Machine Learning library comparisons for testing brain architectures

There are already nice deep-learning/ML software comparisons out there:

1. https://en.wikipedia.org/wiki/Comparison_of_deep_learning_software
2. Comparative Study of Deep Learning Software Frameworks, https://arxiv.org/abs/1511.06435
   Outdated, but as of v3 Mar 2016, Theano was the most extensible, Torch was fastest, while Tensorflow was extensible and best for multi-GPUs, but not as fast as the others. [Given this, Keras seems a good choice as it supports both Theano and Tensorflow backends, supports fast prototyping, is modular, and claims to be easily extensible and suitable for advanced research.]
3. http://www.pcquest.com/15-top-open-source-artificial-intelligence-tools/ Magazine style brief write-ups about 15 frameworks.
4. http://deeplearning.net/software_links/ List of deep learning frameworks.
5. https://github.com/josephmisiti/awesome-machine-learning
   Curated list of broader machine learning frameworks, not just deep learning.

Here, we compare free deep-learning/ML and bio-realistic libraries/frameworks for suitability of use in archibrain.
Particularly, we want
0. python / julia support

1. arbitrary neurons (GRU, LSTM, spiking, reLU)
2. auto-differentiation of arbitrary functions
3. ease of implementing different learning algos (not just backprop) - typically bio-realistic simulators only allow synaptically-local and online learning rules, while ML/deep-learning frameworks do backprop over multiple layers.
4. can use GPUs (and CPUs) across multiple nodes (openmp/cuda/opencl/multi-node?).
5. ease of implementing both RL and supervised learning tasks.

library	py/ju?	arb nrns?	auto-diff?	arb learn algo?	multi-node GPU/CPU	tasks
tensorflow	py	?	yes	?	cuda,multi-node	?
mxnet	py/ju	?	yes	?	openmp,cuda,multi-node	?
scikit-learn
theano	py	?	yes	?	openmp,cuda,(dev-opencl),multi-node	?
keras	py	?	yes	?	theano-backend,multi-node	?
torch	no (lua)
cntk	py	?	yes	?	openmp,cuda,multi-node	?
caffe	py	?	?	?	cuda,...	no
brian	py	yes	no	only local,online	?	no
nengo	py	lif,lifrate, hack others	no	PES,etc, hack other local,online	opencl,not multi-node	no
numenta/nupic
paddlepaddle

For the moment, let's go with keras and update the above table as we learn/play more, while following an agile software development process.