C++ wrapper for the MKLDNN primitives. pre-requisites: MKL DNN primitives: available in the free community edition (maybe, TODO, check if true) cmake CXX of some kind (that supports openmp maybe)

build instructions: mkdir build cd build CXX=g++ cmake .. make -j

Net - Container for layers. The user should only have to call fwd(), bkd() and update() member functions of this class. Currently only sequencial networks.

• potentiallly graph and sequential supports Layer - Base class for various neural network layers. Conatins the weight and gradient.
• conv
• relu
• maxpool Optimizer - optimization algorithm object to be passed into net.update(). Loader? - should be able to load a network from some sort of configuration file.

Tasks a user will probably do: define a network come up with some hyper parameters give it data press go to train look at learning curve data modify the network or hyper parameter press go to train again save network weights save network go do something else load network load network weights individually run forward look at array of losses individually run backward individually run update