Robust NTM training

[chanhou]

The tanh in the hidden layer is unnecessary since the head param need the result of hidden layer instead of transform by tanh.
Secondly, by using lower learning rate can avoid nan problems, and in my experiment, about 60000 iteration will be converged to a better result.