Introduce momentum to DNN

[gyeongin]

A lot of researches articulate that momentum may accelerate job convergence of DNN. We should introduce momentum to our DNN codebase.

[jsjason]

• (Apparently) the first paper to propose the momentum method: Polyak, "Some methods of speeding up the convergence of iteration methods", 1964.
• A good comparison of the classical momentum method and the Nesterov version: http://www.cs.toronto.edu/~hinton/absps/momentum.pdf, Section 2
• Very short version: http://cs231n.github.io/neural-networks-3/#sgd