Experiment: biased training data and its consequences

The model learns the patterns present in the data — including those you did not intend.

If you train a classifier on 'defective/okay' with 90 images of defective items on a white table and 90 images of okay items on a dark mat, the model learns 'white surface = defective'. A defective item on the mat is classified as 'okay'.

This is called bias in the data. The model is not logically flawed — it found the simplest correlating pattern that separated the categories. The solution is almost always better data, not a more complex model.

Experiment in Teachable Machine:

1. Class A: 20 images of invoices — all scanned against a white background.
2. Class B: 20 images of receipts — all photographed against a dark background.
3. Train. Test with an invoice on a dark background. What does the model say? Often 'receipt'.
4. Repeat with mixed backgrounds for both classes. Test again.

Write down: accuracy on standard images, accuracy on 'swapped background', before and after. This is a controlled experiment: only the data changed.