🌽🌽🌽🌽🌽 | 🌾

Five samples pulled from the same easy-to-reach heap, and one lonely sample from everywhere else.

🎒 What you need

access to a real pile of something varied — grain, mangoes, stones, leaves
paper for tallies
a friend to sample separately from you

✋ Do this

Grab 20 items from the pile without thinking. Tally how many are good and how many are damaged. That is your quick sample.
Now dig properly: take items from the bottom, the edges, the middle, the bit hidden behind. Another 20. Tally again.
Compare the two tallies. They will differ. Say out loud why: the easy grab took what was on top, and what is on top is not random.
Write down three real ways a dataset could get skewed: only photos taken in bright sun, only from one farm, only of the crop somebody already suspected was sick.
Now the serious one. Imagine an app trained only on tidy photos from one region, then used across a whole country. Write down who gets a wrong answer and what it costs them.
Design a fair sampling plan for your pile in five written steps. Then follow it and see whether your tally changes again.

💡 Why it works

Sampling bias is not a rare accident, it is the default outcome of convenient data collection. A crop-disease model trained on well-lit photos of one variety from one area can drop sharply in accuracy on a different soil, a different variety, or in harmattan dust. The fix is boring and unglamorous: go and deliberately collect the awkward cases nobody else bothered with.

🔥 Challenge

Deliberately build a biased dataset that scores brilliantly on its own test and would fail badly in reality. Then explain the trick to someone — and notice how convincing your fake score looked.

📖 New words

sampling biaswhen the examples you collected do not represent the real world

representativea sample that reflects the true variety of what you will meet

generaliseto work well on new data you have never seen

⚖️

Bias Hunter

Tap when you have finished this module.