| |
| |
| |
|
Page: 1 2 3 4 5 6 7
Comments:
<0> auto***oc is where you fold a normal backprop NN to feed input and both input and output and use the middle layers as output
<0> s/and/as
<0> as both...
<1> So, ... does that mean you can 'recall' stuff or what?
<0> supervised learning with examples like x,x
<0> no, it finds features
<0> unsupervised learning
<1> hmh right
<0> I used it for image compression. and it worked
<1> I think that Hopfield net is a content addressable memory, .. so it's not alike indeed.
<2> what did you use for image compression?
<0> the power that comes from an algo that can find a solution on its own is not to be underestimated
<0> auto***oc backprop
<2> DanF_DrC, I think anli was doing something like that.
<2> How much was the compression?
<0> depending on quality not unlike jpeg
<2> Cool.
<0> very
<3> DanF_DrC: Hopfield is auto***ociative.
<2> find: ./Documents and Settings/Atari/Local Settings/Temporary Internet Files/Content.IE5/096nghmj/?>:: Input/output error
<2> Ach, not good is it?
<3> Unless what I had read and presently see on Google is wrong
<0> x00q, not in the same sense
<3> Differing meanings of auto***ociative?
<1> Andares: Dunno, that filename looks weird. Try deleting it .p?
<2> Defty, there are hundreds ;o.
<1> That's just a cache directory. Feel free to delete everything in there ,p
<0> x00q, spend energy on understanding instead of p+roving wrong
<2> Defty, but the "Input/output error" part scares me.
<1> Andares: Try to read them with something else than find?
<2> Meh, I don't care that much.
<2> I'm just trying to find a needle in a haystack.
<1> x00q: Maybe what you refer to is plain '***ociative'?
<1> Hopfield is described as "content-addressable ***ociative memory".
<0> some do call it auto***ociative because it recalls
<0> but it's not the same
<3> DanF_DrC: lol, spend time on Googling rather than arguing
<3> Defty: Maybe
<2> DanF_DrC, auto***ociative is setting the output to the input and picking up the real output in the middle of the net, right?
<1> x00q: in wikipedia, :-)
<3> Ah
<1> Andares: You should consider using 'updatedb' and then 'locate', instead of 'fidn'.
<1> *find
<1> Andares: updatedb as root.
<0> Andares, yep
<3> x00q: Strangely, when I looked up auto***ociative both in my brain and on Google, I found Hopfield
<3> Thus, so I spent time time trying to understand by asking you :P
<0> for instance 16 input and 16 outputs but 4 neurons in the middle hidden layer. perhaps 16 40 40 4 40 40 16
<2> Defty, updatedb explodes the computer.
<0> I called hip learning when I rediscovered it
<0> it
<1> Andares: Huh? It just takes a while to run :-D
<2> As in, I come back in 3 days to find the thing locked up with a 85 degree (celsius) CPU temp
<2> Completely unresponsive.
<1> Ok, you have some file system problem then.
<2> Proooobably yeah.
<1> Andares: You can tell it not to search given partition (such as windows etc.)
<1> Andares: Dunno where, I guess the manual tells you.
<2> Defty, see, it's on the windows partition..
<2> the file is.
<1> DanF_DrC: So.. um.. both the input and output layers are used as input or what=
<1> And how do you then backpropagate?
<1> From the middle?
<1> heh sorry
<1> you don't, :-D
<1> uhmn.. or do you..
<0> Defty, no reverse. input is inout and input is target output as well. the unit function
<3> DanF_DrC: Got a good link on auto***ociative backprop?
<1> DanF_DrC: Oh.
<2> It's the same as regular backprop, isn't it?
<1> Andares: seems so.
<0> x00q: not sure it's good and there is nothing more to it than I have already told other than appreciation of it but : http://www.ncl.ac.uk/cpact/neural_networks_nf.html
<4> do those things propogate?
<0> scroll down to auto***oc
<0> Buggaboo, converge? sure
<4> do hopnets allow cyclic structures?
