Not good on "small datasets", well in theory, but in practice transfer learning given a similar training distribution can take you quite far.