To transform the predictor variables so they have constant variance.

To map the expected value of the response variable, $\mu = E[Y]$, to the linear predictor $X\beta$.

To eliminate the need for maximum likelihood estimation.

To change the distribution of the errors from normal to non-normal.