Software Release Announcement: SNARLI, free/open-source Java package for neural nets
Simon Levy
levys at wlu.edu
Fri Jul 18 16:03:04 EDT 2003
Dear Connectionists,
I would like to announce the release of a free, open-source Java package
that may be of interest to members of this list. This package is
currently available at http://snarli.sourceforge.net, and is described
below.
Please feel free to download this package, and contact me with question,
criticism, or suggestions. I am especially interested in hearing from
educators and researchers who find the package useful in their work, and
anyone who has a feature or neural architecture that they would like to
see implemented.
Thanks,
Simon
========================
Simon D. Levy
Assistant Professor
Computer Science Department
Washington & Lee University
Lexington, VA 24450
540-458-8419 (voice)
540-458-8479 (fax)
levys at wlu.edu
http://www.cs.wlu.edu/~levy
*SNARLI* (*/S/*imple */N/*eural */AR/*chitecture */LI/*brary) is a Java
package containing two classes: BPLayer, a general back-prop layer
class, and SOM, a class for the Kohonen Self-Organizing Map. BPLayer
also supports sigma-pi connections and back-prop-through-time, allowing
you to build just about any kind of back-prop network found in the
literature.
*SNARLI* differs from existing neural-net packages in two important
ways: First, it is /not/ GUI-based. Instead, it is meant as a code
resource that can be linked directly to new or existing Java-based
projects, for those who want to try a neural-network approach without
having to write a lot of new code. Given the variety of platforms that
currently interface to Java, from HTML to Matlab
<http://www.mathworks.com/>, it made more sense to me to focus on the
neural net algorithms, and leave the GUI development to others.
Second, *SNARLI* gets a great deal of mileage out of a single class
(BPLayer), instead of adding a new class for each type of network. Using
this class, my students and I have been able to construct a large
variety of back-prop networks, from simple perceptrons through Pollack's
RAAM <http://demo.cs.brandeis.edu/papers/long.html#raam>, with very
little additional coding. We have used these networks successfully in
coursework <http://www.cs.wlu.edu/%7Elevy/cs397/>, thesis projects
<http://www.cs.wlu.edu/%7Elevy/grieco_thesis.pdf>, and research
<http://www.cs.wlu.edu/%7Elevy/papers>.
Future versions of *SNARLI* may include classes to support other popular
architectures, such as Support Vector Machines
<http://research.microsoft.com/%7Ejplatt/svm.html> (SVMs), Hopfield Nets
<http://documents.wolfram.com/applications/neuralnetworks/index9.html>,
and Long Short-Term Memory
<http://www.inf.ethz.ch/%7Eschraudo/NNcourse/lstm.html> (LSTM), as user
interest <mailto:levys at wlu.edu> dictates.
More information about the Connectionists
mailing list