ML Seminar: Self-training Avoids Using Spurious Features Under Domain Shift (Yining Chen, Stanford University)