So I think it's clear that this isn't a turn I would make, though a lot of people would: "the key capacity that computers and robots should possess in order to emulate human cognition and (ethical) consciousness is the capacity to learn and apply 'coherent webs-of-theories'... where classic AI has been, in essence, 'data-driven', the greatest leap forward would be 'theory-driven' AI." The important concept discussed in this paper (36 page PDF) is the Conscious-through-Monitoring-through-Theories (CMT) Model, and it's used to argue that "all truly human-like AI will somehow have to be theory-driven, rather than purely data-driven, as is now the case. A similar claim has, in essence, been made by cognitive scientists." There's some good discussion about whether neural networks could develop these theories themselves, including a recognition that there is probably no general model for theory generation.
Today: 5 Total: 97 [Direct link] [Share]
