The InfoNCE loss in self-supervised learning