Comparative Study of Caffe, Neon, Theano, and Torch for Deep Learning
arxiv.org
arxiv.org
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).
That said, I would start with Python based frameworks:
- Caffe is more for production, and Theano and family for experimentation.
- You can play with Theano or Tensorflow using Keras which abstracts them
- Python scientific stack makes it easier to visualize the results later
Caffe is C++, and the rest are Python (Neon, Theano and TensorFlow).
I know a guy who wrote a fast SAT solver in C, though called it with various slow bash scripts. I asked him why he uses bash, he mentioned it doesn't matter. In the end Microsoft's SAT solver beat him, exactly because they cared about such details too.