| |
| |
| |
|
Page: 1 2 3 4 5 6 7 8
Comments:
<0> Anyone still have that "quick matrix" link? I want to re-read it =]
<0> nevermind, found it
<1> i wanted to write a book, but I forget what was about... :|
<2> The backprop might find the needed weight might be .34288888888888888888888888888888888888888888888888888888888...
<2> but it has it only gets .342888888888888889
<3> sure
<3> quite
<2> so the computer rounding and not giveing #inf precision is why I tell it to find a output of 1 and it only finds .99999994523
<2> ?
<3> no that's another matter
<3> for sigmoid to produce 1 the input would have to be infinity
<3> same with 0
<2> yeah.
<3> but when I train on images I use the full range from 0-1 and it doesn't seem to be a problem. you could limit it to 0.1-0.9 and then scale it up after
<4> Hi DanF_DrC.
<2> or you can get precision, upto like the 20 place then round it and call it aday :P
<3> hi Andares
<3> Shadow_mil, hehe you will get far less precision than that and you will love it :)
<2> well find it upto the 20 then round to the nearest 15th, for a image, that sould give you more precision then you need.
<3> think of a neural nets as a sheet of cloth draped over a cityscape. it is not a perfect copy
<3> net*
<2> nor should it be
<3> the more refined the net the greater the precision
<3> perfect would be nice but so far not happening. and it's quite useful anyway
<4> DanF_DrC, are there any neural nets that attempt to model human brain behavior?
<3> not unlike our mental butterfingers
<4> *attempt*
<2> How do the learning rate effect it? I know it makes the weight change less perfect, if you have a learning rate of 1 would you only have to train it once?
<3> the term behavior I don't like. but yes ANN tries and has shown some interestingly similar 'behavior'
<2> Andares: model yes, but far far far from the human brain complexy
<3> Shadow_mil, the rate is not a matter of precision. the basic concept is stepwise gradient descent. if the step is too big you miss the hole
<3> trying to reduce the error you can overshoot. which is what a too high learning rate does
<2> yeah. I get that
<3> max rate depends on the target function
<3> some are sensitive. some work at 0.5
<2> the learning rate has to be between 0 and 1 right?
<4> Shadow_mil, so..
<3> Shadow_mil, not sure really but probably
<4> Shadow_mil, I thought the complexity of the human brain lay in the some two billion neurons it sports.
<3> 1 is not a sharp upper limit I think
<4> not in the complexity of any neuron itself.
<3> 0 is though
<2> Hmmm
<2> I would say the best way if the first training make it 1, then the rest make it like .0001
<5> lo
<2> I think if we could have inf percision in math one would hit the mark the first time
<3> that may be too much but it's common to start high and let it fall during training
<3> I often start around 0.5 and fall to 0.3
<2> if my theory is correct
<3> but afaik there is no exact science on that yet
<5> are we discussing some sort of reinforcement learning?
<3> oh no, precision wouldn't help you
<3> _death, ust backprop
<3> just
<4> Shadow_mil, anyway.
<5> i don't completely understand backprop yet
<4> _death, join the club..
<4> We have NAMETAGS!
<4> w00t.
<2> Well the way I look at learning rate, all it is, is a scale of the real value, you make it like .5 it meets 1/2 of the distance between the starting value and the real one, then the next time it meets it 1/2 again, and slower gets closer
<2> how I am seeing it
<3> _death, join the club. but it's largely the same as setting b in y=bx for pairs of x and y
<3> stepwise
<3> adjusting weights in the direction of reducing the error
<3> backprop is merely an efficient way of that for multilayer nets
<3> why it works for non linear functions is as of yet beyond me
<3> probably something good to be learned trying to understand that
<0> Which part? The math?
<5> there are a lot of ways of setting b
<3> _death, this is stepwise though
<5> and you can extrapolate with b
<5> oh, right
<6> <danielbar@ef> wtf
<6> <danielbar@ef> stop that ****
<6> <danielbar@ef> !
<3> if you should discover a better way for non linear functions you would have improved upon backprop
<6> <danielbar@ef> enough with this !
