Deep learning versus linear models across all input distributions

Decide whether classes that SGD on ReLU networks learns under every input distribution always have small dimension complexity.