Octave's neural network package / News: Recent posts

Hi all users of the octave's neural network package

Jhonny Fu was asking me for help with a problem and see, a bug in newff.
It's removed now in available in the repository. It's only a bug if you use only 2 input arguments, otherwise it doesn't matter.

Michael

Hi all

After a long time happens something in this package.
Now we can use 4 new functions:
mapstd: a function used in the newest toolbox version of matlab
ind2vec:
vec2ind:
dividerand: seems to be similar to my "subset"

This functions are contributed from Luiz Angelo Daros de Luca.

Hi all

Yes, this project is still alive and I don't think that I will stop it! Eventhough I'm very slow because my heavy workload.

Ok, I updated the User's Guide, you can find the new one on the google-page, but not yet at the source-forge package. I will do this tomorrow.

Than, I realised that the matlab interface of the command "newff" has changed. Matlab provides now some more arguments to pre- and or postprocess the datas. I can't test it because I don't have any access to Matlab. It seems similar to subset ... but not so nice :-)... read more

Hi all

I put a new example into the user's Guide.
Only a short example about the use of subset.

Today, the new release 0.1.9 is released (what for a sentence:-) )
As soon as the package is updated on octave-forge, both versions should be the same!
New is only the function subset and some docu improvements.

After a long time, I'm happy to establish a new function "subset" which can be used to create and optimize some data sets for a feed forward neural network. More will follow in the documentation ...

regards
Michael

I found a bug in poststd which is now corrected.
I never got a bug report or something similar, so I think no one is really using this package :-(
Or, no one has more than one output :-)

I think most users think about the vitality of this project. Currently I'm writing another software for neural networks with a GUI which won't be open source. But it will include the same algorithm, means Levenberg-Marquardt but written in C++. That's why you can see any progress currently. I think in about 3 or 4 month the next "perceptron" will be implemented in this package. So long thanks for using this package.... read more

Hi all

I'm still writing some tests! I have found some different behaviour of octave between the linux version (self compiled) and the windows version from octave-forge. That's why I need more time than I thought!

Michael

Hi all
Actually, I'm writing a lot of tests. I don't want to risk the functionality of the mlp for the next neural network type.
So I think it will go over a lot of time till the next release.

Michael

* tested logsig transfer function
* some updates in docu
* one more docu, more for developers

actually, I'm updating the octave-forge repository, then I will release nnet-0.1.3 with logsig tested

since some days, a new transfer function is implemented, logsig.
But only in the svn version
that's it

In the svn repository, the bug is removed.
Release on the octnnettb page will be done tonight as soon as all my sons are sleeping.
I don't know if this will happen for the octave-forge page

Today, a new bug was reported. I will remove them as soon as possible.
If you don't use 6 input arguments, you'll get in troubles with undefined variables ...

The bug is removed and you can download version 0.1.1 which will work correctly under octave-2.9.10 ...
no other usefull changes ...

Hi all
Yesterday evening I detected that the octave-forge version of nnet doesn't work! I found the error in:
net.inputs{1}.range = [0 0];

where net.inputs is initialized : net.inputs = cell(0,1);
This was working till octave version 2.9.10
I will correct them soon.

Michael

Hi all

I decided to set this repository to the same content like the repository on the octave-forge page...
A lot of files are renamed, some directories are deleted and some others are new created
BUT NO new functionality...
The release nnet-0.1.0 is the same like on the octave-forge page. NO NEW functionality..
but on this repository I can work from my office so I can do much more than with the cvs repository of octave-forge...... read more

Hi all, still nothing new, but if you install octave >=2.9.10 you will get an package manager to install the nnet package from the octave-forge site ... good luck

Hi all
The complete code redesign is finished and so I decided to commit the code to octave-forge. If you checkout the cvs version you will get the neural network toolbox with. It's in main/nnet ...
I don't generate a new release here because there are no main changes, only code optimizing without speed optimizing.

After some weeks, here is something new in the neural network toolbox. I've forgotten the "poststd" function. With this, you can postprocess data which are preprocessed with prestd before. An example will follow in the next some days..
Michel

Thanks to Paul D. Brown how told me that my release is empty on the sourceforge page!!
Now should be everything ok
Michel

I decided to make a new release.
So the actually release 0.0.2-39 can simulate multylayer perceptrons with more than only one hidden layer...
but all the jobs one "news" before must still be done..
Michel

now it should be everything ok if you download the repository, this means all new m-files are included, nothing else
Michel

yes, now it is possible to create neural networks with more than only one hidden layer. But attention, I tested only with 2 hidden layers, rest will be done...
the following jobs will be done before doing more improvements in developing:
- code redesign
- texinfo completion for octave
- documentation
- more tests

this will need a lot of time, so after the code redesign I will launch a new release...... read more