<6> <danielbar@ef> ENOUGH!
<4> 'Ey daniel! Stop with the profanity!
<6> <danielbar@ef> STOP This crap is f*cken annoying
<2> ...
<4> What's annoying?
<3> probably just a troll. ignore him
<4> DanF_DrC, can't we block people from the Relay?
<4> If not, let me at the source. I'll add it in.
<6> <danielbar@ef> oh
<6> <danielbar@ef> cool
<6> <danielbar@ef> its a relay
<3> trolls are discouraged by not talking to them
<3> best medicine
<6> <danielbar@ef> first time I see this sorta thing
<5> extrapolating with neural networks
<4> danielbar, no, it's a typewriter.
<4> n00b
<6> <danielbar@ef> I thought this AIrelay Was simply a m***ive flooder
<2> IRC is just a muiltplayer notepad :)
<3> shhh
<2> No, we are real people, with real feels, that you are hurting
<6> <danielbar@ef> the bad thing about this is that I cant use the Tab key to shorttype ur names
<5> does anyone know anything about extrapolating with neural networks?
<6> <x00q@ef> Don't block people from the relay!
<7> hey ya'all
<3> hi. any progress?
<0> _death: That's tricky business since you're working with bounded outputs. You have to know the range your extrapolating to ahead of time.
<7> ehhh, a few more lines of code. not much
<4> x00q, but he's being trollish!
<6> <danielbar@ef> are there any chess players here?
<7> real life tends to get in the way
<6> <x00q@ef> But I'm using the relay too!
<7> chess? is that a game? :)
<3> _death, extrapolation is not it's best feature
<4> danielbar, *BEEP* WOULD YOU LIKE TO PLAY A GAME?
<4> HOW ABOUT A NICE GAME OF CHESS?
<3> interpolation
<6> <danielbar@ef> Yep I very much would
<4> x00q, I was talking about blocking just daniel lol
<3> its*
<5> x00q: think to yourself, "is this the best way of doing things?" next time you have a project
<3> Andares, wopr :)
<0> _death: hmm?
<2> DanF_DrC: I guess that step is better two if need two things trained on the network so that way if the true needed value of that weight can be found
<6> <danielbar@ef> where do u want to play?
<6> <danielbar@ef> we can play on FICS
<2> if it was perfect, it would jump back between the two every time
<0> Andares: Ah :P
<4> danielbar... I actually miiiiiiiiiiiiiiiiiiiiiiiiiiiiiight be interested.
<4> danielbar but yeah, it was a joke... from the movie Wargames?
<5> a neural network is like a screwdriver; if you need a set of calipers then it's the wrong tool
<3> Shadow_mil, step is because of the non linearity. if it only had to adapt to something linear it could do it directly in a single step
<7> that's very.....quotable, _death
<3> non linear is more problematic
<5> lol
<0> _death: Well I think that answers your question
<4> _death, can't you hook up a neural network to the net and train it to hack a tool supplier and deliver you calipers?
<5> well...
<6> <danielbar@ef> Andares oh yeah I remember that book
<2> I understand the differance between step and linearity
<4> danielbar.. it's a book? I thought it was just a movie.
<6> <danielbar@ef> so lets go for It
<6> <danielbar@ef> I'm logged to FICS right now
<5> so don't use neural networks for extrapolation?
<6> <danielbar@ef> are u already there?
<4> I dunno what FICS is.
<7> www.freechess.org
<6> <danielbar@ef> the Free Internet Chess Server
<5> hmm
<3> extrapolation can not be done in general. so it's not really a failing of ANNs
<8> mmhmm.
<7> what's your account, danielbar?
<5> that's an interesting way of looking at things, DanF_DrC
Return to
#ai
or
Go to some related
logs:
movw syntax ARM
No catalog found at ERROR: No proposal
#web
rpm initdb can't create transaction lock
packman unixhead
pysh job control
gentoo FAILLOG_ENAB
#oe
ssh2_exec troubleshooting
ubuntu wireless icon tray
|
|