speed up LSTM with DNI

[zuowang]

I found the method described in Deepmind's DNI paper to speed up LSTM is promising. It showed that unrolling the LSTM and decoupling with DNI would get 2x faster. Can we implement it on paddle?

The paper is Decoupled Neural Interfaces using Synthetic Gradients

[reyoung]

We received an email for this feature request. The speed up is not the most amazing thing, it can also train a very very large model on multiple devices. Related to #141. But how to implement it in current Paddle framework is still under thinking right now.

In this paper, the datasets are quite small(cifar, mnist). Is this algorithm still work for a larger, more complex dataset with a lot of noises, which could make gradient/error prediction very difficult?

@pengli09 in our team is also very interested in this method.