|
Thread Rules
1. This is not a "do my homework for me" thread. If you have specific questions, ask, but don't post an assignment or homework problem and expect an exact solution.
2. No recruiting for your cockamamie projects (you won't replace facebook with 3 dudes you found on the internet and $20)
3. If you can't articulate why a language is bad, don't start slinging shit about it. Just remember that nothing is worse than making CSS IE6 compatible.
4. Use [code] tags to format code blocks. |
On April 23 2013 01:40 icystorage wrote:
@mcc
im actually working on the perceptron-like multi-layer network with backpropagation and im kinda relieved it matched your suggestion (means im on the right track)
what im trying to do is use NN to predict stock prices (yes it is impossible but i want to test the NN with these data)
stock has 4 prices, open, low, high and close. i want it to predict the closing price. now, is the training (weight adjustment) separate from the predicting or should i input all the necessary data (technicalities aside, just the procedure) then expect the predicted value?
I am not sure how much experience you have with NNs so I will start with saying that if you are fairly new to them, start with something simpler and when you are well acquainted with the processes move to stock predictions  I would actually start with something completely simple like function approximation or some simple classification tasks.
I am not entirely sure what you are asking, but hopefully I guessed correctly.
As for training. You should definitely split your data into at least two sets. One group for training, another one for testing. There is whole science around correct setup of training data, but general answer is, do not mix them. Ideally you would want even more sets for different iterations and their evaluation, but you can ignore that at the start, or read up on it properly
You also need to think quite a lot about input of your network. How much data per prediction do you want it to be based on. Basically how far into the past should the NN look for predicting today's prices. And how detailed it should be. For example you can go 2 last days with all four points and before that use only closing price.
The general training process is :
1) setup weights randomly (or another starting setup)
2) while you are not satisfied
-input one vector of data and run the network
-calculate error based on actual and expected output
-use backpropagation to change weights
3) now test it on testing data and calculate error
4) if satisfied you are done if not
5) either redesign training data or redesign your NN topology/other attributes
|
On April 23 2013 00:18 mcc wrote:Show nested quote +On April 22 2013 23:39 icystorage wrote:
Um okay
1) you can use neural networks to predict an output right?
2) how do you make it predict? do you have to train it?
3) by training it, it adjusts its weight right?
4) if i give NN a pattern of 1,1,1,1,1 will it predict the next number will be 1?
5) heres my general understanding in predicting neural networks
- train the network by feeding it data (it adjusts its weight and compares its output to your EXPECTED, not predicted, output)
- with the current configuration of the weights you now use it to get the predicted output
sorry if its confusing but im also very confused atm
1) yes 2) You can either train it or set the weights "correctly" at the start by using some other process (but you can also call it training) 3) yes 4) Not necessarily. It depends on what patterns you trained it to recognize and also how well you trained itNote that there is plenty of NN types and they have sometimes quite different ways of setting up weights. So you need to specify which type you are talking about, most likely some perceptron-like multi-layer network with backpropagation ?
Can you elaborate on how this is possible?
Sorry. I'm not challenging you. I barely know enough about NN's so I am curious myself.
A neural network for stock prices sounds logical but as you're aware stocks change very dramatically to inputs you may not have guessed and people have already done this and are paid to do this work. Still, I'm sure your results will be interesting.
6) Here's one more question from me: Do neural networks only work for linear correlations?
And to icystorage:
How many days ahead are you trying to predict? Using how much training data?
If you are only going to look at the past 2 days of data as input it may be easier to use a moving average.
|
On April 23 2013 02:21 obesechicken13 wrote:Show nested quote +On April 23 2013 00:18 mcc wrote:On April 22 2013 23:39 icystorage wrote:
Um okay
1) you can use neural networks to predict an output right?
2) how do you make it predict? do you have to train it?
3) by training it, it adjusts its weight right?
4) if i give NN a pattern of 1,1,1,1,1 will it predict the next number will be 1?
