Computer science
Training data & fairness
A machine only learns what is in its data, for better or worse.
What you need first
Your video app suggests clips, the camera detects faces, and the voice assistant understands your questions. All of that works because machines have learned from examples. But what happens if the examples are picked in a one-sided way? Then the machine takes on exactly that one-sidedness, and real people are the ones who feel it.
The machine only learns from data
A learning program is given many examples with the correct answer, say a thousand photos labelled dog or cat. From these it finds patterns that it uses to sort new cases. You can picture this as a dividing line: everything on one side counts as dog, everything on the other as cat. The key point: the machine does not understand what a dog is. It only knows the patterns from its .
The machine learns a boundary from examples. A perfect split is often impossible.
Skewed data, skewed results
Imagine a voice app trained almost only on adult voices. It will struggle with children's voices, simply because it has hardly seen any. Or a goalkeeper AI trained only in sunshine fails in the rain. Such distortions are called . They rarely happen on purpose, but because groups or situations are missing when the data is collected. The machine cannot notice this gap itself, it takes its data for the whole world.
Humans stay responsible
That is why machine results must be checked, by humans. If a model reports 100 percent accuracy, that is more suspicious than impressive: maybe the test was too easy, or the model memorised the answers instead of learning real patterns. Good teams therefore test with new, diverse data and ask: who does the system work poorly for? And when an AI makes an unfair decision, it is not the machine that carries the responsibility, but the people who built and deployed it.
Exercises
0 of 6 solvedTime to try it yourself. You can't break anything, every attempt counts.
Where does a learning program get its knowledge from?
A training dataset contains 90 dog photos and 10 cat photos. How many photos is that in total?
What is likely to happen if a model was trained almost only on dog photos?
When whole groups or situations are missing from the training data, a distortion arises that is called ….
A model recognises 45 out of 50 dog photos correctly. What percentage is that?
Match each case to what follows from it.
Where this leads