Understanding Deep Learning Still Requires Rethinking Generalization

• unclear what variables trigger phase transitions
  • sometimes many variables can triger
• generalization depends on optimization and regularization