From: Ben Goertzel (email@example.com)
Date: Sat May 11 2002 - 10:26:24 MDT
> Okay, thanks; that does help. Next question: What is it about Novamente
> that makes this problem space more tractable for Novamente than for
> traditional applications? Is there any way you can describe this in human
A very abstract way to say it is..
1) greedy search doesn't work for this kind of application, you need to find
complex patterns whose component parts are not simple patterns
2) evolutionary search won't work because the problem space is too big
3) you need, in some way, to use some type of reasoning to guide various
phases of search [e.g. you search for good sets of genes, in a non-greedy
way, but where "goodness" is defined by the results of another search, for
prob. logical patterns binding together the set of genes]
Now you can do this sort of thing by tying together variants traditional AI
methods, but an useful *way* of tying them together is apparently hard to
come up with, unless you have an overall framework like novamente to guide
the tying-together process...
> Does Novamente learn which variables are worth paying attention to, then
> perform correlation on those variables instead of all variables in the
> problem space? Does Novamente try to notice certain classes of
> that data-mining algorithms don't notice because they're not directly
> useful, then use these useless correlations to spread attention
> and thereby
> find a dataset among which there are likely to be useful
yes to both
> Novamente learned heuristics that enable it to connect locally visible
> properties of a variable to its probability of distant entanglement?
No, it hasn't learned any of its own heuristics at this point
> does Novamente notice correlations in such a way that the
> intermediate steps
> not only have no humanly comprehensible semantics on a case-by-case basis,
> but also in such a way that there is no known explanation for what the
> intermediate steps are, in general, doing?
Yes, this is of course true, but this is also true for NN's and GA's ...
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:38 MDT