Give the AI a few examples of two different things. Watch it train live in your browser. Then test it. See exactly what training data means — and why more examples produce better results.
// Train the model first to see the explanation here.
The model can only learn what you show it. If your Cat A images are all orange cats, it may fail on black cats. That is bias — baked in from the data.
3 images gives a weak model. 20 images gives a much stronger one. Real AI models train on millions. Scale is the difference.
This lab uses MobileNet — a model already trained on millions of images — as a starting point. Your training only adjusts the final layer. That is why it works with so few images.
Even a 95% confidence score means 1 in 20 predictions will be wrong. Real AI systems make wrong predictions. High confidence wrong predictions are particularly dangerous.
Try an image that is somewhere between your two classes. The model must still pick one — it cannot say "I don't know." That is a fundamental limitation.
You cannot see exactly why it made a decision. It found patterns in the features — but which patterns? This is the explainability problem in real AI systems.