<1> Buggaboo: they're recurrent by definition I think
<4> I need to read up again, this stuff is 3 years ago for me (and I didn't pay attention then :)
<1> Buggaboo: It seems to be kind of fully connected, also in the recurrent direction...
<1> Just looking at a picture.
<1> Buggaboo: I mean, there's no layers.
<1> Buggaboo: There's just nodes, and each node has input from all the others.
<1> (In addition to an input and an output to feed & recall)
<1> Buggaboo: And outputs are either -1,0,1 depending on a threshold and the type (either 1 or -1)
<4> ah.
<4> so how do the weights adapt?
<1> curiously no idea. there's an energy function and it converges to states of local minima of that energy
<1> the states are binary.
<1> i.e. each node is binary, so ..
<1> hmh btw they're either from 1 to -1 or from 1 to 0
<1> i mixed up previously
<1> (The output values, ..)
<1> And I'm not *sure* if the states are binary ;-)
<1> Yeah, states are binary and you lower the energy of states that should be remembered :p
<1> so I guess you compute the gradient and use whatever descent method
<4> arctan, sigmoid, hard threshold?
<1> harsh
<1> hard
<1> I mean ;-)
<1> It's either 0 or 1 (or -1 or 1)
<1> you sum state_j for all weights
<1> and see if it crosses the threshold, that's the input
<1> so you feed some binary number and get the one that is "remembered" ...
<4> okay, so what patterns does it learn?
<1> or maybe you don't even have to feed binary, dunno
<1> Different configurations of states :p
<1> if you have 10 nodes, you have 2^10 possible configurations etc.
<1> oh, and w_ij = w_ji so they're symmetric
<4> okay
<5> DanF_DrC hi I am back
<1> Anyway, when you select a node at random, the net moves to some minimal state, for the state and its neighbors..
<1> Buggaboo: I should also learn about this in more depth, to get a more rigorous picture ,p
<5> sorry that took so long, had to move about .7 ton of gr*** over 30 yards uphill
<4> explain it to me to make sure you really got it okay? :P
<1> maybe later ;)
<1> I'm watching some stupid series, going to have to sleep and I have work to do during days... finally got a summer job. Though I'm not sure if it's only for the summer :-p
<4> good luck with that
<1> sure
<6> "cognition is a chimera contrived by an observer who is necessarily biased by his/her own perspective on the environment. (Brooks 1991) As an entirely subjective fabrication of the observer, cognition cannot be measured or modeled scientifically."
<2> Yet it exists. :/
<6> that's pretty extreme
<6> this is the same guy who said that we "athropomorphize fellow humans"
<2> Haha.
<7> Aaah, doctrinaire behaviorism.
<6> or philosophical or possibly scientific reasons, some people believe that human intelligence is fundamentally rational, and can best be represented by logical systems incorporating truth maintenance. Others believe that human intelligence consists of a m*** of learned or evolved hacks, not necessarily having internal consistency or any unifying organizational framework.
<6> people sit around talking about 'semantic nets' and 'knowledge representations' while they dont even have a basic intuitive notion of how an insect's brain works.
<6> its idiotic
<6> Im clearly a Scruffy
<6> http://en.wikipedia.org/wiki/Scruffies
<2> :O
<7> evolution is nothing BUT hacks.
<2> hax omg.
<2> Sotek, hey, don't dis evolution. It built you. :D
<2> Though I prefer NN's
<7> yes, and? .p
<7> it's such an amazing piece of hackery...
<2> Haha
<2> I just remembered.
<2> I met a guy with THREE kidneys today.
<2> And all three functioned.
<2> He's a lucky guy.
<2> He can give up two kidneys and live.
<6> his urine must be clear as ice
<2> XD/
<2> *.
<2> Thanks for the mental images, paros.
<6> youre welcome
<0> evolution didn't build ****
<0> but understandable devotion
<2> Hi "De4thClaW"
Return to
#ai
or
Go to some related
logs:
You must have the XTest library to build synergy
Python book torrent
dandesousa
#perl
kernel panic - not syncing : VFS : Unable to mount root fs on unknow-block(0,0)
amarok+suse+libslang
#kde
acroread syntax error ubuntu
#perl
honeywell dps6 value
|
|