org.aika.neuron

Class INeuron

java.lang.Object

org.aika.AbstractNode<Neuron>

org.aika.neuron.INeuron

All Implemented Interfaces:

Comparable<INeuron>, Writable

public class INeuron
extends AbstractNode<Neuron>
implements Comparable<INeuron>

The INeuron class represents a internal neuron implementation in Aikas neural network and is connected to other neurons through input synapses and output synapses. The activation value of a neuron is calculated by computing the weighted sum (input act. value * synapse weight) of the input synapses, adding the bias to it and sending the resulting value through a transfer function (the upper part of tanh).

The neuron does not store its activations by itself. The activation objects are stored within the logic nodes. To access the activations of this neuron simply use the member variable node or use the method getFinalActivations(Document doc) to ge the final activations of this neuron.

Author:

Lukas Molzberger

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	INeuron.NormWeight
static class	INeuron.ThreadState The ThreadState is a thread local data structure containing the activations of a single document for a specific logic node.
static class	INeuron.Type

Field Summary

Fields

Modifier and Type	Field and Description
ActivationFunction	activationFunction
String	activationFunctionKey
static boolean	ALLOW_WEAK_NEGATIVE_WEIGHTS
double	bias
double	biasDelta
double	biasSum
double	biasSumDelta
TreeMap<Synapse,Synapse>	inputSynapses
String	label
ReadWriteLock	lock
static int	MAX_SELF_REFERENCING_DEPTH
double	maxRecurrentSum
double	metaBias
double	negDirSum
double	negRecSum
Provider<OrNode>	node
TreeMap<Synapse.Key,Provider<InputNode>>	outputNodes
TreeMap<Synapse,Synapse>	outputSynapses
String	outputText
double	posDirSum
double	posRecSum
Writable	statistic
INeuron.ThreadState[]	threads
static double	TOLERANCE
INeuron.Type	type
static double	WEIGHT_TOLERANCE

Fields inherited from class org.aika.AbstractNode

lastUsedDocumentId, provider

Constructor Summary

Constructors

Constructor and Description
INeuron(Model m)
INeuron(Model m, String label)
INeuron(Model m, String label, String outputText)

Method Summary

Modifier and Type	Method and Description
Activation	addInput(Document doc, Activation.Builder input) Propagate an input activation into the network.
void	clearActivations()
void	clearActivations(Document doc)
void	clearActivations(int threadId)
int	compareTo(INeuron n)
void	computeBounds(Activation act)
Activation.State	computeWeight(int round, Activation act)
Collection<Activation>	getActivations(Document doc)
Collection<Activation>	getAllActivations(Document doc)
Collection<Activation>	getFinalActivations(Document doc) getFinalActivations is a convenience method to retrieve all activations of the given neuron that are part of the final interpretation.
Activation	getFirstActivation(Document doc)
INeuron.ThreadState	getThreadState(int threadId, boolean create)
void	linkNeuronRelations(Document doc, Activation act) Sets the incoming and outgoing links between neuron activations.
void	propagateAddedActivation(Document doc, Activation act)
void	reactivate()
void	readFields(DataInput in, Model m) Deserialize the fields of this object from in.
static INeuron	readNeuron(DataInput in, Neuron p)
void	remove()
void	suspend()
String	toString()
String	toStringWithSynapses()
static boolean	update(Model m, int threadId, Neuron pn, double biasDelta, Collection<Synapse> modifiedSynapses)
void	write(DataOutput out) Serialize the fields of this object to out.

