MinimalRNN: Toward More Interpretable and Trainable Recurrent Neural Networks