CNNs actually have a fairly major weakness: they are not orientation invariant, which means that if you were to feed the same image in, but upside down, the network is likely to not recognize it at all. One of the ways we can ensure this is not the case is to train the model with different rotations; however, there are better architectures that can solve this problem, which we will discuss later in this book.

They are also not scale invariant. Feeding it the same image much smaller or much larger makes it likely to fail. If you think back to why this is the case, it's because we are building the model based on a filter of a very specific size on a very specific group of pixels.

You have also seen that the model is very slow to train in general, especially on the CPU. We can get around this somewhat by using the GPU instead, but overall, it is an expensive process and can take several days to complete.