Introduction to NeuNet Pro
Table Of Contents


What is Neural Net?

Suggested Uses

SFAM Classification

Back Propagation

NeuNet Overview

SFAM Classification

Classification is a type of problem where one attempts to predict the correct class or category for a given pattern where two or more classes are possible. NeuNet Pro will allow up to 256 possible classes in an SFAM project.

Many problems involve a choice between only two possible classes.
For example:

  • {YES or NO}
  • {TRUE or FALSE}
  • {PASS or FAIL}

Other problems can involve dozens or hundreds of classes.
For example:

  • {a certain letter of the alphabet}
  • {a certain numerical digit}
  • {a certain range of price}
  • {a certain state name}
  • {a certain disease name}

The SFAM Algorithm
SFAM is short for "Simplified Fuzzy Adaptive Resonance Theory Map". The algorithm was described in an excellent article by Tom Kasuba in AI Expert Magazine in November, 1993. The following points reflect our thoughts after several years experience using SFAM.

SFAM Strengths:

  • Training is extremely fast.
  • Training requires almost no user intervention.
  • Interpolation between clean data points is excellent.

SFAM Weaknesses:

  • There is a tendency to memorize the training data. This tendency becomes a problem if the training data contains anomalies. Training data should be as clean as possible.
  • Training data must contain no blatant contradictions in the class. These contradictions result in the creation of redundant nodes on every training cycle.
  • These weaknesses can be overcome by early termination of the training process -- A feature of NeuNet Pro.

A Typical SFAM Neural Network

(Used For Predicting Classes)

Training begins with just one hidden node whose weights are set equal to the first record and prediction is set equal to the class of the first record. Similarly, whenever a new class is encountered a new node is created. The node whose weights best match the current input supplies the prediction, provided the degree of match exceeds the vigilance threshold value. If this prediction is correct, the weights of this winning node are adjusted toward this input. If the prediction is wrong or vigilance threshold is not achieved, a new node is created with weights and prediction equal to this record.

A Complete Neural Network Development System

CorMac Technologies Inc.
34 North Cumberland Street ~ Thunder Bay ON P7A 4L3 ~ Canada
E m a i l