ANNOUNCE: Version 0.03 of Neural module for Perl5 
Author Message
 ANNOUNCE: Version 0.03 of Neural module for Perl5

(sent to, comp.lang.perl.announce)

This purpose of this message is to announce the availability of
Gann version 0.03, a copylefted artificial neural network simulator.

This software is in the alpha stage of development so don't
expect everything to be finished.

The purpose of announcing Gann at this early stage is to solicit
comments on the programming and user interfaces to the simulator.
Currently, Gann only contains routines to do back-propagation with
gradient descent or momentum descent. However, the interfaces are
very generic and adding new algorithms is very easy.

Gann is copylefted, see the file COPYING in the distribution for details.

What's new?

  - Took out the most presentable parts of 0.02 and put them up for FTP
v0.02: (never released)
  - Cleaned up the network internals and interface to perl
    so that adding new group types shouldn't require recompilation (!!!)
  - Added a shorthand definition for group types so that
    only short snippets of code and hints about the group
    have to be provided.
  - Added new, 'tiled' group type for doing convolutions
  - Somewhat better documentation
  -  with all this, introduced a slew of new bugs.
  - Split the oiginal huge package to smaller subpackages
     with clearly defined interfaces.
  - Name is now Neural::
  - Graphical user interface, requires Tk-b10 (available from CPAN)
    see http://www.*-*-*.com/ ~lukka/gann.html for screen shots.
  - Nomenclature: Changed all package names to be under Math::Neural.
  - Convergence detection in minimization.
  - Temp{*filter*}difference algorithm (TD(lambda))
  - Bug fixes
  - Working backprop

Gann is implemented as a Perl module using C++ for the speed-critical
parts and Perl for everything else, for maximum flexibility.
You need perl version 5.002 or higher (perl 5.001 is rumoured to work but..).

The package contains an example program demonstrating the learning
of the 'xor' function.

If you are interested in seeing the package developed in a certain
direction, please send me email. I'm especially interested in comments
about the following issues
 - What is good and what is bad about the current interfaces in Gann
 - What other net types than backprop should I include?
   (for example, Kohonen etc. Which types of nets are currently
   'hot' and which are not so interesting)
 - What minimization algorithms should I include, what algorithms
   do you have good and/or bad experiences with?

If possible, please include references to publications, or code
on the net.

I'd like to make Gann a general grab-bag of neural network algorithms
containing well-documented code and examples for any algorithms
one might want to use, implemented in an object-oriented fashion to encourage
reuse and interesting multi-network experiments.

The next revision will probably happen in a few weeks, after
I've had time to consider the shape of the interface based on the
comments obtained about this version.

In the future, I intend to keep the release rate high in order to make
new code and network types accessible as fast as possible.


Gann is available from
or any other CPAN site near you.  see
for information about mirrors near you.

Tue, 20 Oct 1998 03:00:00 GMT  
 [ 1 post ] 

 Relevant Pages 

1. ANNOUNCE: (DBD::)XBase version 0.03

2. ANNOUNCE: Math::Polynomial Version 0.03

3. ANNOUNCE: Gann-0.00 neural network simulator perl module


5. ANNOUNCE: WebFS::FileCopy 0.03

6. ANNOUNCE:: DateTime::Precise 0.03

7. ANNOUNCE: RadiusPerl 0.03, alpha release

8. ANNOUNCE: Font::TFM, TeX::DVI, TeX::Hyphen, all 0.03

9. ANNOUNCE: Net::DNS 0.03 available

10. ANNOUNCE: Net::DNS 0.03 available (repost)

11. Announce: String::Multibyte 0.03 is available

12. ANNOUNCE: Bio::PSU 0.03


Powered by phpBB® Forum Software