Sequential data only?

An important issue is that in many deep learning methods we assume that the input and output data can be treated as independent and identically distributed, normally abbreviated to iid. This means that the data we use can be seen as mutually independent.

This is however not the case for most data sets used in RNNs since we are dealing with sequences of data with strong inter-dependencies. This applies in particular to time series, which are sequential by contruction.