org.simbrain.network.subnetworks

Class SimpleRecurrentNetwork

java.lang.Object

org.simbrain.network.groups.Group

org.simbrain.network.groups.Subnetwork

org.simbrain.network.subnetworks.SimpleRecurrentNetwork

All Implemented Interfaces:

Trainable

public final class SimpleRecurrentNetwork
extends Subnetwork
implements Trainable

Implements a simple recurrent network (See, e.g, Elman 1991). While the SRN behavior could be implemented more efficiently by using a recurrent hidden layer and a buffered update, a context layer is here for educational and visualization purposes, as well as to more closely follow Elman's original design.

Author:

Zach Tosi, Jeff Yoshimi

Constructor Summary

Constructors

Constructor and Description
SimpleRecurrentNetwork(Network network, int numInputNodes, int numHiddenNodes, int numOutputNodes) Build an SRN with default activation rules and initial position.
SimpleRecurrentNetwork(Network network, int numInputNodes, int numHiddenNodes, int numOutputNodes, NeuronUpdateRule hiddenNeuronType, NeuronUpdateRule outputNeuronType, java.awt.geom.Point2D initialPosition) Constructor specifying root network, and number of nodes in each layer.

Method Summary

Modifier and Type	Method and Description
NeuronGroup	getContextLayer()
NeuronGroup	getHiddenLayer()
NeuronGroup	getInputLayer()
java.util.List<Neuron>	getInputNeurons() Returns the list of input neurons.
Group	getNetwork() Returns a reference to the subnetwork or synapse group being trained.
java.util.List<java.util.List<Neuron>>	getNeuronGroupsAsList() Return neuron groups as a list.
NeuronGroup	getOutputLayer()
java.util.List<Neuron>	getOutputNeurons() Returns the list of output neurons.
TrainingSet	getTrainingSet() Returns the the training set, which contains input and target data.
java.lang.String	getUpdateMethodDesecription() Returns a description of this group's update method, which is displayed in the update manager panel.
void	initNetwork() Initialize the network.
void	update() Update this group.

Methods inherited from class org.simbrain.network.groups.Subnetwork

addAndLabelSynapseGroup, addNeuronGroup, addRowToTrainingSet, addSynapseGroup, clearActivations, connectNeuronGroups, connectNeuronGroups, connectNeuronGroups, delete, displayNeuronGroups, getEnabled, getFlatNeuronList, getFlatSynapseList, getIndexOfNeuronGroup, getLongDescription, getModifiableNeuronList, getNeuronGroup, getNeuronGroup, getNeuronGroupByLabel, getNeuronGroupCount, getNeuronGroupList, getSynapseGroup, getSynapseGroup, getSynapseGroupByLabel, getSynapseGroupCount, getSynapseGroupList, isEmpty, recursivelySetIds, removeNeuronGroup, removeSynapseGroup, setDisplayNeuronGroups, setEnabled, size, toString

Methods inherited from class org.simbrain.network.groups.Group

getId, getLabel, getParentGroup, getParentNetwork, getStateInfo, hasParentGroup, isMarkedForDeletion, isTopLevelGroup, setId, setLabel, setMarkedForDeletion, setParentGroup, setStateInfo

Methods inherited from class java.lang.Object

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

Constructor Detail

SimpleRecurrentNetwork

public SimpleRecurrentNetwork(Network network,
                              int numInputNodes,
                              int numHiddenNodes,
                              int numOutputNodes)

Build an SRN with default activation rules and initial position.

Parameters:

network - underlying network

numInputNodes - number of nodes in the input layer

numHiddenNodes - number of nodes in the hidden and context layers

numOutputNodes - number of output nodes

SimpleRecurrentNetwork

public SimpleRecurrentNetwork(Network network,
                              int numInputNodes,
                              int numHiddenNodes,
                              int numOutputNodes,
                              NeuronUpdateRule hiddenNeuronType,
                              NeuronUpdateRule outputNeuronType,
                              java.awt.geom.Point2D initialPosition)

Constructor specifying root network, and number of nodes in each layer.

Parameters:

network - underlying network

numInputNodes - number of nodes in the input layer

numHiddenNodes - number of nodes in the hidden and context layers

numOutputNodes - number of output nodes

hiddenNeuronType - update rule for hidden nodes

outputNeuronType - update rule for hidden nodes

initialPosition - where to position the network (upper left)

Method Detail

initNetwork

public void initNetwork()

Description copied from interface: Trainable

Initialize the network. E.g. used before a training run for SRN's to clear activations.

Specified by:

initNetwork in interface Trainable

update

public void update()

Description copied from class: Group

Update this group.

Overrides:

update in class Subnetwork

getNeuronGroupsAsList

public java.util.List<java.util.List<Neuron>> getNeuronGroupsAsList()

Return neuron groups as a list. Used in backprop trainer.

Overrides:

getNeuronGroupsAsList in class Subnetwork

Returns:

layers list

getInputNeurons

public java.util.List<Neuron> getInputNeurons()

Returns the list of input neurons.

Specified by:

getInputNeurons in interface Trainable

Returns:

the list of input neurons.

getOutputNeurons

public java.util.List<Neuron> getOutputNeurons()

Returns the list of output neurons.

Specified by:

getOutputNeurons in interface Trainable

Returns:

list of output neurons

getTrainingSet

public TrainingSet getTrainingSet()

Returns the the training set, which contains input and target data.

Specified by:

getTrainingSet in interface Trainable

Returns:

the training set.

getContextLayer

public NeuronGroup getContextLayer()

Returns:

the contextLayer

getHiddenLayer

public NeuronGroup getHiddenLayer()

Returns:

the hiddenLayer

getInputLayer

public NeuronGroup getInputLayer()

Returns:

the input

getOutputLayer

public NeuronGroup getOutputLayer()

Returns:

the input

getUpdateMethodDesecription

public java.lang.String getUpdateMethodDesecription()

Returns a description of this group's update method, which is displayed in the update manager panel.

Overrides:

getUpdateMethodDesecription in class Subnetwork

Returns:

a description of the update method.

getNetwork

public Group getNetwork()

Returns a reference to the subnetwork or synapse group being trained.

Specified by:

getNetwork in interface Trainable

Returns:

the network being trained.