MAE

If we assume the error $\epsilon = y - \hat{y}$ follows a normal distribution, the log likelihood is $\log p(\epsilon) = -\frac{|\epsilon|}{b} +\text{const}$. Maximizing this leads to the derivation of MAE.