5) heres my general understanding in predicting neural networks
- train the network by feeding it data (it adjusts its weight and compares its output to your EXPECTED, not predicted, output)
- with the current configuration of the weights you now use it to get the predicted output
sorry if its confusing but im also very confused atm
1) yes 2) You can either train it or set the weights "correctly" at the start by using some other process (but you can also call it training) 3) yes 4) Not necessarily. It depends on what patterns you trained it to recognize and also how well you trained itNote that there is plenty of NN types and they have sometimes quite different ways of setting up weights. So you need to specify which type you are talking about, most likely some perceptron-like multi-layer network with backpropagation ? Can you elaborate on how this is possible? Sorry. I'm not challenging you. I barely know enough about NN's so I am curious myself. A neural network for stock prices sounds logical but as you're aware stocks change very dramatically to inputs you may not have guessed and people have already done this and are paid to do this work. Still, I'm sure your results will be interesting. 6) Here's one more question from me: Do neural networks only work for linear correlations? And to icystorage: How many days ahead are you trying to predict? Using how much training data? If you are only going to look at the past 2 days of data as input it may be easier to use a moving average.
Neural networks are specifically good at non-linear problems. It's basically consecutive matrix multiplication in order to reveal non-linear relationships between inputs and outputs.
|
On April 23 2013 02:30 RoyGBiv_13 wrote:Show nested quote +On April 23 2013 02:21 obesechicken13 wrote:On April 23 2013 00:18 mcc wrote:On April 22 2013 23:39 icystorage wrote:
Um okay
1) you can use neural networks to predict an output right?
2) how do you make it predict? do you have to train it?
3) by training it, it adjusts its weight right?
4) if i give NN a pattern of 1,1,1,1,1 will it predict the next number will be 1?
5) heres my general understanding in predicting neural networks
- train the network by feeding it data (it adjusts its weight and compares its output to your EXPECTED, not predicted, output)
- with the current configuration of the weights you now use it to get the predicted output
sorry if its confusing but im also very confused atm
1) yes 2) You can either train it or set the weights "correctly" at the start by using some other process (but you can also call it training) 3) yes 4) Not necessarily. It depends on what patterns you trained it to recognize and also how well you trained itNote that there is plenty of NN types and they have sometimes quite different ways of setting up weights. So you need to specify which type you are talking about, most likely some perceptron-like multi-layer network with backpropagation ? Can you elaborate on how this is possible? Sorry. I'm not challenging you. I barely know enough about NN's so I am curious myself. A neural network for stock prices sounds logical but as you're aware stocks change very dramatically to inputs you may not have guessed and people have already done this and are paid to do this work. Still, I'm sure your results will be interesting. 6) Here's one more question from me: Do neural networks only work for linear correlations? And to icystorage: How many days ahead are you trying to predict? Using how much training data? If you are only going to look at the past 2 days of data as input it may be easier to use a moving average. Neural networks are specifically good at non-linear problems. It's basically consecutive matrix multiplication in order to reveal non-linear relationships between inputs and outputs.
I may have used the wrong words.
What if you wanted to predict something that you knew had a roughly exponential growth?
Like say you were given a set of data that matched closed to the input of x^2 + 2y^2 + 5 = z
but you were not given the equation and you wanted to use Neural networks to find it. Could you?
|
hi, this is my first post on this thread
I'd like to know what do I need to learn to make an app for android (I'd like to do something similar to a social network)
I'm a mathematics student and there I have done two subjects on programming (the first one was introductory, the second one was on algorithmics: algorithmic efficiency, sorting...), on both of them we programmed using C++. I also know a bit of java, I watched the videos of this course: http://www.academicearth.org/courses/programming-methodology .
