How does a model recognise an object?

The model does not see a cat. It analyses thousands of pixels and looks for patterns that were often present in the images it was trained on: pointed shapes (ears), stripes, round eyes, a specific silhouette.

This is why it can make mistakes:

• a dog with pointed ears may be classified as a 'cat',
• a cat seen from behind or in the dark lacks the usual patterns,
• if all training images were indoors, a 'sofa' might become a cue — and a sofa without a cat could be classified as a 'cat'.

Experiment with Teachable Machine (free, in the browser): train 'cat' / 'not cat' with 20 images each. Then test:

1. A normal cat → high accuracy.
2. An upside-down cat → lower.
3. A toy cat → ?
4. Just an empty sofa → sometimes 'cat', if many training images had a sofa.

Note what tricked the model. This shows what the model has actually learned.