Training Methods

We show that carefully tuning standard deep learning components can achieve state-of-the-art performance on class-imbalanced datasets without specialized techniques.