| |
| |
| |
|
Page: 1 2 3 4 5 6
Comments:
<0> If you have a robot with a kohonen net inside of it
<0> what granularity is best?
<0> If its too fine, the robot will make distinctions between irrelevant stimuli.
<0> If its too coarse of a network, the robot will never pick up required distinctions.
<0> SO its a goldilocks problem.
<0> There is some middle-ground granularity that is best for the robot.
<1> I totally disagree with that part. Think about it.. If the stimuli is generated by a computer, the faster computing speeds the more it can take in over time.
<0> So MORE is not always better.
<1> You could stimulate it with a hundred DVD's a second
<1> it would grow that much faster methinks
<0> Middle-ground issues are solvable via genetic algorithms.
<1> but if you only used stimulation from the real world.. I would agree with you.
<0> Those are built to find "good/best middle grounds" of certain things.
<0> My idea is to have the robots contain all different kinds of kohonens maps and then select for the most successful.
<0> This is something pfeifer didnt even do
<1> that's a good idea
<0> Well I think 3d games could look pretty realistic and even have physics. So maybe a sped-up simulated world is the way to go
<1> have you tried programming that approach? with a multitude of kohonen maps?
<0> ive really been avoiding that kind of thing
<0> maybe its time i get into it
<1> haha.. how come, that's the fun part
<0> yeh i think its time
<0> I have ideas on how to extend this architecture as well
<1> well, if you need a programming slave, let me know. i've got endless hours just waiting to be spent not on IRC! =p
<2> ive got a business idea Yobert maybe I could tell you about
<3> <ce_muslim@ef> hallo aidha
<3> <ce_muslim@ef> hallo ai
<3> <Elixir_@ef> I'm trying to research models of AI. If one developed a system that emulated the human brain, what would call the system?
<3> <Elixir_@ef> Where you had a subsytem for the amygdala, cortex, etc
<4> that would be more in the direction of psychology and cognitive science. I'm not familiar with those terms.
<3> <Elixir_@ef> cwenner: thanks
<3> <Elixir_@ef> I'll look in to that
<5> hi all
<5> i hav problam wit bak propigaton an neurol natwirk
<5> can uz H3LP?
<6> AJC, shoot
<5> nah, i was kidding... seems we only get people in here asking that question
<5> i thought the purposefully bad spelling gave that away :-)
<7> IRC needs a /last command
<5> <5> i hav problam wit bak propigaton an neurol natwirk <5> can uz H3LP?
<7> heh :)
<8> chessguy: irssi has a /last command :/
<5> icez, even when you were not connected when the /last thing was said?
<8> uhm, that probably not:)
<6> AJC, in all fairness to them, NN is the only base thing of worth in AI
<5> what is, and what isn't a "NN" is rather fluffy these days
<5> a bayesian system is a "network"
<6> that's just the probabilistic camp trying to steal thunder to cover for the uselessness
<5> a NN is just a bayesian process with ***umptions built in
<5> actually, i think "gaussian processes" is the field that encapsulates all that
<5> e.g. http://citeseer.ist.psu.edu/williams97prediction.html
<5> this one I've started reading before, good but it gets challenging towards the end :-)
<5> http://www.inference.phy.cam.ac.uk/mackay/gpB.pdf
<6> well if the ***umption is that bayesian is useless and we use NN instead then yes NN is bayesian with ***umptions
<6> otherwise they are completely different and bayes brings bubkes to the table
<6> afaik
<5> GP are a superset of both
<6> hehe ok now you are just trolling
<5> read the ****ing PDF
<9> AJC: don't bother offering material, published or not, that contradicts DanF_DrC.
<6> I think not. I've heard similar conclusions before. would it be completely unreasonable to say that you don't understand the reasoning but merely accept the conclusion of the papers?
<9> He's opinion won't be swayed, no matter how absurd
<9> s/He's/His
<5> well, since i actually read part of those papers DanF_DrC, i guess i understand it more than you.
<6> if you do then could you perhaps explain just in loose terms how gradient descent emerges as a special case of BN?
<6> sigmoid activation function, weighting, summing
<5> not BN, GP.
<6> <5> a NN is just a bayesian process with ***umptions built in
<6> but try with GP if you will
<5> <5> actually, i think "gaussian processes" is the field that encapsulates all that
<5> that's what the PDF covers
<6> hehe which you say you understand parts of
<5> you have the material you need, now all you need is the will to learn.
<6> I believe there is nothing to learn for me in this. and wasting my time on the pdf when I understand how gross the differences are
<5> you're confusing BN and GP
<5> you may understand BN intuitively, but you've never heard of GP until today
<5> otherwise, you wouldn't be making such stupid claims...