What should I try to learn next?
thanks in advance ^^
|
On April 23 2013 06:41 ekra wrote:hi, this is my first post on this thread I'd like to know what do I need to learn to make an app for android (I'd like to do something similar to a social network) I'm a mathematics student and there I have done two subjects on programming (the first one was introductory, the second one was on algorithmics: algorithmic efficiency, sorting...), on both of them we programmed using C++. I also know a bit of java, I watched the videos of this course: http://www.academicearth.org/courses/programming-methodology . What should I try to learn next? thanks in advance ^^
You should learn making an app for android. If nothing else their tutorials will give you more prereqs to do.
|
On April 23 2013 06:41 ekra wrote:hi, this is my first post on this thread I'd like to know what do I need to learn to make an app for android (I'd like to do something similar to a social network) I'm a mathematics student and there I have done two subjects on programming (the first one was introductory, the second one was on algorithmics: algorithmic efficiency, sorting...), on both of them we programmed using C++. I also know a bit of java, I watched the videos of this course: http://www.academicearth.org/courses/programming-methodology . What should I try to learn next? thanks in advance ^^
http://developer.android.com/training/basics/firstapp/index.html
Very very easy, and complete documentation.
|
@mcc
yes i am really new and just trying to learn about NNs through reading. thank you for that elaborate answer, you have really cleared the air inside the confusion in my brain. +1
@obesechicken
we were planning to use 15 years worth of daily data. is that too much for training? overtraining? and a month worth of data for testing.
|
On April 23 2013 05:12 obesechicken13 wrote:Show nested quote +On April 23 2013 02:30 RoyGBiv_13 wrote:On April 23 2013 02:21 obesechicken13 wrote:On April 23 2013 00:18 mcc wrote:On April 22 2013 23:39 icystorage wrote:
Um okay
1) you can use neural networks to predict an output right?
2) how do you make it predict? do you have to train it?
3) by training it, it adjusts its weight right?
4) if i give NN a pattern of 1,1,1,1,1 will it predict the next number will be 1?
5) heres my general understanding in predicting neural networks
- train the network by feeding it data (it adjusts its weight and compares its output to your EXPECTED, not predicted, output)
- with the current configuration of the weights you now use it to get the predicted output
sorry if its confusing but im also very confused atm
1) yes 2) You can either train it or set the weights "correctly" at the start by using some other process (but you can also call it training) 3) yes 4) Not necessarily. It depends on what patterns you trained it to recognize and also how well you trained itNote that there is plenty of NN types and they have sometimes quite different ways of setting up weights. So you need to specify which type you are talking about, most likely some perceptron-like multi-layer network with backpropagation ? Can you elaborate on how this is possible? Sorry. I'm not challenging you. I barely know enough about NN's so I am curious myself. A neural network for stock prices sounds logical but as you're aware stocks change very dramatically to inputs you may not have guessed and people have already done this and are paid to do this work. Still, I'm sure your results will be interesting. 6) Here's one more question from me: Do neural networks only work for linear correlations? And to icystorage: How many days ahead are you trying to predict? Using how much training data? If you are only going to look at the past 2 days of data as input it may be easier to use a moving average. Neural networks are specifically good at non-linear problems. It's basically consecutive matrix multiplication in order to reveal non-linear relationships between inputs and outputs. I may have used the wrong words. What if you wanted to predict something that you knew had a roughly exponential growth? Like say you were given a set of data that matched closed to the input of x^2 + 2y^2 + 5 = z but you were not given the equation and you wanted to use Neural networks to find it. Could you?
Okay, the question to ask is...what is your actual end goal? Is using a Neural Network a requirement in your mind? Or are you primarily trying to determine a formula to match given data points, and then extrapolate the data further?
Yes, you could, technically, use a Neural Network to do such a thing. However, it would be extremely inefficient to do so, given that your time would be much better spent looking into Discrete Mathematics, where there are already established theorems and equations for such tasks. I think the study is called Recurrence relations...it's been awhile though, so I could be wrong.
EDIT: Sorry, I appear to have randomly jumped into the conversation at an odd point. My point really only applies to deriving a formula from data, not to any sort of predictive software.
|
On April 22 2013 10:20 obesechicken13 wrote:
I'm starting to dislike coursera again. I hate formal courses. I was able to do a project today: a rock paper scissors lizard spock game, but I didn't like having to watch video lectures.
