When to use RNN or LSTM

Question

What is the difference between RNNs and LSTMs and when do we use either of these models?

Can you explain your question better? In what case are you talking? Seq2Seq applies for when you have encoder and decoder architecture. RNN, LSTM have that architecture so they are of that type. What is the question? — ウィエム, May 13 '19 at 05:32
yes i mean encoder and decoder architecture.can it RNN or LSTM, is it implementor choice — subho, May 14 '19 at 06:04

ウィエム · Accepted Answer · 2019-05-14T11:44:21.267

Not exactly. RNN, Recurrent Neural Networks works well for some tasks but because some issues arise from using RNN such as The Exploding Gradient or The Vanishing Gradient Problem, it became standard to use LSTM, which is a type of RNN that solves these problems.

Exploding/vanishing gradients happen when you are back-propagating through a long network and because you keep multiplying the gradient as you go, it can either grow as much and overflow, or it can become very very small that it will have no effect anymore.

LSTM solves this by using a mechanism called Gated Recurrent Unit. Because it solves the problems usually encountered by the standard RNN, LSTM is more popularly used.

It is is actually important for you to understand these topics so I would take a look at this blog, it explains them perfectly with graphs and examples. https://skymind.ai/wiki/lstm

When to use RNN or LSTM

1 Answers1