ftp.nice.ch/pub/next/developer/objc/ai/NeuralNetwork.README

This is the README for NeuralNetwork.N.bs.tar.gz [Download] [Browse] [Up]


	Neural Network Simulation Classes
	for the NeXT Computer
		By: Ralph Zazula
		University of Arizona
		zazula@pri.com (NeXT Mail)		

July, 29 1992
		
	This package contains Classes that can be used to create software simulations of neural network architectures.  These were written to facilitate rapid creation of code to explore various network topologies and learning algorithms.   The Neuron class implements a number of different types of "activation" functions including: Sigmoidal, Sgn, Tanh and Binary and also allows the use a pseudo-temperature to create a stochastic network for use in things such as simulated-annealing.  The Neuron class uses a linked-list structure to store and retrieve weights and connections amongst the Neurons resulting in very fast performance.  Also included in the distribution is the Neuron-HT class.  This variation of the Neuron class uses a HashTable structure to store the connections and weights.  The HashTable data-structure provides constant-time retrieval of the weights and connections but also introduces some overhead.  Experimental results showed that the HashTable exhibited faster performance for very large networks (>256 fully-connected Neurons).

The classes contained are:

  Neuron		Implements a number of different types of neurons
  Neuron-HT	Same as above with HashTable data-structure
  BackPropEngine	Implements a back-propagation network

The Example files are:

bptest	An example of using the BackPropEngine (BPE) to perform the identity function

boltz	An example of using stochastic Neurons to implement a Boltzman machine.  This example implements the XOR function

hoptest	An example of a Hopfield associative-memory network

weighted_matching	An neural-network solution to the weighted-matching problem

Other documentation can be found in the Class descriptions, code or example programs.   If you have any questions/comments, feel free to write or e-mail me.

An excellent reference on Neural-Networks is: Introduction to the Theory of Neural Computation by Hertz, Krogh and Palmer, 1991 - Addison Wesley, Publisher (ISBN 0-201-50395-6)

Ralph Zazula
zazula@pri.com (NeXT Mail)
(602)577-1090
4901 E. Sunrise Dr. Apt. 406
Tucson, AZ 85718

These are the contents of the former NiCE NeXT User Group NeXTSTEP/OpenStep software archive, currently hosted by Netfuture.ch.