Continual Learning in Deep Networks: an Analysis of the Last Layer