What makes for good views for contrastive representation learning