The advantages of lstm compared with general RNN

Lstm can only avoid rnn gradient disappearance (gradient vanishing), but not against the gradient explosion (exploding gradient). Gradient expansion (gradient explosion) is not a serious problem, usually by cutting the optimization algorithm can be solved, such as gradient clipping (if the gradient of the norm is greater than a given value, the gradient will shrink year by year).
The gradient tailoring method generally has two kinds: 1. One is when a gradient of the absolute value of a dimension is greater than a certain limit, the clipping is the upper limit.
2. Another is the gradient of the L2 norm greater than the upper limit, so that the gradient divided by the norm, avoid too large.
Lstm How to avoid gradients disappearing.