This whole feud thing with Ben and Matt is a little awkward and strange, but i guess that happens on email.
Anyway, isolated recognizers are getting quite amazing these days, but they all suffer from a lack of generality. In particular, they focus, of necessity, on features specific to their problems, instead of the kind of things we use--features learned over an entire lifetime in addition to specific ones learned for specific problems. And people have a completely separate, I'm going to say, system of language that adds communication ability in there, with surely some kind of benefit. And for sure, we have other special sorts of modules that help out. Plus, yes, other information sources like color and motion. Those speak to improving accuracy. Somewhat related topic. Turns out we also have algorithmic improvements that can improve over straight brain emulation in terms of computational efficiency. SVMs do better than neural nets in recognition problems, but they aren't so flexible in taking just any old inputs available.. The processing used by complicated connectionist operations needed for massively parallel neural networks can be quite bit simpler with matrix manipulations and linear algebra. That just isn't a biological option, but it's available to AGI designers. And while I'm not a big fan of customized data structures to handle some internal process communications in an AGI system, if a sufficiently flexible and effective design is worked out, there plenty of room for optimization over the limited brain mechanisms. Plus, who knows how you can balance the reliability and speed compromises the brain had to make with the incredible accuracy, reliability, and speed of computers. So, I would suggest that complexity calculations only based on brain analogies are simply not good enough. That somewhat extends to isolated recognizers. andi Matt: >> Anyway, I would like opinions on the computational complexity of human >> vision. Specifically, how would you optimize Google's cat face >> recognizer and bring it up to human level? >> http://128.84.158.119/abs/1112.6209v3 > Ben: > I wouldn't try to optimize that algorithm; I would take a different > approach that couples a visual hierarchy with a structurally and > dynamically richer cognitive system... > > But I'm not going to try to pack the details of my AGI thinking into an > email... ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
