Adam vs SGD Generalization Gap

Adam optimizer typically converges faster than SGD but often generalizes worse on held-out data. Explain the mechanism behind this generalization gap.