COMPGROUPS.NET | Search | Post Question | Groups | Stream | About | Register

### Performance of neural networks

• Email
• Follow

Hello guys,

I have a question that might seem simple but I don't seem to be able to find the answer.

I have a very simple neural network which I train using the Neural Network Toolbox GUI. My performance criterion is the MSE.

I am wondering why each time I train the network I get different results for the performance.

What can I report as the performance of the network. Should I run the algorithm for ten or more times and pick the best performing one?

Thank you,

George
 0
Reply mavromatidis.g (12) 6/26/2012 2:39:07 PM

See related articles to this posting

"Georgios Mavromatidis" <mavromatidis.g@gmail.com> wrote in message <jschib\$36f\$1@newscl01ah.mathworks.com>...
> Hello guys,
>
> I have a question that might seem simple but I don't seem to be able to find the answer.
>
> I have a very simple neural network which I train using the Neural Network Toolbox GUI. My performance criterion is the MSE.
>
> I am wondering why each time I train the network I get different results for the performance.

The initial weights are chosen randomly. This results in different local minima solutions.

> What can I report as the performance of the network. Should I run the algorithm for ten or more times and pick the best performing one?

I typically run the algorithm Ntrials = 10 or more times for each candidate value of the number of hidden nodes, H. Then, for each value of H I tabulate the summary statistics
min, median, mean, std, max.

Be sure to specify the initial RNG seed or state so that you can replicate the run.

Typically, the net with the smallest H that meets the performance goal is chosen. I typically use MSEgoal = 0.01*mean(var(t')) to obtain R^2 <~ 0.99.

Sometimes a weighted average of the outputs from several good nets are combined  to form an ensemble.

Hope this helps.

Greg
 0
Reply heath (3990) 6/26/2012 3:26:07 PM

Hello Greg,

Best,

George

"Greg Heath" <heath@alumni.brown.edu> wrote in message <jsckaf\$g8s\$1@newscl01ah.mathworks.com>...
> "Georgios Mavromatidis" <mavromatidis.g@gmail.com> wrote in message <jschib\$36f\$1@newscl01ah.mathworks.com>...
> > Hello guys,
> >
> > I have a question that might seem simple but I don't seem to be able to find the answer.
> >
> > I have a very simple neural network which I train using the Neural Network Toolbox GUI. My performance criterion is the MSE.
> >
> > I am wondering why each time I train the network I get different results for the performance.
>
> The initial weights are chosen randomly. This results in different local minima solutions.
>
> > What can I report as the performance of the network. Should I run the algorithm for ten or more times and pick the best performing one?
>
> I typically run the algorithm Ntrials = 10 or more times for each candidate value of the number of hidden nodes, H. Then, for each value of H I tabulate the summary statistics
> min, median, mean, std, max.
>
> Be sure to specify the initial RNG seed or state so that you can replicate the run.
>
> Typically, the net with the smallest H that meets the performance goal is chosen. I typically use MSEgoal = 0.01*mean(var(t')) to obtain R^2 <~ 0.99.
>
> Sometimes a weighted average of the outputs from several good nets are combined  to form an ensemble.
>
> Hope this helps.
>
> Greg
 0
Reply mavromatidis.g (12) 6/27/2012 11:35:08 AM

2 Replies
25 Views

Similar Articles

12/12/2013 1:28:21 AM
[PageSpeed]

Similar Artilces:

Neural network performance variation
Hello, I am currently trying to understand neural networks and after a little digging in the documentations, I made a simple code to create a network for a XOR gate. input=[0 0 1 1;0 1 0 1]; output=[0 1 1 0]; net=feedforwardnet(5,'traingd'); % net.performFcn='msereg'; net.trainParam.epochs=100000; net.trainParam.goal=0; net.trainParam.show=25; net.trainParam.lr=0.001; net.trainParam.min_grad=1e-10; net.trainParam.max_fail=20; net=train(net,input,output); The funny thing is, sometimes when the NN training starts, it stops immediately, saying validation stop and sometimes the ...