Haha, ended up signing up for coursera's intro interaactive python after reading a few comments in this thread.
I haven't actually gone through any of the material/lectures yet, but it looks like they have the concepts in written form if you go to Class > Video Lectures > Week# > Week# Concepts (button that looks like 'i').
|
|
|
On April 23 2013 10:26 Ame wrote:Show nested quote +On April 22 2013 10:20 obesechicken13 wrote:
I'm starting to dislike coursera again. I hate formal courses. I was able to do a project today: a rock paper scissors lizard spock game, but I didn't like having to watch video lectures. Haha, ended up signing up for coursera's intro interaactive python after reading a few comments in this thread. I haven't actually gone through any of the material/lectures yet, but it looks like they have the concepts in written form if you go to Class > Video Lectures > Week# > Week# Concepts (button that looks like 'i').
Well for me, I really don't mind the lectures. I realize that most of these professors are well more educated and actually has the ability to teach other than doing research (unlike most of my professors at a state university).
I'm currently taking the scala course, and there's never a time when I'm bored by the lectures. But at the same time, it could be due to the way I take the online course: I spend 20-30min worth of lectures a day, so I don't burn myself out since I have other classwork to do. And plus, it's just easier to retain what I learn in chunks. It sounds corny but I also take notes and pause the video at almost every slide until I can understand the damn thing in my own words lol.
Considering the class you're taking is fairly introductory, I can see why you get bored of the lectures. But maybe try taking it by a little each day--you have the whole week to finish it. I can't guarantee anything, but it's worth a shot.
|
On April 23 2013 17:11 billy5000 wrote:
Well for me, I really don't mind the lectures. I realize that most of these professors are well more educated and actually has the ability to teach other than doing research (unlike most of my professors at a state university).
I'm currently taking the scala course, and there's never a time when I'm bored by the lectures. But at the same time, it could be due to the way I take the online course: I spend 20-30min worth of lectures a day, so I don't burn myself out since I have other classwork to do. And plus, it's just easier to retain what I learn in chunks. It sounds corny but I also take notes and pause the video at almost every slide until I can understand the damn thing in my own words lol.
Considering the class you're taking is fairly introductory, I can see why you get bored of the lectures. But maybe try taking it by a little each day--you have the whole week to finish it. I can't guarantee anything, but it's worth a shot.
For me, given my current lifestyle, going through the lectures is probably one of the more productive things I can do anyways (repetition aiding learning etc etc). So yep, I might as well watch them.
Scala course eh. Might add/take that one too. :<
|
On April 23 2013 02:21 obesechicken13 wrote:Show nested quote +On April 23 2013 00:18 mcc wrote:On April 22 2013 23:39 icystorage wrote:
Um okay
1) you can use neural networks to predict an output right?
2) how do you make it predict? do you have to train it?
3) by training it, it adjusts its weight right?
4) if i give NN a pattern of 1,1,1,1,1 will it predict the next number will be 1?
5) heres my general understanding in predicting neural networks
- train the network by feeding it data (it adjusts its weight and compares its output to your EXPECTED, not predicted, output)
- with the current configuration of the weights you now use it to get the predicted output
sorry if its confusing but im also very confused atm
1) yes 2) You can either train it or set the weights "correctly" at the start by using some other process (but you can also call it training) 3) yes 4) Not necessarily. It depends on what patterns you trained it to recognize and also how well you trained itNote that there is plenty of NN types and they have sometimes quite different ways of setting up weights. So you need to specify which type you are talking about, most likely some perceptron-like multi-layer network with backpropagation ? Can you elaborate on how this is possible? Sorry. I'm not challenging you. I barely know enough about NN's so I am curious myself. A neural network for stock prices sounds logical but as you're aware stocks change very dramatically to inputs you may not have guessed and people have already done this and are paid to do this work. Still, I'm sure your results will be interesting. 6) Here's one more question from me: Do neural networks only work for linear correlations? And to icystorage: How many days ahead are you trying to predict? Using how much training data? If you are only going to look at the past 2 days of data as input it may be easier to use a moving average.
