Project C: collect data and train a model

Project C: the whole chain — from an idea to a measured model.

1. Pick a question that can be answered by classification (which kind of waste? which hand sign? which sound?).
2. Collect the data yourself: at least 3 classes × 40 examples. Enter it in a spreadsheet: file, class, who collected it, the conditions.
3. Split it: 30 training / 10 test per class. The test set is not touched until the end.
4. Train (Teachable Machine or similar).
5. Measure on the test set. Count per class — one class can be much worse than the average.
6. Report it as an experiment note plus the spreadsheet.

A confusion table — the best way to see which mistakes the model makes:

actual \ guessed	paper	plastic	metal
paper	9	1	0
plastic	2	7	1
metal	0	1	9

Accuracy: (9+7+9)/30 = 83 %. But plastic is only 70 % — and gets confused with paper. What in the data could explain that? (Transparent plastic against a white background?) That is your improvement.