Neural Network Architecture with MATLAB Neural Network Toolbox
Dear neural network experts, I would like to make a neural network model for a situation when some of the input variables can interact, while others don't have any interactions. This information is based on logical considerations and I have to include it into the architecture of the net. For example, I have 3 different groups of input variables. Within each groups the variables are disjunct (i.e. no interaction). Variables of different groups may interact. This would result in either - restrictions to some weights of a network consisting of completely inter-linked neurons (i.e. some weight...

Wich Neural Network program got comitee network ? We have to construct
Wich Neural Network program got comitee network ? We have to construct it by ourself or this program have automated this process ? I did not see this kind of network in Nerosolution, MathLab, EasyNN, SAS and Statistica. Which program got it ? Thanks you. On 14 Sep 2005 06:44:33 -0700, "Kheor" <kheor_@hotmail.com> wrote: >Wich Neural Network program got comitee network ? We have to construct >it by ourself or this program have automated this process ? > >I did not see this kind of network in Nerosolution, MathLab, EasyNN, >SAS and Statistica. Which program go...

PCA Neural Networks
Im interested in using a PCA neural network to initialise my SOM along the first principal component (Linear Initialisation).... I tried to learn standard PCA, but was finding it far too hard, I cant get a grip on finding eigenvectors and eigenvalues for my covariance matrix so im going to try the PCA NN method... which Oja introduced I believe? I am finding it hard to find resources... can anyone help me with the learning process for it? thanks ... Chris milesy@blueyonder.co.uk (Milesy) wrote in message news:<0307241707401@news-lhr.cableinet.net>... > Im interested in using a P...

logic from a neural network
I was wondering if anyone has figured out how to perform sequential logic using a neural net. Is there any theory, conjecture, design methods, or experiments pursuant to this question? It seems a miracle to me, that the human brain, a gigantic squishy imprecise adaptive neural net, is able to play chess, prove theorems, generate plans, and so forth (at least, some brains can do so). Can anyone explain this, or replicate it artificially? -- Rich RichD wrote: > I was wondering if anyone has figured out how to > perform sequential logic using a neural net. Is there > any theory,...

Neural Networks Ensemble
Hello, I would appreciate if any one direct me to online good explanation of Neural Networks Ensemble Regards, -- Motaz K. Saad On 14 Oct, 09:47, "Motaz K. Saad" <motaz.s...@gmail.com> wrote: > Hello, > > I would appreciate if any one direct me to online good explanation of > Neural Networks Ensemble > > Regards, > -- > Motaz K. Saad Not specific to neural nets, but there is a nn ensemble in there. http://www.tiberius.biz/pakdd07.html On Oct 14, 8:04=A0am, baldrick <philbrier...@hotmail.com> wrote: > On 14 Oct, 09:47, "Motaz K. Saad&...

Neural network opencv
I am using neural network of opencv in order to make face detection, but i have some problem to use this neural network.the fase of training is very fast and i try to train the net with more of 300 samples,the net does not work more. Someone knows to me to help?You know a good neural netwiork library in c++? The performance of this neural network implemented with ipencv,as they are respect to other nets? thank you On Jul 11, 7:21 am, g.trovalu...@libero.it wrote: > I am using neural network of opencv From my own experience you have to reinvent a lot of wheels. Other tools can help to insp...

Dear all; I want to build a neural network for weather forecast, through the draft and I read in this area a number of issues, but now how can I draw my mind at the beginning of building the network?? Can you help me in this regard, or give me some sources Thanks ...

SAS Neural Network
I know that SAS has neural network functionality built in to its enterprise data miner. Can its NN select the variables (through optimal brain damage, etc) that most impact the target? Thanks. Alan The NN automatically does not select the variables that impact the target the most. But you can select the significant variables by using a Tree node in the SAS EM, followed by a neural network node. In the options for the Tree, select the option that will forward all signifant variables to the next node (I don't exactly remember the wordings of this option, but they are present in the Tree ...

We have a need to develop an application that can identify the elements present in spectra obtained from a scintillation detector. A given sample can contain multiple elements sometimes with peaks very close to one another or embedded in one another and very small signals or far away sources etc sometimes yield very small peaks only a few counts above the background. We need to achieve a high degree of accuracy in our detection system. We have pursued a number of approaches to this problem. We have a standard least squares peak fitting algorithm that we have used with some degree of succe...

