RNN and the gradient vanishing-exploding problem