There are now lots of deep learning toolkits, maybe over 100. A new one is announced every other week, it seems. And each one wants to become the "standard" f

There are now lots of deep learning toolkits, maybe over 100. A new one is announced every other week, it seems. And each one wants to become the "standard" for deep learning, trying to win the hearts and minds of developers, developers, developers! So far, I'd say these are winning the competition for mindshare: 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.