Octave's Neural Network Package

© by Michael Schmid
Last changes: 2010-04-11


SourceForge.net Logo

Introduction:

You find a more detailed description of Octave's Neural Network Package on this page. The bold letters are used in the unix short name of this project octnnetpkg for the sourceforge page.

Currently the part to commit the octnnetpkg code to the octave-forge repository is done. This means if you download the repository version of octave-forge you will also download the newest octnnetpkg.

I have to programe the complete code in my spare free time, so I'm not very fast, but one point is very important for me, testing testing and testing. So I won't release often. And this is also the reason for this simple not very stylish page ;-)

Net description:

Right now it exists only one realised neural network type. This is the multi-layer network type short MLP. This is a simple feed-forward network. So no feedback path exist. Calculation is straight forward from the input neurons to the output neurons.

Most current changes:

__trainlm: The progress messages appears as soon as one is available. This depends on the parameter "net.trainParam.show". Still only available in the repository
newff: Since version 3.2.4 the code doesn't work properly. So I had to change a detail. Still only available in the repository

Function description:

The first neural network type which will be realised is the feed-forward multi-layer perceptron. It should be possible now to use hidden layers as much as you want.
One more target is to be matlab compatible where it makes any sense!

Documentation:

Download from here:

Download:

It's inadvisable to download the package of the sourceforge server ... the package is to old.
Please download from the octave-forge project.

Known Bugs:

Independent of the operating system, the version 3.0.0 is working very well. Some problems with the versions 3.2.x occures.

  1. error(nargchk(1,2,2)) seems to be a problem, see more under http://n4.nabble.com/nargchk-always-fails-td1673882.html
  2. net.IW{2:nLayers,1} = []; seems to be the next problem under the 3.2.x versions.
The first error is fixed in version 3.2.4, so please use this version.
The second problem is fixed in the repository, so please download the octave-forge repository and copy the file "newff.m" into the corresponding directory.

Thanks to:

The list below is in order of the date. The newest is on top.

Bibliographie:

Following literature is used to build this package ...