Elucidating robust learning with uncertainty-aware corruption pattern estimation