New training method ignores corrupted data points

Original: Quantile-based Loss Filtering for Outlier-Robust Stochastic Gradient Descent

Source: arxiv.org ↗