#focal-loss

Focal loss replaces the cross-entropy $-\log p_j$ by $-\alpha (1 - p_j)^{\gamma} \log p_j$, where $p_j$ is the predicted probability of the true class. With $\alpha = 1$ and $\gamma = 2$, by what factor is the loss of a well-classified example with $p_j = 0.9$ smaller than its cross-entropy?