This was written in 2015. I don't recommend Theano to learn DL. Keras is probably the best option (which runs on top of Theano or TensorFlow), and if you want t

This was written in 2015. I don't recommend Theano to learn DL. Keras is probably the best option (which runs on top of Theano or TensorFlow), and if you want to look under the hood, use TensorFlow directly. TensorFlow is probably the most popular DL framework today. Two other frameworks have recently appeared: Caffe2 and PyTorch. In my opinion, PyTorch is the most intuitive to use (similar to coding in Numpy). However, it does not support Windows yet. It's a Python version of the popular Torch DL framework, and is probably the most flexible tool for experimenting with DL today. Most bleeding edge research published today seems to be using Torch (should be easy to reproduce in PyTorch). Caffe2 is the newest addition. It builds on a popular Caffe framework, so should be fast and scalable. It's still very new though, so expect issues.