cross-entropy

for training, minimize $L=-logP(y_{true})$

goal: maximize probability for the correct class

softmax classifier (multinomial logistic regression)