How does this simulation work?
The program does not analyse image pixels. It uses predefined data about species, background and ear shape. It compares two prepared rules and selects the one that makes fewer errors on the training set.
Tie-breaking rule: Both rules (background and ears) work without errors on the initial examples. Algorithms commonly use a predefined tie-breaking rule. In this program, ties are resolved by choosing the background rule.
About the collar: The collar is not evaluated in this simulation. The program only compares the background and ear-shape rules. In a real neural network, a prominent element like this could become an unintended misleading feature.
Class discussion questions
- Why did the model achieve 100% in the first stage but fail in the second?
- Do more similar examples always help a model learn better?
- What might happen if we trained AI to recognise apples and pears, but all training images showed apples on white plates and pears on blue plates?