*What is Thought?* argues that the road to general intelligence comes by evolutionary computing via Occam's razor.
20+ years of CS research has shown that concept learning comes from a formalized version of Occam's razor: a compact enough neural net, say, or program that is consistent with enough examples of a concept is so constrained that it will generalize to new, not yet seen, examples of the concept. (And there are also pretty good, though not quite as rigorous, arguments that this is roughly speaking the only route to concept learning.) *What is Thought?* extrapolates this to general intelligence, arguing that "understanding" comes from having a compact enough program that that behaves properly-- for example producing a very compact genome that knows how to construct a mind that knows how to do the right kinds of calculations. Such a program can only be so compact because it has learned how to exploit the underlying structure of the world. It will be compact by virtue of code reuse, i.e. have a modular structure with modules corresponding to real concepts, reusing the modules in combinatoric ways to get great power. What it means to "exploit the compact underlying structure" of the world and why this is equivalent to understanding is discussed at some length and detail. Roughly speaking, if you are seeking a general intelligence, you are seeking understanding, and, in my view, you won't get there without exploiting Occam in this way. Some of you seem familiar with my Hayek results, which are something of an illustration. Hayek managed to exploit the compact underlying structure of Blocks World to solve it by producing a relatively compact modular program. I argue that evolution extracted computational biases which cause us to automatically develop much of this modular structure, enough so that our learning and reasoning are constrained to deal with meaningful concepts (i.e. correspond to modules exploiting real underlying concepts in the world). As I discuss at some length, this is very different than hand coding in knowledge, e.g. building a big expert system. Complexity theory has told us that finding compact, Occam representations is an NP-hard problem. It is too hard to be solved by human programmers anymore than a human can, by inspection, solve a huge Travelling Salesman Problem. AI programs typically do not "understand" because they do not exploit Occam's razor in the way natural intelligences do. I discuss how computer science approaches, for example computer chess programs, can exploit structure in ways different than humans which are still interesting. But achieving anything like general intelligence will, in my view, require a massive evolutionary effort. We do not now have, nor will have in the forseeable future, computational resources comparable to those evolution threw at the problem, so it is not at all obvious we can succeed. The Hayek work showed that we can impose rules that facilitate fast evolution, so there is some hope, but I am nonetheless not optimistic. Please pardon any unclearness or outright confusion in these remarks, which I'm just typing in flow of consciousness in reaction to seeing there was some discussion of my previous post. Obviously, the book *What is Thought?* is much, much clearer on all these points and contains a lot of other stuff besides. There's now a website with more information: http://www.whatisthought.com Some of my other papers, including Hayek papers, are posted at http://www.whatisthought.com/eric.html ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
