For the first lesson, I experimented with a bunch of images in the cat classifier, and while it was very good at identifying cats as cats and other mammals as not cats, other results were… interesting.
This is a very common issue that people new to ML run into. A classifier will only work correctly on data that is similar to the training data. If you give it data that is way different (known as out-of-distribution or OOD data), the results won’t make any sense.
You need to have a few extra examples of completely different images in the dataset. If you only have animals in the dataset, then it’s probably looking for very specific features that may be present in other images but not in the current context. If you have some more diverse images, then the model can better learn the concept of a cat.