about which arrangement works better.
• Learning to ride a bike works similarly: actions that lead to falling discourage those actions.

The basic reinforcement-learning cycle is:

1. An agent takes an action in an environment.
2. The action changes the agent’s state, such as its location.
3. The environment may provide a reward or a negative outcome.
4. The agent uses these outcomes to guide future actions.

A maze-running animal illustrates the same idea: it learns which routes lead to rewards.

3. Unsupervised learning

• The learner receives unlabeled data and no instructions about the correct answer.
• Its task is to discover patterns or groupings within the data.
• For example, books might be grouped by genre, alphabetically, or by color without being told which arrangement is correct.
• The video also compares this to people picking up patterns in facial expressions without being explicitly taught what each expression means.

4. Semi-supervised learning

• This approach combines supervised and unsupervised learning.
• It uses a small amount of labeled data alongside a larger amount of unlabeled data.
• In the book example, labeled books provide initial guidance on fiction versus non-fiction, while the many unlabeled books can help reveal additional patterns.
• Those patterns may help reassess uncertain classifications. For example, noticing that an author’s other books are almost all non-fiction may cast doubt on classifying one of that author’s books as fiction.

Main takeaway

The speaker encourages viewers to understand machine-learning concepts intuitively rather than relying only on memorized definitions. The four approaches differ mainly in the kind of feedback or data available: labeled answers, rewards from actions, unlabeled patterns, or a combination of labeled and unlabeled data.

Speaker and source

• Narrator: Daniel Gibbon, the video’s presenter. The subtitles contain one continuous speaker; no other speakers are featured.