And gate in Neural Network
Hi People, I am interested to build AND gate (as neural network) using FPGA. I am using Xilinx webpack 7. If anyone can advice on how can I build one using basic components such as adder, multiplier.. lookup table etc... thanks in advance. John ...

Neural Network 3D
turn the 2D surface network into a 3D spatial topology network. Hints: the topology basically doesn't change at all - just have to introduce a third dimension to your array (that's really it for the topology). But forget about the mesh as a representation - you will have to come up with a way to draw lines between the points/ indices of the array. I already showed that to some of you. When you draw it, you will have to draw it separately in all three directions (others, who have done that already can experiment with triangular 3D networks, for example): First: For i... For j... ...

Network performance testing
Hi all, does anyone know of a network performance testing tool primarily to simulate SMB / NetBIOS / file transfers for Windows 2000 / 2003? i know of varios utilities like MS Web Application Stress Tool etc but all the network performance monitoring utilites are usually only for db / web applications. If anyone does know of a open source file transfer simulation agent to test network thoughput please let me know. Cheers ...

Neural Networks Basics
Hi, I've recently started reading about neural-nets and very new to the field. I came across a sample network that represents some Boolean functions: http://www.comp.nus.edu.sg/~pris/ArtificialNeuralNetworks/Images/NAND_OR_AND.gif .. Could someone please briefly explain why the weights shown are the way they are? Why is -1, -1, with a bias of -1.5 is the same as NAND? Any help will be much appreciated. Cheers, Michael dayzman@hotmail.com wrote: > Hi, > > I've recently started reading about neural-nets and very new to the > field. I came across a sample network that ...

Neural Networks and Time
Hi, Is there much research on the implementation of time and rhythm in the brain? It does occur to me that "keeping time" in a musical sense is not really a basic function of the neural network as such. Has anybody looked into how it is implemented in people? Regards, Ralph Ralph.little@tribaltechnology.co.uk schrieb: > Is there much research on the implementation of time and rhythm in the > brain? > It does occur to me that "keeping time" in a musical sense is not > really a basic function of the neural network as such. > > Has anybody looked into ho...

New to Neural networks
Hi to all I am very new to this is enterprising and exciting field. I have learnt a few basics about the Neural networks. Presently i am trying to collect materilas on hoe to create a handwriting recognition system. This is intended to produce as a project for my course. So I would love to have it as simple as possible. Currently I have an idea about how to interpret images and feed as input to the i/p layer the pixels. But i am struggling to strike the appropriate network, algorithms and other related stuffs. I would be extremely happy to receive direc...

Neural network forecasting
Hi, I'm trying to create a neural network to forecast the S&P 500 Index. I Have a few questions: 1)I want to use a training set, a validation set and a testing set ; I Imported the input data P and the output data T, then I pre-processed them with the command [p,minp,maxp,t,mint,maxt]= premnmx(P,T) I then divided both datasets, p and t, into three subsets: INPUTtr, INPUTval, INPUTtest, OUTPUTtr, OUTPUTval, OUTPUTtest. Is this procedure correct or should I pre-process only the training data, then standardize the validation and test data using the statistic...

Neural networks #18
Hello, I'm trying to train a neural network using the adapt function. The thing is that I want to introduce the error myself and I'm not able to change it. In other words, I don't want to have an error being the difference between the output and the target but an error which comes from the difference of other 2 signals. Anyone can give a hand? "PABLO" wrote in message <kppiho\$a5h\$1@newscl01ah.mathworks.com>... > Hello, > > I'm trying to train a neural network using the adapt function. The thing is that I want to introduce the error myself and I...

Help on Neural Networks
Hi All Can anyone provide information on pros and cons of supervised or unsupervised Neural Net methodology and how to implement the same in SAS ( how we need to prepare the data for the same ?, what procs do we use etc.) -- Thanks in advance Vishal Jain ...

Is the Internet a Neural Network?
I tend to think it's not because all the data we find on the internet was outputted directly by humans. The internet could not have gathered information only to "learn" and converge on a single output. Tell me your opinions and reason why! Thanks. On 5 Mar 2007 20:22:07 -0800, harb37@gmail.com wrote: >I tend to think it's not because all the data we find on the internet >was outputted directly by humans. The internet could not have gathered >information only to "learn" and converge on a single output. > >Tell me your opinions and reason why! Thank...