ML Seminar: Theoretical Analysis of Self-Training with Deep Networks on Unlabeled Data (Colin Wei, Stanford University)