Well if you have training data like this [1,1,1,1,1 -> 0] then properly trained network will answer 0 to input [1,1,1,1,1] My point was more that you can train the network to answer whatever you want, so asking if neural network will answer 1 to [1,1,1,1,1] is dependent on what you trained it to do.
Of course some NNs use some self-organization mechanisms instead of typical training to set their weights and they work somewhat differently.
On April 23 2013 05:12 obesechicken13 wrote:Show nested quote +On April 23 2013 02:30 RoyGBiv_13 wrote:On April 23 2013 02:21 obesechicken13 wrote:On April 23 2013 00:18 mcc wrote:On April 22 2013 23:39 icystorage wrote:
Um okay
1) you can use neural networks to predict an output right?
2) how do you make it predict? do you have to train it?
3) by training it, it adjusts its weight right?
4) if i give NN a pattern of 1,1,1,1,1 will it predict the next number will be 1?
5) heres my general understanding in predicting neural networks
- train the network by feeding it data (it adjusts its weight and compares its output to your EXPECTED, not predicted, output)
- with the current configuration of the weights you now use it to get the predicted output
sorry if its confusing but im also very confused atm
1) yes 2) You can either train it or set the weights "correctly" at the start by using some other process (but you can also call it training) 3) yes 4) Not necessarily. It depends on what patterns you trained it to recognize and also how well you trained itNote that there is plenty of NN types and they have sometimes quite different ways of setting up weights. So you need to specify which type you are talking about, most likely some perceptron-like multi-layer network with backpropagation ? Can you elaborate on how this is possible? Sorry. I'm not challenging you. I barely know enough about NN's so I am curious myself. A neural network for stock prices sounds logical but as you're aware stocks change very dramatically to inputs you may not have guessed and people have already done this and are paid to do this work. Still, I'm sure your results will be interesting. 6) Here's one more question from me: Do neural networks only work for linear correlations? And to icystorage: How many days ahead are you trying to predict? Using how much training data? If you are only going to look at the past 2 days of data as input it may be easier to use a moving average. Neural networks are specifically good at non-linear problems. It's basically consecutive matrix multiplication in order to reveal non-linear relationships between inputs and outputs. I may have used the wrong words. What if you wanted to predict something that you knew had a roughly exponential growth? Like say you were given a set of data that matched closed to the input of x^2 + 2y^2 + 5 = z but you were not given the equation and you wanted to use Neural networks to find it. Could you?
Are you asking about neural network approximating the function or are you asking about neural network recognizing the function ?
As for approximation there is actually reasonable number of theoretical results. Basically you can say that perceptron-like networks with at least one hidden layer are "universal approximators" as long as activation function is "nice". If I recall correctly that basically means anything but polynomial activation function. As for classes of functions that can be "approximated" - continuous functions can and even functions that are not continuous in some points if they satisfy some criteria.
As for "recognizing the function" as in providing you with symbolic representation, that depends what exactly you want. You can train it to recognize general type of function, or you can train it to provide you with parameters of the specific class of functions, but then it only works if on the input there will be instances of that class of functions. But as someone pointed out below there are most likely better tools for doing that.
|
On April 23 2013 22:14 mcc wrote:Show nested quote +On April 23 2013 05:12 obesechicken13 wrote:On April 23 2013 02:30 RoyGBiv_13 wrote:On April 23 2013 02:21 obesechicken13 wrote:On April 23 2013 00:18 mcc wrote:On April 22 2013 23:39 icystorage wrote:
Um okay
1) you can use neural networks to predict an output right?
2) how do you make it predict? do you have to train it?
3) by training it, it adjusts its weight right?
4) if i give NN a pattern of 1,1,1,1,1 will it predict the next number will be 1?
