A vision network trained via large-scale multi-task self-supervised learning.