You’re definitely not alone on that. It turns out that Logistic Regression is not as powerful as real “neural networks”, but we’ll do a similar thing in Week 4 and find that even the more powerful model we develop there also doesn’t do so well with new images. The problem is that because of the resource constraints of the environment here, we are forced to use a very small training set and that means we don’t get a very “general” model.
Here’s another thread of someone who noticed the same issue. Well, in that case, the picture is obviously a cat, but the algorithm says no.
Here’s another image that is pretty strange that gets recognized as a “cat”.