Course 0 · Lesson 03/10 · Interactive guide

Data Is AI's Food

Models learn patterns from the examples we provide. Quality is not simply more data; it means enough coverage, balance, and clean signals for the task.

Lesson
03 / 10
Time
About 8 minutes
Level
Beginner · 0/5
Prerequisite
Lesson 02 · How AI Differs From Regular Software

01 · Start with a box of examples

A model only sees what is represented in training

To classify road signs, we provide example images with correct answers. During training, the model is adjusted so patterns in those examples help it predict signs in new images.

01Examples

Many sign images

02Dataset

Collected with answers

03Training

Adjust from errors

04Prediction

Try an unseen image

Do not take the title literally: data is not food, and a model does not “eat” or understand images. It is an analogy for remembering that training examples affect the behavior we get.

02 · Analogy and real cases

Practicing from an incomplete study guide

A dataset is like a set of practice questionsA thousand questions do not always help if they are all the same type, contain bad answer keys, or omit topics on the real test. A smaller, well-checked set with useful coverage may be better.

Road signs

No night or rain

A model trained only on bright daytime images may be brittle when reflections, rain, or darkness alter the scene.

Message filter

Language is not covered

A formal-text-only dataset may struggle with abbreviations, spelling variants, or languages that real users send.

Product images

Classes are imbalanced

If shoes greatly outnumber bags, a model may lean toward the class it saw much more often.

03 · Adjust a toy dataset

How coverage, balance, and noise change reliability

This learning toy is not a real model-accuracy score. Each slider also works with the keyboard arrow keys.

Toy dataset quality signalBrittle

The model may memorize daytime conditions or the majority class while receiving wrong training signals. Add missing cases, check balance, and review labels before training.

04 · Terms and limitations

Data that fits the task matters more than “a lot”

Coverage

Whether examples include the conditions and groups the system will actually meet: light, angle, language, device, and rare cases.

Label

The answer used in supervised learning, such as “stop sign.” A wrong label sends the model a wrong training signal.

Noise

Corruption, broken images, inconsistent text, or unreliable labels that obscure the pattern needed for the task.

Imbalance

Some groups have far more examples than others, so an overall score can look good while a less common group performs poorly.

Trade-off: collecting and reviewing more data takes time and money and can create privacy risks. More is not an automatic answer. Define the task, collect only what is justified, check usage rights, and evaluate important conditions separately.

05 · Remember these three

Before saying “just add more data”

  1. 01

    Examples define the training boundary. Missing conditions can become brittle spots in real use.

  2. 02

    Quality has several dimensions. Check coverage, balance, labels, and noise—not only the count.

  3. 03

    Test away from training data. Inspect important groups and conditions instead of trusting one overall score.

06 · Knowledge check

Three questions with visible explanations

1. You have ten thousand sign images, all taken in daylight. What is the main problem?

2. What is label noise?

3. Why is “more data” not always the answer?