ref：https://www.mathworks.com/content/dam/mathworks/mathworks-dot-com/images/events/matlabexpo/cn/2018/deep-learning-for-non-image-applications-cn-2018.pdf
ref:https://karpathy.github.io/2015/05/21/rnn-effectiveness/
If training vanilla neural nets is optimization over functions, training recurrent nets is optimization over programs.
如果说训练平凡神经网络是对函数的优化，那么训练循环网络就是对程序的优化。
RNN是 Turing-Complete
