Historically the XOR function has been the simple example that many ML algorithms can't handle. Just imagine a higher dimensional XOR with outliers, and you have a pretty good use case for DL with limited predictors.