Textbook of AI

Abstract. Introduces focal loss, a reweighted cross-entropy that addresses extreme class imbalance in dense object detection. The loss multiplies standard cross-entropy by a modulating factor $(1-p_t)^\gamma$ that down-weights well-classified easy examples and focuses gradient on hard, misclassified ones. With $\gamma=2$, focal loss reduced the gap between one-stage detectors (RetinaNet) and the more accurate two-stage detectors (Faster R-CNN). Focal loss has since become a standard ingredient in long-tailed classification, dense prediction and any setting where the easy-positive class swamps the gradient.