Two Wrongs Don't Make a Right: Combating Confirmation Bias in Learning with Label Noise