@TAbraham I'm no expert on neural networks but I'm interested in learning about them to answer the genetic algorithms question so I'm happy to look things up as we go... :)
@trichoplax Yeah, almost finished it.. would be cool if u use that to answer the question.. although it may be redundant to use genetic algorithims to train a neural network to use a genome in a genetic algorithim!
Making a neural network
Challenge
As you can see from the title, you are supposed to make a neural network with the least possible characters. The images for the cat pictures are in this dataset: http://bit.ly/JgaTcL
5000 pictures will be used to train the neural network. Use ten more pictures ...
@trichoplax wait.. r u talking about your implementation of the neural network to solve the genetic algorithim question? I was talking about using the genetic algorithim to trainthe neural network like in the link I sent u and using that genetic algorithim-trained neural network to learn how to use the genome in the lab rat race question to solve it... does that seem redundant?
Oh yeah, what r ur thoughts on my question in the Sandox?
@TAbraham I've had a look at the question in the sandbox previously. It's very difficult to find a good way of asking a neural networks question. You might go through several different ideas before you find one that gets a good response in the sandbox. It reflects the difficulty of defining it
@TAbraham That sentence might not mean what you intend it to. "Training" and "learning" amount to the same thing with a neural network.
If you train a neural network that means it learns something. So what you describe is the same thing twice
@trichoplax what about asking people to design a training function to train the neural network..Actually, now I think about it, I should make it a popularity contest, not code golf.. code golf makes it too hard.. don't u agree?
@trichoplax ur're right.. what i am trying to say is u are trianging a neural network with a genetic algorithim to solve a genetic algorithim question...I think that' redundant..
@TAbraham I agree that golf may not be the best approach. It depends on how you end up defining the challenge. However, I would steer away from making it a popularity contest. Try to define the challenge specifically enough that a winner can be determined objectively
@TAbraham I don't see how you could train the neural network without using the same genetic algorithm as in the question. Otherwise it would be trained for the wrong task and wouldn't work.
It would be like training it to ride a horse and then asking it to bake a cake
@TAbraham You can define it however you like. The key is to be objective, so it can easily be measured and the winner is agreed by everyone
@TAbraham it will make sense the more you learn about it. You can learn about genetic algorithms separately from neural networks, and you can learn about neural networks separately from genetic algorithms. That might be easier than trying to learn both at once - they work well together but are less confusing one at a time.
If you can look into why you are confused then hopefully you can find out which one you need to learn more about. I'm guessing the confusion is due to a misconception about one or the other
@TAbraham you need a way that everyone can agree on who the winner is. A score is one way of doing that.
I don't think making it a popularity contest will change people's objection to it. Even a popularity contest needs to have criteria so people know what they are voting for.
You might benefit from asking other questions first, to see how the sandbox works and what makes a good question. Then you can come back to the neural networks question and have an easier time of designing it
Not every question I ask makes it out of the sandbox. Sometimes I post to the sandbox, then someone points out why it won't work, and I move on to think of a different question.
(in fact, hardly any of my questions have made it out of the sandbox...)
The trouble with neural networks is that they are solutions rather than problems. As such they can be useful for writing an answer, but are very difficult to write a question for
@TAbraham No that's my point - it's easier to use them for an answer (which I want to do). I wouldn't try to write a question about them because I know it's hard to do (no one has been able to so far)
For example, the GA question isn't asking for neural networks, but people can use neural networks if they want to. Making a question that can't use anything else is difficult to define
If you want to write an image recognition question, you could write it without mentioning neural networks, and then there will be lots of different answers, some of which might use neural networks and others might not
Image recognition is a very hard topic, so you might not get many answers on a question like that. You might want to try simpler challenges than recognising a cat. For example, something like this
@TAbraham then you are free to define your challenge with that in mind - no one is going to stop you, it will just be a lot harder to work out how to define it well.
@TAbraham those were just examples of image processing, to show that even very simplified image processing can be challenging to write answers for
@TAbraham Have you seen any questions that you think could have neural network answers? Maybe you could try writing a neural network answer, then you'll have more insight into how to ask a question that can request such answers
@trichoplax actually, I am writing neural networks now, and it seems very easy.. I already made a neuron with a training function for it...I don't know if it works but it seems easy.. I just need to implement a network ...
@TAbraham Lol I know. But if you showed me your work I wouldn't know if it is right or not. If you make it do something then we can both see whether it is right or not...
It's like making a cake is easy if no one is going to eat it - as long as it looks ok there's nothing to worry about
I wish it were possible to get small, focused answers on meta. All I can seem to get is [Huge prelude] followed by [bulleted list] followed by [obfuscated conclusion] that sounds like "I agree in principle but really I disagree."
I wish it were possible to get small, focused questions on meta. All I seem to get is [Broad Opinion] followed by [sloppy justification] that sounds like "I don't like what just happened on main."
I sent a mail to OSI asking if I could use their logo to promote my Open Source Q&A proposal. Their general manager replied saying they'd like to have an official OSI Open Source Q&A site on Stack Exchange and asked if I'd like to help make that happen O.o
"running open source projects" seems to fall under "Can I do this with X license", but I don't like the "how do I get more people on my project" questions
I do think the "How can I monetize with open source" is a great question though
@Sp3000 I've been playing around with CellularAutomaton a bit. Wire in 81 characters: CellularAutomaton[{{0,2,3,If[0<Count[#,1,2]<3,1,3]}[[#[[2,2]]+1]]&,{},{1,1}},i,n]
In principle you could specify it as a single number, but I think it would have a "few" more digits than this code :D
I was able to encode the game of life in a single number though: CellularAutomaton[{224,{2,{{2,2,2},{2,1,2},{2,2,2}}},{1,1}},i,n] (the other parameters define the neighbourhood)
honestly, I really don't care... but I consider prompts different from stdout... but I think rules that require printing for such a simple challenge are ill-designed anyway.
@Optimizer I don't know... the result of the Print goes to the exact same console as the evaluated expression. and we don't require p or o for CJam either.
In my opinion, such answers should be valid unless the OP states otherwise.
These answers often reflect a different slant to the problem, and we all like variety, don't we? I don't think we should remove this dimension of creativity by default, but if the OP doesn't want such answers, they can o...
most challenge authors who want a printed result don't seem to accept REPL answer in my experience
@PeterTaylor Algorithms question: you're given a tree with N vertices. You're to given each vertex an integer cost in [1,N], such that no two vertices connected by an edge have the same cost. Minimise the overall cost.
It seems there should be a polynomial solution.
Is there such a problem as "find the maximum non-redundant vertex cover?" It might be possible to base a greedy approach on that.
I always feel really awkward calling my team leader over to say "Hey, am I going down the right path?" followed by a short tour of code that I spent the last 7 hours working on.
