Neural Networks

A Tcl extension for neural networks processing using the FANN

library can be found at tcl-fann.

---

(RS would call this A little feed-forward back-propagation learning neural network.)

There came a time when I needed to do a series of simulations of a feed-forward back-propagation (FFBP) neural network

. Rather then spend just an hour doing the homework, I instead spent over 10 hours writing a full-fledged network builder and simulator. As you can see from the screenshot above the result is a little tool that allows the user to graphically draw a network and then run simulations upon it.

Features include:

hidden and bias nodes
choice of the sigmoid (or logistic function) or bipolar squashing function
configurable eta value (learning rate)
learning by way of steepest descent and delta rule

The bipolar function is g(x)=2f(x)-1, where f(x) is the sigmoid function. Conveniently, its derivative is g'(x)=0.5(1+g(x))(1-g(x)). A more general-purpose steepest descent algorithm may be found at Differentiation and steepest-descent, though my FFBP does not use his code.

Usage

Shift-click the canvas to place a node. Shift-drag between nodes to add a weight link. Reposition nodes by dragging them around. Double click a node or weight to change its properties.

The real power of the program is its ability to run a number of trials and to learn after each test datum. To do this requires modifying the code a bit. See the last three functions in ffbp.tcl. The current functions demonstrate how to learn the XOR function; the corresponding network is saved in ffbp.net.

While you are at it, take a look at another kind of neural network, Hopfield Networks.

To do list:

add momentum value
better interface for running trials

Downloads

Version 0.2 - http://tcl.jtang.org/ffbp/ffbp-0.2.tar.gz

This version breaks save file compatibility with previous version!
Remove nodes and weights by double-clicking them, then clicking on 'Remove' button.
Show current weights directly on the canvas.
Adjust weight format using format codes; set string to empty to see raw value.

Version 0.1 - http://tcl.jtang.org/ffbp/ffbp-0.1.tar.gz

Initial hack-ish release.