[johnsutor]

This is the premise behind some more recent self-supervised learning papers (see https://arxiv.org/abs/2303.03307). Turns out, it's a solid analog for classification tasks.