Good & Bad Examples
We know that machines learn from examples. But here is an important question nobody asked yet: does it matter WHICH examples you use? Spoiler alert — it matters a lot! In this lesson we will discover why good examples teach well and bad examples teach badly. And we will find out that YOU are in charge of which kind you give.
What Makes an Example Good?
A good example has three things going for it: 1. It is clear. The thing in the example is easy to see and recognize — not blurry, not hidden. 2. It is varied. Good examples show the thing from different angles, in different colors, in different sizes. This helps the machine learn the whole picture. 3. It is correctly labeled. The label matches what is really in the example. If you put a label that says "cat" on a picture of a rabbit, that is a disaster — the machine will learn the wrong lesson. When your examples have all three of these, the machine learns accurately and fairly.
Your machine will only ever be as good as the examples you give it. Great examples in = great learning out. Messy or wrong examples in = confused and wrong machine out.
Here is a story to show what can go wrong. Sam wants to train a machine to recognize "healthy food." Sam is in a hurry, so they only grab pictures of green vegetables — broccoli, peas, spinach. Not a single fruit, not a grain of rice, not a bowl of soup. The machine trains on those examples and builds a model: healthy food = green and vegetable-shaped. Now Sam shows the machine a picture of a ripe mango. The machine says: NOT healthy. But mangoes ARE healthy! Sam's examples were not varied enough — they missed a huge part of what healthy food really looks like. Sam also accidentally labeled a picture of candy as "broccoli" (they clicked the wrong button). Now part of the machine's brain thinks candy is a vegetable. Oops.
Match each example quality to what it means.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
Scientists use the word bias to describe what happens when your examples only show one side of something. If you train a machine to recognize "faces" using only photos of people who look like you, the machine might struggle with faces that look different. That is a bias — a lopsided view caused by lopsided examples. Bias can cause real harm when AI systems make decisions about people. That is why the people who build AI work very hard to make sure their examples are varied and fair.
A wrong label is worse than no label at all. If you are ever helping to label examples for a machine, slow down and check carefully. One wrong label can teach the machine something false.
Sam's machine thinks mangoes are not healthy. What went wrong?
What is bias in AI training?
Sort the Examples
- Imagine you are building a machine that recognizes 'round things.'
- Write down five examples you would include. Make them varied!
- Now write down one BAD example — something that is NOT round but might be mistaken for one.
- Write down one WRONG LABEL — a round thing you accidentally call by the wrong name.
- Look at your list. Which examples are good? Which are bad?
- Talk about it: how would the bad examples confuse the machine?