Just because I've never heard of it in my 3-decade experience with NNs doesn't mean it can't be done.
For regression and classification the easiest case would be if there were the same number of examples. Then inputs and targets could be vertically concatenated ( e.g., [input1 ; input2 ].
However in most cases, separate training would be faster if not more accurate.
It can also be done for timeseries. However, the delays have to be the same.
I see no advantage and do not recommend it.
Hope this helps.