Self Supervised learning approximates commonsense

Unsupervised learning is a misnomer, because often labels are derived from the data itself. Performing self-supervised learning using a huge amount of data might be one of the few feasible methods of endowing an artificial agent with robust commonsense. The reasoning is that human performance and few-shot learning is possible only with the epistemological background of commonsense acquired during development.

Resources

Backlinks