Depends on what kind of prototyping. At the current state of neural networks DSLs are mainly helpful if you want to tune a network architecture for well-established tasks like image classification for the Imagenet dataset.

Outside of that I see more dynamic alternatives used much more.