Calling this 'universal learning' is a stretch (and in HN speak 'click-bait'). The paper only talks about a particular subclass of learning algorithms in supervised learning domain within PAC framework.<p>Specially they talk about learning algorithms that minimize some error function from training examples. That's not how learning happens in living organisms. Therefore calling it 'universal' is an unwarranted generalization, kind of like 'theory of everything'.<p>A more appropriate name would be 'Distribution-dependent Supervised PAC Learning' or something along those lines. It's a solid work which addresses a particular niche of a particular theory.