Neural Network Designer

[Barbarrosa]

I've made a simple neural network designer. It doesn't have a lot of the more advanced options that Joone or Emergent have, but it can be used to design a neural network with parallel-running neurons controlled in steps. It can receive and even edit input from .CSV files if they are formatted in a particular way. The CSV-reading part of the program uses the LGPL library Java CSV.

I'd like to hear any feedback people have to offer. I'm considering taking this project further, or initiating projects based on this one.

Instructions & information:

To setup the program, download it from here and obtain Java CSV from http://sourceforge.net/projects/javacsv/. Put the "javacsv.jar" file in a folder named "ANNMaker_lib", with "ANNMaker.jar" a folder above that. Then just run "ANNMaker.jar".
Edit: Almost everything still works if you don't set it up. You just can't open CSV files.

The Data Sets the program works with provide information about which neurons will receive what data as input, or what data is expected as output. The input functionality is partially present, but the output-testing and training functionality is not present.

The constructor portion works by placing the objects you click on in the palette, into the canvas when you click there. You can form a connection by clicking on "TopConnection", then on one neuron, then on another neuron, then input a weight. Neurons can restrict how connections relate to them, and refuse or accept connections based on a condition. Double-clicking on a neuron activates whatever that neuron is programmed to do; input neurons ask for an input file, and output neurons display an actively updating list of their outputs. You can run, pause, and save the network from the constructor (saving actually serializes the window).

Any subclass of "Neuron" can be added to the program by adding it to the package where the types of neurons are stored, adding it to the list of available neurons to place, then compiling and running the program. Currently the sigmoid neurons have only a default bias value; they can be modified to make their bias values changeable from the GUI.