Course 0 · Lesson 05/10 · Interactive guide
How AI Learns From Examples
In AI, “learning” means adjusting a model from data or feedback. It does not mean thinking or understanding in the way a person does.
- Lesson
- 05 / 10
- Time
- About 9 minutes
- Level
- Beginner · 0/5
- Prerequisite
- AI, ML, Deep Learning, and Generative AI
No equations required yet
“Learning” means adjusting to make fewer errors on a defined task
Think about practising a throw into a basket. You throw, see where the ball lands, adjust the force or angle, and try again. A model follows a related loop: receive examples, produce an output, receive information about the result, and adjust internal values.
Change the feedback
Same toy task, different signal: how will it learn?
Imagine that we are training a system to handle pictures of fruit. Choose what it receives after one attempt, then watch the learning mode and next step change.
- It receives
- Images paired with correct labels
- It compares
- Its “orange” prediction with the true “apple” label to calculate error
- Next step
- Adjust model values to reduce error, then try the next labelled examples
- Limitation
- People or another process must provide labels that are accurate and cover the task
The actual training loop
The system does not receive the answer to every future case
- 1
Receive training data
Examples may have labels, have no labels, or arrive with reward signals, depending on the method.
- 2
Produce an output
Using its current values, the model calculates a prediction, grouping, or action.
- 3
Measure error or feedback
One task compares an answer key, another measures structure, and another receives reward after actions.
- 4
Adjust and repeat
The training method changes values in small steps and repeats, then evaluation must use data that did not make those adjustments.
Vocabulary to keep
The three methods differ in the signal they receive
Supervised
Examples come with target answers
Grounded examples: emails labelled “spam/not spam” train a classifier, while leaf images paired with disease names can train a screening model.
Trade-off: labels can be expensive, wrong, or incomplete for new cases.
Unsupervised
Find structure without target answers
Grounded examples: group similar shopping patterns for exploration, or flag data points far from a cluster so a person can inspect them.
Trade-off: a mathematically tidy group might not match the category people need.
Reinforcement
Learn from reward after actions
Grounded example: a controller in simulation tries actions and accumulates reward for reaching a target without hitting obstacles.
Trade-off: a poorly designed reward can invite shortcuts, while real-world trial and error can be unsafe.
Remember three things
Find the feedback before asking how an AI learns
The basic loop is examples → output → feedback → adjustment, followed by repetition.
Answers, no answers, and rewards lead to different learning and evaluation methods.
Training success does not guarantee success in a new setting. Test beyond training data and monitor after deployment.
Three-question check