| |
| |
| |
|
Page: 1 2 3 4
Comments:
<0> hi
<0> could someone recommend a good book on image recognition?
<1> DaYZman which ones have you got?
<1> One) Gonzale's and Woods: Digital image processing
<1> 2) shapiro & stockman: computer vision
<1> There're many others
<1> *1) Gonzalez and Woods: Digital image processing
<0> i haven't got any
<0> i'm absolutely fresh to image recognition
<0> or image processing
<1> Depends on your goal
<1> Which specific aspect are you interested in
<0> well, image segmentation and recognition
<0> still images
<1> You mean object recognition
<1> That's considered an advanced topic
<0> i have a fair background in AI but i haven't laid my hands on image recognition
<1> "Visual perception" by bruce, green & georgeson is a very good background book
<0> how about the first one you suggested, Gonzales and Woods?
<1> That's more down to earth
<0> oh okay
<1> Object recognition in that book is just one intorductory chapter
<0> Visual Perception: Physiology, Psychology and Ecology?
<1> Yeah
<1> This one is more like general talk
<1> Very high-level
<0> yeah, the title looks like it
<0> i'll find it in the library
<0> thanks
<1> You can't expect a book to tell you how to do object recognition right away... that's an unsolved problem!
<2> GP is considered to be AI? :(
<3> indeed
<2> but GP doesn't try to model human thought or the human mind.
<2> basically all the inspiration comes from genetics, as opposed to neuroscience. there's no cognitive aspect to it. it's an optimization procedure.
<3> I disagree. Haven't you ever found yourself thinking "well, what if I try this, and than that, and then...". That's what GP is. A simplistic form of trial and error.
<3> Not that AI is merely concerning with modeling the human mind. We're not even sure how the mind works.
<2> i don't like the term AI because as you've hinted at, it's ill-defined and has a stigma of failure surrounding it.
<2> finding better GP operators is a concrete goal. "figuring out how the mind works" is ill-posed
<3> True. I prefer to define it simply as making computers solve problems without explicitly telling them how.
<2> so all declarative languages constitute AI? like SQL? :)
<3> I see where you're going, but no :)
<2> as for the topic in #machinelearning
<2> i've thought it would be interesting to look at EDGAR data myself
<2> another data set i've wanted to analyze is federal contracts
<3> It's true, even most AI code is written in procedural languages. But AI code, like GP, isn't necessarily gauranteed to find the right answer, or accomplish a single pre-defined goal.
<2> e.g., studying patterns of no-bid contracts
<4> i like that definition too, Cerin
<3> chessguy: Hey. Did you ever play around with the code I sent you?
<4> haha
<4> that's funny
<3> ?
<4> i've barely had a spare minute to catch my breath this week :)
<3> psshh, you college people and your *work* :)
<3> wkh: What does the data represent?
<4> i did learn something new about python last night
<3> chessguy: oh?
<4> i never knew it was whitespace-dependent
<2> cerin: besides just the absolute dollar amounts, there's tons of other fields on the bidding process that use acronyms which aren't defined anywhere in the data dictionary they provide and that googling couldn't provide satisfactory definitions of
<2> they had a 100-page data dictionary that 'explained' what all the different fields meant
<3> chessguy: yes, helps keep your code cleanly formatted, and saves you the time typing useless {,},begin,end, etc
<2> it was clearly only for people who were directly involved in the bidding process
<4> yes, it's a very interesting idea
<3> chessguy: It was one of the hardest things for me to accept about the language, coming from a C++ background, but now I prefer significant whitespace.
<4> i remember an old column-based language called RPG
<4> i guess we've come full circle now :)
<3> wkh: What are you trying to do with all that data?
<4> anyway, tonight's the first chance i've had to relax for a while, so i taught myself some haskell. very fun
<5> Heh, I'm fooling around with facial recognition now :)
<2> cerin: i wanted to see who was getting no-bid contracts in Iraq.
<2> i found interesting stuff, like $150 million here and there going to people with no discernable name and for no discernable purpose.
<4> must be a jewish conspiracy
<3> heh, only$150 mil? Seems lite.
<2> it's not m***ive amounts of money, but $30-150 million here and there (i remember seeing several) adds up
<2> this was about a year ago when i was first learnign about data mining and neural networks
<4> oh well, bedtime
<4> ta
<6> hi guys...how do we debug neural networks? i am really not able to find anything wrong with my code. any help is appreciated.