<6> ah no I was confusing GP with genetic programming. but you did say bayesian process once
<5> i did, say bayesian, yes... part of the theory behind gaussian processes is involved with bayesian inference.
<6> I glanced it. I'm still not feeling it.
<6> I still get the sense that they offer nothing better but merely tries to get credit for NNs
<5> i think it has merit in a sense that the hard-coded behaviour of your neurons is often problematic.
<5> e.g. the two spirals cl***ification problem
<5> and it's not like NN have been that "succesful" anyway :-)
<6> I can't rule out improvements on backprop but actually it has yes. unlike any other approach I might add
<5> genetic programming has already done more than NN ever could
<6> hehe yeah right!
<5> NN just does cl***ification, prediction... not much else really.
<6> let's call it learning
<5> don't get me wrong, i'm not a fan of Koza or genetic programming, but he has a point.
<5> http://www.genetic-programming.com/humancompetitive.html
<6> it's infinitely better than GP. it's beyond comparison. NN learned to drive a car across USA with a very simple net. it's used for OCR, sorting wood in factories by vision. and GP has done what?
<6> trite circuit boards. single motor jerky bots. bubkes
<6> nothing simple brute force couldn't do better
<5> fair enough.
<5> here's something for you to read.
<5> http://www.paul-almond.com/OccamsRazorPart09.htm
<5> i've not wrapped my head around it yet, didn't have time so far.
<6> so why should I read it?
<7> the links to the previous articles on that page don't work
<6> hi chessguy. is your project moving along?
<5> chessguy, from the main page they seem to be ok.
<7> no, i've been out of town and away from the computer for a few days
<5> DanF_DrC, he's a smart guy, asked me to look at his work. and it seems you don't have much else to do :-)
<6> AJC, NN owns so far with no challengers
<6> should anyone claim to beat NN in general, I'll take a look
<7> ah, found it. thanks
<9> DanF_DrC: Did you see: http://www.genetic-programming.com/humancompetitive.html ? Also, do you have a reference to back up the " NN learned to drive a car across USA with a very simple net."? I am not familar with that.
<6> x00q, a good old example in the topic
<6> and yes I've seen it before
<6> it's a long list of trite results that simple brute force could do better
<7> AJC: doesn't genetic programming refute occam's razor? :)
<9> DanF_DrC: You really should publish that conclusion. You'd be famous.
<6> brute force is quite human competitive in narrow searches
<5> chessguy, no, in fact the GP guys go a long way to reduce the complexity of their programs
<6> x00q, hehe no fame depends on general appreciation
<9> DanF_DrC: Yeah, in *narrow searches* sure
<6> you ***ume incorrectly that people respond to reason
<6> x00q, those GP results are very narrow searches
<7> mmm, but the fact is that genetic diversity is inherently necessary in the population. and not just diversity, but redundancy
<9> DanF_DrC: I don't ***ume that. This is evidences by arguing with you.
<6> x00q, how would you know
<7> necessary for improving, that is
<5> DanF_DrC, they are not narrow searches. they are much broader search spaces than NN
<6> take a look at the good old example x00q
<5> it took months on huge clusters to find solutions to those problems
<9> From the tokyo lectures?
<6> AJC, because they didn't use brute force
<9> DanF_DrC: How would I know what? That you don't respond to reason?
<6> x00q, no not from the tokyo leectures. 'good old example' read the topic dude. it's not that many lines
<5> DanF_DrC, don't you think that if it was possible with brute force, everyone would be doing it?!
<6> AJC, no
<5> 'cos you have to be smart to do brute force?
<7> AJC: no, because the government is suppressing all scientific advancement
<5> chessguy, touche!
<9> hahaha
<6> hehe idiots
<5> rofl
<6> not only stupid but flaunting it
<5> DanF_DrC, why not email Koza and tell him that brute force is better?
<7> that one was just too easy :)
<4> if they are so stupid, how about actually giving proper arguments on their terms and beating them in their own game? i cannot say i've witnessed much from you in this area
<9> DanF_DrC doesn't have to, he was gifted with a great brain and he is just exercising it.
<5> cwenner, this channel has 54 people who've been disgruntled yet not proven false by you-know-who
<6> AJC, basically two possibilites, he already knows or he's too stupid to realize it even if I tell him.
<7> hey, anyone actually familiar with the theory of genetic programming? the schema theory, etc.
<6> not unlike a few people here : )
<4> that's hardly an argument against the propositions
Return to
#ai
or
Go to some related
logs:
synergy ubuntu startup
#perl
file associations +gentoo file manager
#php
qkemu layer
[blocks B ] sys-apps/pam-login (is blocking sys-apps/shadow-4.0.15-r2)
#centos
losetup -d ioctl: LOOP_CLR_FD: Device or resource busy
claimed by usblp while 'scanimage'
kde 3.5.2 automount
|
|