Stochastic Re-weighted Gradient Descent via Distributionally Robust Optimization

We design simple, explicit, and flexible per-sample re-weighting techniques for learning deep neural networks inspired by distributionally robust optimization.