TensorFlow – Google’s latest machine learning system, open sourced for everyone
googleresearch.blogspot.com
googleresearch.blogspot.com
1 Google's TensorFLow: looks like Theano 2.0 with easier asynchronous computation (e.g., multi-GPU), a friendlier API, insanely good documentation, lots of tutorials for beginners, and lots of sample code. Comes with Python bindings.
2 UC Berkley's Caffe: a Google team used it for GoogLeNet last year. Probably the best choice if you work with images and want high performance in a ready-to-deploy package. Comes with Python bindings.
3 U of Montreal's Theano: it's more powerful and more widely used than people realize. For example, Keras, Lasagne, and Blocks are essentially friendly "layer-by-layer" APIs built atop Theano's "calculation graph" API. Comes with Python bindings.
4 FaceBook's Torch: people who've used it like it a lot, but its reliance on Lua (Lua!?) prevents it from being more widely adopted. FaceBook people use it because that's where Mr. Le Cun works.
But by far the biggest winner is Python, which is now clearly the #1 scripting language for deep learning.