MAE

• loss converges to median of targets
• issues if gradient is close to 0
• $L(y,\hat{y})=n^{-1}{\Sigma }_i|{\hat{y}}_i-y_i|$