Methods inherited from class org.aika.AbstractNode

read, setModified

Methods inherited from class java.lang.Object

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

Field Detail

ALLOW_WEAK_NEGATIVE_WEIGHTS

public static boolean ALLOW_WEAK_NEGATIVE_WEIGHTS

WEIGHT_TOLERANCE

public static double WEIGHT_TOLERANCE

TOLERANCE

public static double TOLERANCE

MAX_SELF_REFERENCING_DEPTH

public static int MAX_SELF_REFERENCING_DEPTH

label

public String label

type

public INeuron.Type type

outputText

public String outputText

bias

public volatile double bias

biasDelta

public volatile double biasDelta

biasSum

public volatile double biasSum

biasSumDelta

public volatile double biasSumDelta

metaBias

public volatile double metaBias

posDirSum

public volatile double posDirSum

negDirSum

public volatile double negDirSum

negRecSum

public volatile double negRecSum

posRecSum

public volatile double posRecSum

maxRecurrentSum

public volatile double maxRecurrentSum

statistic

public Writable statistic

activationFunction

public ActivationFunction activationFunction

activationFunctionKey

public String activationFunctionKey

inputSynapses

public TreeMap<Synapse,Synapse> inputSynapses

outputSynapses

public TreeMap<Synapse,Synapse> outputSynapses

outputNodes

public TreeMap<Synapse.Key,Provider<InputNode>> outputNodes

node

public Provider<OrNode> node

lock

public ReadWriteLock lock

threads

public INeuron.ThreadState[] threads

Constructor Detail

INeuron

public INeuron(Model m)

INeuron

public INeuron(Model m,
               String label)

INeuron

public INeuron(Model m,
               String label,
               String outputText)

Method Detail

getThreadState

public INeuron.ThreadState getThreadState(int threadId,
                                          boolean create)

addInput

public Activation addInput(Document doc,
                           Activation.Builder input)

Propagate an input activation into the network.

Parameters:

doc - The current document

input -

remove

public void remove()

propagateAddedActivation

public void propagateAddedActivation(Document doc,
                                     Activation act)

computeBounds

public void computeBounds(Activation act)

computeWeight

public Activation.State computeWeight(int round,
                                      Activation act)

linkNeuronRelations

public void linkNeuronRelations(Document doc,
                                Activation act)

Sets the incoming and outgoing links between neuron activations.

Parameters:

doc -

act -

getActivations

public Collection<Activation> getActivations(Document doc)

getFirstActivation

public Activation getFirstActivation(Document doc)

clearActivations

public void clearActivations()

clearActivations

public void clearActivations(Document doc)

clearActivations

public void clearActivations(int threadId)

compareTo

public int compareTo(INeuron n)

Specified by:

compareTo in interface Comparable<INeuron>

write

public void write(DataOutput out)
           throws IOException

Description copied from interface: Writable

Serialize the fields of this object to out.

Specified by:

write in interface Writable

Parameters:

out - DataOuput to serialize this object into.

Throws:

IOException

readFields

public void readFields(DataInput in,
                       Model m)
                throws IOException

Description copied from interface: Writable

Deserialize the fields of this object from in.

For efficiency, implementations should attempt to re-use storage in the existing object where possible.

Specified by:

readFields in interface Writable

Parameters:

in - DataInput to deseriablize this object from.

Throws:

IOException

suspend

public void suspend()

Overrides:

suspend in class AbstractNode<Neuron>

reactivate

public void reactivate()

Overrides:

reactivate in class AbstractNode<Neuron>

update

public static boolean update(Model m,
                             int threadId,
                             Neuron pn,
                             double biasDelta,
                             Collection<Synapse> modifiedSynapses)

readNeuron

public static INeuron readNeuron(DataInput in,
                                 Neuron p)
                          throws IOException

Throws:

IOException

toString

public String toString()

Overrides:

toString in class Object

toStringWithSynapses

public String toStringWithSynapses()

getFinalActivations

public Collection<Activation> getFinalActivations(Document doc)

getFinalActivations is a convenience method to retrieve all activations of the given neuron that are part of the final interpretation. Before calling this method, the doc.process() needs to be called first.

Parameters:

doc - The current document

Returns:

A collection with all final activations of this neuron.

getAllActivations

public Collection<Activation> getAllActivations(Document doc)