Does /g/ into Artificial Neural Networks?
Are ANN more powerful than perceptron with kernel trick ? Can someone give an example were you'll prefer ANN ?
>>58960588
>>58960940
ruski go away
>>58960667
a perceptron is an ANN
>>58960667
ANNs are more powerful where the target task is more complex. This complexity makes the objective function a lot less convex. The perceptron with kernel trick (or an SVM) incorrectly assumes the function to be estimated to be smooth.
Basically, if the target task is difficult enough, layers of nonlinear models outperform pretty much anything else.
>>58960588
/g/ is only into pajeet and Linux memes.
>>58960588
What software do you guys ANN with?
>>58961622
tensorflow is the only logical choice as of now
evolutinaory reinforcement learning is neat, neural nets are a meme
>>58960588
Yes I predict the weather using a mesh of sensors for agriculture purposes. Shit is cash and with that I mean literal money
>>58961622
Tensorflow
Few years ago matConvNet but it kind of died since tensorflow became so great
>>58961994
Are you using a cnn for that?
>>58962024
There used to be a lot of image data but we don't really use that anymore, so I think the cNN is decomissioned
Can any of you smart dudes point me to any /g/ approved learning materials for this subject? I'm kinda a dumbass but I find this stuff fascinating.
>>58963496
your universities intro to ai class ?
When would you use an SVM over an ANN or vice versa?
>>58960588
yeah i just finished writing one that writes raps.
ANN's are so interesting it's crazy - too bad we can't really see what's going on inside of them.
http://boingboing.net/2017/02/14/rhymes-from-a-machine-learning.html
>high schooler
i'm t-totally 18 g-guys...
>>58963496
>Artificial Intelligence: A Modern Approach by Stuart Russel & Peter Norvig
this is literally one of the most used book for AI courses in university
>>58960667
Both are turing-equivalent. Only one doesn't take infinite time to work.
>>58961874
t. never did anything remotely related to ANNs
>>58964732
Also completely irrelevant to the topic at hand, just like K&R is not going to help someone learn python.
>>58960940
dat London look
>>58965042
well it's a book about AI. it's co authored by the director of research of google and has a introduction to AI and contains the most methods of machine learning. and neural networks are of course part of it. it's theoretic, but provides to understand the topic. don't know what's the problem right there
>>58965183
Literally (You).
>>58963576
if you have hand engineered features and a small dataset, use svm. if you want a black box that takes raw data and gives results, use ann.
>>58963496
don't bother with this >>58964732
I think cs231n is applied enough for a beginner. It's completely focused on vision though.
>>58965032
hey dudebro you're the one calling them ANNs
>>58961874
Theano and Torch are just as cutting-edge, and way more research is done on those two compared to tensorflow.
>>58967830
Tensor flow is being adopted more, however, and has better benchmarks, github activity, and support.
>>58967852
And the other two are more popular. It is logical to go with the popular choice.
Tensorflow 1.0 came out today. Soon there will be skynet.
>>58963496
http://neuralnetworksanddeeplearning.com/
This is what my ML class follows. Is good enough for most tasks.