New paper w @BDRoads, "Learning as the Unsupervised Alignment of Conceptual Systems". Supervised learning tasks can be solved by purely unsupervised means by exploiting correspondences across systems (e.g., text, images, etc.). 1/5 nature.com/articles/s4225… 
How can supervised tasks be solved without supervision? Concepts have unique signatures that hold across unsupervised systems. E.g., The words pen and pencil pattern similarly in text corpora and are also visually similar. Likewise, both are different to lions, cars, etc. 2/5
Unsupervised alignment only gets better with more concepts and systems. Each concept's signature becomes more unique the more concepts there are in each system (see below). Each system provides a slightly different viewpoint on reality such that adding more systems helps too. 3/5
From a developmental perspective, children's early concepts form readily aligned systems, which in principle allow for purely unsupervised learning without image-label co-occurrence, child directed speech, etc. 4/5 nature.com/articles/s4225…
We've shown that supervised learning can be solved by purely unsupervised means. The big challenge now is to devise algorithms to efficiently exploit this information across conceptual systems. We (and I hope others) are working to meet this challenge. 5/5 nature.com/articles/s4225…
Oh, if your institution doesn't have access, it's all here, arxiv.org/abs/1906.09012
