On the Computation of the Fisher Information in Continual Learning