Correcting Model Misspecification via Generative Adversarial Networks