5) heres my general understanding in predicting neural networks
- train the network by feeding it data (it adjusts its weight and compares its output to your EXPECTED, not predicted, output)
- with the current configuration of the weights you now use it to get the predicted output
sorry if its confusing but im also very confused atm
1) yes 2) You can either train it or set the weights "correctly" at the start by using some other process (but you can also call it training) 3) yes 4) Not necessarily. It depends on what patterns you trained it to recognize and also how well you trained itNote that there is plenty of NN types and they have sometimes quite different ways of setting up weights. So you need to specify which type you are talking about, most likely some perceptron-like multi-layer network with backpropagation ? Can you elaborate on how this is possible? Sorry. I'm not challenging you. I barely know enough about NN's so I am curious myself. A neural network for stock prices sounds logical but as you're aware stocks change very dramatically to inputs you may not have guessed and people have already done this and are paid to do this work. Still, I'm sure your results will be interesting. 6) Here's one more question from me: Do neural networks only work for linear correlations? And to icystorage: How many days ahead are you trying to predict? Using how much training data? If you are only going to look at the past 2 days of data as input it may be easier to use a moving average. Neural networks are specifically good at non-linear problems. It's basically consecutive matrix multiplication in order to reveal non-linear relationships between inputs and outputs. I may have used the wrong words. What if you wanted to predict something that you knew had a roughly exponential growth? Like say you were given a set of data that matched closed to the input of x^2 + 2y^2 + 5 = z but you were not given the equation and you wanted to use Neural networks to find it. Could you? Are you asking about neural network approximating the function or are you asking about neural network recognizing the function ? As for approximation there is actually reasonable number of theoretical results. Basically you can say that perceptron-like networks with at least one hidden layer are "universal approximators" as long as activation function is "nice". If I recall correctly that basically means anything but polynomial activation function. As for classes of functions that can be "approximated" - continuous functions can and even functions that are not continuous in some points if they satisfy some criteria. As for "recognizing the function" as in providing you with symbolic representation, that depends what exactly you want. You can train it to recognize general type of function, or you can train it to provide you with parameters of the specific class of functions, but then it only works if on the input there will be instances of that class of functions. But as someone pointed out below there are most likely better tools for doing that.
I meant for approximating. If I recall correctly that basically means anything but polynomial activation function. . I see... I see. Now explain it to me like I am 5?
|
On April 23 2013 22:50 obesechicken13 wrote:Show nested quote +On April 23 2013 22:14 mcc wrote:On April 23 2013 05:12 obesechicken13 wrote:On April 23 2013 02:30 RoyGBiv_13 wrote:On April 23 2013 02:21 obesechicken13 wrote:On April 23 2013 00:18 mcc wrote:On April 22 2013 23:39 icystorage wrote:
Um okay
1) you can use neural networks to predict an output right?
2) how do you make it predict? do you have to train it?
3) by training it, it adjusts its weight right?
4) if i give NN a pattern of 1,1,1,1,1 will it predict the next number will be 1?
