| |
| |
| |
|
Comments:
<0> What is the model that always bets on the majority cl*** called?
<0> It's something like Majority Vote Model
<0> although i believe it was a different term than 'vote'
<1> winner takes all?
<0> Majority X Model
<0> I said it in this channel a few months ago but cannot recall it anymore
<1> you can search the logs
<1> too bad brains dont keep logs
<0> i find the brain's way of doing it very exciting and all but inefficient, given a limited storage that is
<1> you can search google for "Majority * Model"
<0> i think i found the term in Mitchell's
<0> could yoiu check the registry?
<0> register*
<1> huh
<0> Mitchell's Machine Learning
<0> i believe the term is in it. could you check the index?
<1> i dont have that book
<0> i'd recommend it
<0> although so does danf :)
<1> yeah i know the basic stuff already
<1> and the real challenge of ai isnt finding ML solution to fixed domain problems
<0> fixed domain?
<0> do you mean stationary environments?
<1> dunno
<0> what do you mean by fixed domain?
<1> i meant like
<1> finding an algorithm for some specific problem
<1> nevermind
<0> sorry, i'm not trying to be hostile, just want to know waht you mean. you're probably right
<1> no i just dont know what i mean myself
<1> most of the time
<0> so you mean that real AI is general AI and narrow AI isn't very important?
<1> id just like to see all the narrow ones in a big one
<0> personally I believe we are getting there, we make discoveries in subproblems and utilize them in larger problems
<2> I'm pretty sure that real AI is not a collection of narrow ones
<2> one general one applicable to all the skills
<1> fys03cwe: yeah
<0> the narrow ones wouldn't nec. have to be indepedent and called separately
<0> they could be integrated to produce what we would normally call one system, and not many
<1> yeah, and that part is interesting
<1> not finding some algorithm that is 10% better at some task that is already solved
<0> 10% is quite a lot in my mind
<0> i notice when I make systems how the accuracy of low-level components is crucial
<0> for instance, if my dependency graph parser makes a single error at any of the involved nodes, a proof may no longer be possible
<0> in practice, perhaps this puts an upper bound of the square of the accuracy of dependency graph parser (just a guess)
<1> yeah but you can handle slow or wrong execution
<1> if you have smart control structures that notice it
<1> and try other things
<1> so it's more effective to spend your time on these things than improve an algorithm 10%
<0> possibly. i enjoy metareasoning, how to select which of several reasoning acts is the most rewarding
<0> (for instance)
<0> that would be more of the more naive syustem we had above, with separate modules
<0> still unable to find much resources on it however
<0> but you may want to look more into Artificial General Intelligence as it's called
<0> the hutter and schmidhuber that we constantly speak of in here for instance
<1> i only know a couple of systems that try to do it
<1> who are they
<0> people that do research on AGI
<0> there are many more though
<0> you must have heard of Novamente for instance
<0> google for AGI
<0> on videos
<1> i think it's more gonna be a question of combining a lot of things together
<1> and engineering a good system
<1> as opposed to "Optimal Ordered Problem Solver" and "an even more general optimal search algorithm"
<1> just reading some abstract of them :)
<1> why would you need it to be optimal
<0> cause you find suboptimality to be too easy?
<3> any comments? cennywenner.com/elfrte.pdf
Return to
#ai
or
Go to some related
logs:
clint
#dns
zope gc.collect
EPIA-5000 Debian Extremely broken BIOS
#suse
t
totem clean open location
#lisp
gnoomeeting
pinyin qwertz
|
|