A method for training machine-learning models when the data contain arbitrary outliers. A modified Black-Rangarajan duality expresses robust losses from both M-estimation and deep-learning risk minimization through a single robust loss kernel, and the Adaptive Alternation Algorithm then trains the model by repeatedly minimizing a re-weighted ordinary loss while updating each sample's weight as an inlier probability, removing the need for manual parameter tuning. Students learn how re-weighting connects to robust losses and how robust kernels enlarge the region of convergence to outlier-free optima.
Unifies two separate traditions of training machine-learning models to be robust to outliers: M-estimation (common in robotics and computer vision) and robust risk-minimization losses (common in deep…