boone.structure

Class NetFactory

java.lang.Object

boone.structure.NetFactory

public class NetFactory
extends java.lang.Object

A factory producing various neural network types. Note that you may adapt the generated networks to your needs, or produce your own networks by constructing a NeuralNet and adding Neuron and Link objects.

Version:

$Id$

Author:

August Mayer, Helmut A. Mayer

Constructor Summary

Constructors

Constructor and Description
NetFactory()

Method Summary

Modifier and Type	Method and Description
static SpikingNeuralNet	createFeedForward(int[] neuronsPerLayer, boolean full, SpikingNeuron neuron, SpikingLink link) Create a feed-forward spiking net.
static NeuralNet	createFeedForward(int[] neuronsPerLayer, boolean full, Trainer trainer, Neuron neuron, Link link) Create a standard feed-forward network.
static NeuralNet	createFeedForward(java.util.List<Layer> layers, Trainer trainer, boolean random) Creates a feed-forward network with connections between neighboring layers.
static SpikingNeuralNet	createHopfield(int numNeurons, Function neuronAF, Function linkAF, Trainer trainer, Neuron neuron, Link link) Create a recurrent spiking neural network (e.g.
static NeuralNet	createHopfield(int numNeurons, Function activationFunctionTemplate, Trainer trainerTemplate, Neuron neuronTemplate, Link linkTemplate) Creates a a Hopfield network, where each neuron is connected to every other.
static NeuralNet	createSOM(SOMTrainer trainer, PatternSet patterns) The master method for creating a SOM net.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

NetFactory

public NetFactory()

Method Detail

createFeedForward

public static NeuralNet createFeedForward(int[] neuronsPerLayer,
                                          boolean full,
                                          Trainer trainer,
                                          Neuron neuron,
                                          Link link)

Create a standard feed-forward network. Bias is switched off for the input layer, and switched on for the others. Weights and biases are randomized.

Parameters:

neuronsPerLayer - number of neurons per layer, the first and the last layer are input and output layers, and the layers in between are hidden.

full - if true, the network will be fully connected, i.e. all potential feed-forward connections are made resulting in a GMLP (i.e. for a 3-layer-net: input - hidden, hidden - output, input - output).

trainer - template for the network trainer if null, an RPropTrainer will be created as the network trainer

neuron - template for creating the neurons. The network's neurons will be created with the clone() method. If null, a standard neuron will be created

link - template for creating the links, the network's links will be created with the clone() method, if null, a standard link will be created

Returns:

a feed-forward network with the given parameters

createFeedForward

public static NeuralNet createFeedForward(java.util.List<Layer> layers,
                                          Trainer trainer,
                                          boolean random)

Creates a feed-forward network with connections between neighboring layers. The layers must be set up correctly, e.g., an input layer must have input neurons. If 'random' weights and biases are randomized.

Parameters:

layers - a list of layers starting with the input layer

trainer - template for the network trainer if null, an RPropTrainer will be created as the network trainer

random - true, if weights and biases are to be randomized

Returns:

a feed-forward network with the given parameters

createFeedForward

public static SpikingNeuralNet createFeedForward(int[] neuronsPerLayer,
                                                 boolean full,
                                                 SpikingNeuron neuron,
                                                 SpikingLink link)

Create a feed-forward spiking net. Weights will be initialized with 1. Delays for the links will be initialized with random values in the specified interval.

Parameters:

neuronsPerLayer - Number of neurons per layer. the first and the last layer are input and output layers, and the layers in between are hidden.

full - If true, the network will be fully connected, i.e. all layers will be interconnected in the forward direction. (i.e. for a 3-layer-net: input - hidden, hidden - output, input - output).

neuron - Template for creating the spiking neurons. The network's neurons will be created with the clone() method (if you use a derived spiking neuron class, please remember to make a suitable one!). If null, a SpikingNeuron will be created.

link - Template for creating the spiking links. The network's links will be created with the clone() method (if you use a derived spiking link class, please remember to make a suitable one!). If null, a SpikingLink will be created.

Returns:

Standard feed-forward network with the given number of layers and some default parameters.

createSOM

public static NeuralNet createSOM(SOMTrainer trainer,
                                  PatternSet patterns)

The master method for creating a SOM net.

Parameters:

trainer - a SOM trainer

patterns - the patterns to be mapped

Returns:

a SOM

createHopfield

public static NeuralNet createHopfield(int numNeurons,
                                       Function activationFunctionTemplate,
                                       Trainer trainerTemplate,
                                       Neuron neuronTemplate,
                                       Link linkTemplate)

Creates a a Hopfield network, where each neuron is connected to every other. Sets maxCycles to 100. Every link has the directed flag enabled. Weights will be initialized with values in [-0.1, 0.1).

Parameters:

numNeurons - number of neurons

activationFunctionTemplate - Template for the activation functions of the neurons. If null, then the default (AtLeast 0) will be used

trainerTemplate - Template object for the network trainer. If null, a HopfieldDeltaTrainer will be used.

neuronTemplate - A template for creating the neurons. The network's neurons will be created with the clone() method. If null, a standard neuron will be created.

linkTemplate - A template for creating the links. The network's links will be created with the clone() method. If null, a standard link will be created.

Returns:

a Hopfield network

createHopfield

public static SpikingNeuralNet createHopfield(int numNeurons,
                                              Function neuronAF,
                                              Function linkAF,
                                              Trainer trainer,
                                              Neuron neuron,
                                              Link link)

Create a recurrent spiking neural network (e.g. a Hopfield network). This returns a network where every neuron is connected to every other.

Sets maxCycles to 100.

Every link has the directed flag enabled. Weights will be initialized with 1. Delays for the links will be initialized with random values in the specified interval.

Parameters:

numNeurons - How many neurons should be created.

neuronAF - Template object for the activation functions of the neurons. If null, then the Spike function will be used.

linkAF - Template object for the activation functions of the links. If null, then the PostsynapticPotential function will be used.

trainer - the trainer (there is no Boone trainer for SNN currently)

neuron - Template for creating the spiking neurons. The network's neurons will be created with the clone() method (if you use a derived spiking neuron class, please remember to make a suitable one!). If null, a SpikingNeuron will be created.

link - Template for creating the spiking links. The network's links will be created with the clone() method (if you use a derived spiking link class, please remember to make a suitable one!). If null, a SpikingLink will be created. @return a recurrent network.