hard example mining

**Hard Example Mining** is a **training strategy that focuses the model's learning on the most difficult (highest-loss) examples** — instead of treating all training samples equally, hard mining identifies and over-represents the challenging examples that drive the most learning. **Hard Mining Methods** - **Offline**: After each epoch, rank all examples by loss and create a new training set biased toward high-loss examples. - **Online**: Within each mini-batch, compute loss on all samples but backpropagate only the top-K hardest. - **Semi-Hard**: Focus on examples that are hard but not too hard — avoid outliers and mislabeled data. - **Triplet Mining**: For metric learning, mine the hardest positive/negative pairs. **Why It Matters** - **Efficiency**: Easy examples contribute little to gradient updates — hard mining focuses compute where it matters. - **Imbalanced Data**: In defect detection (rare events), hard mining ensures the model focuses on the rare, important cases. - **Convergence**: Hard mining accelerates convergence by prioritizing informative gradient updates. **Hard Example Mining** is **learning from mistakes** — focusing training effort on the examples the model finds most challenging.