🎒 What you need

about 30 beans, maize grains or groundnuts, ideally a mix of good and damaged
a strip of paper marked in finger-widths as a ruler
paper and pen, or chalk and a flat stone

✋ Do this

Pick one grain. Write down three numbers about it: length, width, and number of dark marks. That row of numbers is called a feature vector.
Do this for 20 grains. Mark each row G for good or D for damaged. You now have a dataset — a real one, that you made.
Draw two axes on the ground: length across, width up. Plot every grain as a dot, writing G or D beside it.
Stand back and look. Do the G dots cluster away from the D dots? If they do, your features are useful. If they are all mixed together, they are not.
Now throw away width and try 'number of dark marks' instead. Re-plot. Which pair of features separates them better? Point at the better picture.
Write one sentence: 'The feature that separates good from damaged best is ___.' That sentence is the whole of feature selection.

💡 Why it works

Every image classifier begins by turning a picture into numbers. Modern networks learn which numbers to extract instead of being told, but the principle you just used is unchanged: if your chosen features do not separate the classes, no amount of clever maths afterwards will save you. Choosing what to measure is the highest-leverage decision in the whole pipeline.

🔥 Challenge

Find a pair of features where the two groups overlap almost completely. Then explain why that pair is useless, using your own plot as evidence.

📖 New words

featureone measurable property of a thing, expressed as a number

feature vectorthe full list of numbers describing one example

dataseta collection of examples, each with its features and its true label

📊

Feature Finder

Tap when you have finished this module.