| |
| |
| |
|
Comments:
<0> fuso, I was serious. and I did get banned for no good reason so that's not true
<1> can anyone of you trace back "dnsa's exit from channel" and tell me what it says besides?
<1> I have some problems with connection
<1> ok, it's fine
<0> * dnsa has quit IRC (lem.freenode.net irc.freenode.net)
<0> there are no exits in my log
<2> DanF_DrC: thank you. I figured out some other people had problems also because 5 people entered channel at nearly the same time. And now I also have a message from freenode staff in #1.
<3> hi there...
<4> I discovered AI today, I've got it and you don't. ( hums to the tune of nanny-nanny boo-boo)
<4> The singularity is near and I will dominate all!
<5> I need help with the xor neural net!! I've been taught that a network that decodes xor should contain 2 input neurons two hidden, and an output. However now i have a net composed of only ONE hidden neuron.
<5> is that possible? isn't that impossible according to the linear separability of the sigmoid neuron?
<5> please help ... :(
<3> no patience huascar
<3> for future reference, one neuron is not possible, as there are two "borders" in the XOR problem, so we need to decision planes, hence two neurons
<6> hi people, i've implemented the error backpropogating neural network algorithm in C. the problem is that for some reason the weights of the hidden layer aren't getting updated. i've tried searching for tutorials which are more implementation, rather than theory oriented, but i havent been able to find such tutorials. anybody know some good resources ( books, links etc... ) which can help me out?
<3> basix: are you actually applying the training procedure to the hidden layer?
<3> is the error gradient zero or something?
<6> yes i am applying it to the hidden layer
<6> any pointers links, books??
<6> i need to cross check my formulae
<7> knowing the theory, you should be able to track the bug down merely by verifying values at various points
<7> if they do not change at all, obviously either the code isn't running or the addition is zero. check which it is, et.c.
<6> alrighty
<7> basix: i find this book explaining backprop very well: http://ai.stanford.edu/~nilsson/mlbook.html
<6> thx a lot :)
<8> <Koryu@ef> hello
<8> <Koryu@ef> i have a bot that is an operator
<8> <Koryu@ef> how do i use it?
<8> <Koryu@ef> can anyone help me
<9> it's hard to help without a problem
<3> basix, any luck?
<6> i'm reading that book :)
<6> AJC, my basic concept is clear, yet, i've prolly made some logical error while coding
<6> i hated calculus
<6> haha
<3> you can **** at anything and still manage to get it working with TDD
<6> lol thats true...
<10> there's quite a difference between not updating the weights at all and not learning. for the former, it's probably better to find the interval where the program and you disagree. knowing you agree at A but not at B tends to limit what sort and what parts of the code the error lies in
<6> hmmm
<6> thanks...i guess i should've implemented the algo in stages...i did it all at once...i think i'll go in for a complete rewrite. this book is giving me a better understanding of the algo
<3> basix, the algorithm definitely has two p***es. one to compute the errors, the other to adjust the weights.
<3> you have to do two p***es, as one goes backwards
<3> the other you can do forwards
<6> AJC, yep yep i understand all that
<3> actually, almost three p***es
<3> simulate, propagate back, then adjust
<6> first i setup the weights to random values between -0.5 and +0.5
<6> next, the forward p***. the input signal propogates till the outputs
<6> then i calculate the error in the output layer
<6> propogate it back
<6> and then adjust weights
<3> sounds sensible
<6> its just those damn formulae :P
<10> verifying the functionality a single neurode would probably be wise
<11> zeeeee: MIT:D
<12> icez: hi
<6> cwenner, hmm...its difficult to do that with my code :/ i'll try to do it though.
<6> guys thanks for all your help
<6> but, i'll be back :)
<6> byee
<10> it seems to me, and some searching seems to support that a Partial Order Planning solution using simple search potentially runs into infinite recursion (a->b, b->a, goal: b as a simplified case). AIMA doesn't seem to address this do, saying that in Difference to the propositional POP case, HTN planning may recur infinitely. am i wrong to consider their described POP solver (excl. the planning graph heuristics) to have
Return to
#ai
or
Go to some related
logs:
#css
ERR Hotmail said you must pay money to have WebDAV access
install supermegaspoof ubuntu
zerodni
#perl
#perl
how to mount usb webcam + ubuntu
libcairo FT_GlyphSlot_Embolden
initramfs compile kernel xconfig
#ldap
|
|