For those interested in learning about the field: focus on the concepts, not the frameworks. This is true especially of machine learning, which is highly conceptual. I don't think these frameworks are suitable for people not familiar with the concepts. For example, I think it's a good exercise for beginners to implement a basic (no convolution) two layer neural net in plain Python using matrix operations in numpy.
Regardless, these frameworks are more similar than different (one exception is Caffe, which is very focused on computer vision and makes it really easy to do stuff in that area, but makes it really hard to do anything else). In the course of exploring the area, you will probably touch all of them at some point (when you're trying to use a researcher's code).