<2> are you implementing a common architecture yourself? if so, maybe switch to using a pre-existing library
<2> that's kind of the copout approach and probably the sanest thing to do
<6> wkh, i dont want to use a library. i know many exist. but i want to learn programming a neural network. the thing is that i just want pointers as to how to debug such networks. i have implemented the bpn algorithm for solving the XOR problem. i want to construct a dummy network on paper so i can check the outputs at each neuron. but i dont know whether this is the right approach or is there a better approach to do it.
<2> um. well. i found that getting backpropagation right for a cl***ic 80s-style three-layer perceptron, it was kind of a pain in the ***
<6> hmm...well. the network is pretty flexible so i can change it. but i guess this would be the simplest network to implement.
<2> um
<2> haykin's book on neural nets gives the solution for a set fo weights that solves the XOR problem with 2 neurons
<2> when the activation function is tanh, if i remember correctly
<6> i have haykin's book. yes it explains the working for tanh as well as for sigmoid. but i dont think it gives the solution for it..
<2> i seem to remember it being an example
<2> in chapter 4?
<2> haven't read the book in several months
<2> the 2nd edition
<2> um
<2> when i implemented backpropagation in C, i made typos
<2> in other words the bugs were transfer errors
<6> well. i have it in front of me right now. yes. its in chapter 4. but not much of an explanation. only one graph.
<2> it took me about 3 hours of staring at the code to debug
<6> yes i have the 2nd edition. and yes i have implemented it in C.
<6> its good for two inputs but not for the other two inputs! :/
<2> isolate what the potential problems are by implementing it in pieces
<2> make sure your IO stuff isn't incorrect
<2> are you using proper input values?
<2> what's your learning rate
<2> if you're trying to use momentum, get rid of that
<6> i'm not using momentum atm
<6> learning rate is 0.5
<2> lower it to 0.05
<6> ok
<2> (machine learning is just a hobby and i'm not an expert)
<6> lol. me neither.
<2> it's funny, i've always used really small learning rates in my own code while everyoen else seems to use enormous ones
<2> like i used a learning rate of 0.0005 to start with and then worked downward
<6> after 1000 epochs, all inputs are producing almost the same output
<6> ok...
<6> no change in the behavior :/
<2> make sure you have the algorithm right
<2> that was my problem, i had the actual algorithm wrong
<2> the traditional idea of tracing through input and comparing expected to actual values doesn't seem to apply with neural nets because it's all floating point **** and you have no idea how much it should change at each point
<6> i understand.
<6> i'll trace out the forward and backward p***.
<6> i am sure my forward p*** works correctly as i had tested that with dummy inputs.
<6> i guess i will have to isolate the backward p***
<2> yes, that's what took me 3 hours to get right
<6> hmm...must be some programatic error.
<6> i've written down the algorithm, equation by equation and studied it to make sure i've gotten all the parameters right.
<2> for instance, i had the first layer of weights mixed up with the second layer of weights at one point
<2> i also had the indices wrong
<6> ah
<6> btw how do u handle the bias?
<2> are you not including a bias?
<6> i am including a bias
<2> i added an extra node in the hidden layer
<6> it gets updated with the rest of the weights
<6> yes but that node isnt supposed to be connected with any of the input layer nodes, right?
<2> i forget.
<6> its not supposed to be connected. i confirmed
<7> [ - hello - ] [ - we are invading your planet - ] [ - or something - ]
<2> do we get to wear shiny leather clothing and tell each other "GO!!" when we want them to go and do something immediately
<4> woohoo!
<4> everybody ready for the discussion?
<6> where is everybody? :P
<4> i'm here
<6> i c
<4> are you going to be here for the discussion, basix?
<6> i am a n00b so i'm gonna watch :P
<4> have you read the papers?
<6> hey chessguy is there a channel specifically for neural networks?
<4> not that i know of
<6> not really...
<8> there's #neuralnetworks but it's inactive
<4> oh, they're worth reading
<4> hey cwenner!
Return to
#ai
or
Go to some related
logs:
qtorrent: cannot connect to X server
detected depreciated pcmcia ioctl
python bncs
#oe
#mysql
hardware+homologation+fedora
xine-lib xine-config package apt-get
fail to initialize ASIC in the kernel
spca rpm ubuntu
rug-zmd suse
|
|