Welcome to NI-Labs Neural Nets

[GregMo]

Welcome to the NI-Labs Neural Nets project. Thanks for taking the time to look into using Neural Nets with LabVIEW. This thread is intended to foster discussion about the project, so please post any questions, comments, bug reports, or anything else relevant to the project here.

The NI-Labs Neural Net project consists of an API for training and implementing a multi-layer perceptron with back-propagation training. The API contains tools for managing training and neural net data, for training the network, and for implementing a trained network on LabVIEW desktop, real-time, and embedded computers. You simply need to supply a set of training data pairs {input, desired output} that are relevant to your application.

The project also contains an example application: a simple handwriting recognition system. The example contains VIs showing how to produce training data (sample training data is supplied,) how to train and validate the network using the training data, and how to implement the network to perform the simple handwriting recognition (a sample fully-trained network is supplied.)

There is a document installed with the project that provides a brief introduction to neural nets, as well as a detailed explanation of the API and examples.

Message Edited by Support on 08-01-2008 02:54 PM

[altenbach]

According to the NI-Labs main page, it requires:

LabVIEW 8.6

LabVIEW RT 8.6

IT does not say if it is AND or OR. 😉 Does it require both or is LabVIEW 8.6 sufficient? Form what I can tell, the demo runs just fine in 8.6 (non-RT).

As a matter of fact, the example seems to be in 8.5 and everything runs fine in 8.5. Is there anything that would prevent people from trying this in 8.5?

[GregMo]

You are absolutely correct. It can run on LabVIEW desktop or LabVIEW RT or LabVIEW Microprocessor SDK. It is written in LabVIEW 8.5, so it should be fine for any version later than that.

Thanks for pointing that out. 

[pretender.yl]

It's what I want to have

I think you should add a normlization SubVI. It's very important to ANN.

May I have you Email? GerMo. If I have other question, I can send a email to you.

My email: pretender.yl@gmail.com

[GregMo]

Thanks for the comments.

I am interested to hear more about "normalization functions." Do you mean to normalize inputs? The more specific your comments, the better!

The best way to reach me is by posting on this message board.

Thanks again,

GregMo

[pretender.yl]

The stability of ANN is sensitivity to initializing weight matrices and input data. Yout activation function is sigmoid,so then the output  < 0.1 or  >0.9, the weight matrices must change a lot .It's will use many time. I think the input and output data for trainning ANN must  normalize. Make the data between 0.1 to 0.9 is easy to train and use. I wrote a VI for that. But I can't upload in this BBS.

Furthermore，I'm very interesting in  "40-20-40" rule, May you give me your reference paper? I don't know how you get the 0.3.

[GregMo]

Thanks for the information about normalization.

Regarding the 40-20-40 rule, it is sometimes called "Fahlman's rule". Try this link, for example, or search for "fahlman 40-20-40 neural nets"

http://ieeexplore.ieee.org/Xplore/login.jsp?url=/iel5/72/22161/01031939.pdf?arnumber=1031939

Cheers,

GregMo

[Paul_Torek]

I obtained a good improvement in speed by replacing the Read XML and Write XML functions in the neural net code, with the Read/Write Anything VIs available free on the Moore Good Ideas website.  If anyone wants more info, post a reply and we can move the discussion to the Labview Advanced Virtual Architects website - because I don't seem to be able to post attachments here.

[viScience]

Is it possible to utilize ANN for such things as adaptive temperature control?  I have a very non-linear application controlling a Heat Exchanger and have tried both PID and Fuzzy Logic.  The PID works

well within certain operating ranges (such as inlet water temperature) but would need quite a bit of gain scheduling to work across all conditions.  Fuzzy Logic is more flexible but I have it set up as a 

'bias-less' controller and thus does not have very good response.