negative log-likelihood loss

The negative log likelihood loss (aka surprise, aka categorical cross-entropy loss):

$$L_i=-\log (p_{y_{i,k}})$$

where $p_{y_i}$ is the predicted probability of the true class label $y_{i,k}$, for the $i^{\text{th}}$ example.

This is a simpler version of cross-entropy loss, also called categorical cross-entropy.

References

Understanding softmax and the negative log-likelihood