5) heres my general understanding in predicting neural networks
- train the network by feeding it data (it adjusts its weight and compares its output to your EXPECTED, not predicted, output)
- with the current configuration of the weights you now use it to get the predicted output
sorry if its confusing but im also very confused atm
1) yes 2) You can either train it or set the weights "correctly" at the start by using some other process (but you can also call it training) 3) yes 4) Not necessarily. It depends on what patterns you trained it to recognize and also how well you trained itNote that there is plenty of NN types and they have sometimes quite different ways of setting up weights. So you need to specify which type you are talking about, most likely some perceptron-like multi-layer network with backpropagation ? Can you elaborate on how this is possible? Sorry. I'm not challenging you. I barely know enough about NN's so I am curious myself. A neural network for stock prices sounds logical but as you're aware stocks change very dramatically to inputs you may not have guessed and people have already done this and are paid to do this work. Still, I'm sure your results will be interesting. 6) Here's one more question from me: Do neural networks only work for linear correlations? And to icystorage: How many days ahead are you trying to predict? Using how much training data? If you are only going to look at the past 2 days of data as input it may be easier to use a moving average. Neural networks are specifically good at non-linear problems. It's basically consecutive matrix multiplication in order to reveal non-linear relationships between inputs and outputs. I may have used the wrong words. What if you wanted to predict something that you knew had a roughly exponential growth? Like say you were given a set of data that matched closed to the input of x^2 + 2y^2 + 5 = z but you were not given the equation and you wanted to use Neural networks to find it. Could you? Are you asking about neural network approximating the function or are you asking about neural network recognizing the function ? As for approximation there is actually reasonable number of theoretical results. Basically you can say that perceptron-like networks with at least one hidden layer are "universal approximators" as long as activation function is "nice". If I recall correctly that basically means anything but polynomial activation function. As for classes of functions that can be "approximated" - continuous functions can and even functions that are not continuous in some points if they satisfy some criteria. As for "recognizing the function" as in providing you with symbolic representation, that depends what exactly you want. You can train it to recognize general type of function, or you can train it to provide you with parameters of the specific class of functions, but then it only works if on the input there will be instances of that class of functions. But as someone pointed out below there are most likely better tools for doing that. I meant for approximating. Show nested quote +If I recall correctly that basically means anything but polynomial activation function. . I see... I see. Now explain it to me like I am 5?
I do not think that is possible, unless you are very gifted 5-year old
I will try. Neurons in the neural network have weighted input, output of the neuron is f(x) where x is that input and f is activation function. Most commonly it is sigmoid function, but that is not necessary. The result says that any activation function is good enough for the network to be universal approximator as long as f is non-polynomial.
In general it is described here : Universal approximation theorem.
The specific result I talked about is in link .
Hopefully the link works.
EDIT: for completeness, being non-polynomial is not the only requirement, it needs to be "nice" in other ways, but those are fairly common requirements, unlike being non-polynomial. The other requirements are "locally bounded piecewise continuous".
|
I'm pretty new to responsive design so I was wondering if someone could recommend me a lightweight CSS framework that is i.e.a 960 grid system, 7+ compatible, cross browser compatible, has all the major media queries covered, that maybe comes with a very barebones JS function that covers all the crappy old browsers. I'm designing a very basic responsive site because a lot of the users are iphone people so I just need to hide certain page elements at certain widths etc. I've triedaphy messing with bootstrap but it's a little bloated for my use (I'm also using Genesis for wordpress so I'm making a child theme). Best one I like so far is Gumby. Anyone else have any other suggestions?
As a bonus, are there any preset typography/styles out there that look nice and is web friendly?
Eta: I think I found exactly what I was looking for in skeleton.
http://www.getskeleton.com/#
|
As for approximation there is actually reasonable number of theoretical results. Basically you can say that perceptron-like networks with at least one hidden layer are "universal approximators" as long as activation function is "nice". If I recall correctly that basically means anything but polynomial activation function. As for classes of functions that can be "approximated" - continuous functions can and even functions that are not continuous in some points if they satisfy some criteria.
|
On April 24 2013 03:58 lannisport wrote:I'm pretty new to responsive design so I was wondering if someone could recommend me a lightweight CSS framework that is i.e.a 960 grid system, 7+ compatible, cross browser compatible, has all the major media queries covered, that maybe comes with a very barebones JS function that covers all the crappy old browsers. I'm designing a very basic responsive site because a lot of the users are iphone people so I just need to hide certain page elements at certain widths etc. I've triedaphy messing with bootstrap but it's a little bloated for my use (I'm also using Genesis for wordpress so I'm making a child theme). Best one I like so far is Gumby. Anyone else have any other suggestions? As a bonus, are there any preset typography/styles out there that look nice and is web friendly? Eta: I think I found exactly what I was looking for in skeleton. http://www.getskeleton.com/#
http://960.gs/ is popular as well.
|
1019 Posts
Is it possible to teach yourself HTML and be pretty good at it? I'm planning on studying intensively on my own this summer.
Also, should I learn HTML first or PHP, or should I study both at the same time?
|
|
|
|
|
|
EPT
