//Constructor
/// <summary>
/// Creates an initialized instance
/// </summary>
/// <param name="inputFieldIdx">Index of correspondent reservoir input field.</param>
/// <param name="inputRange">
/// Range of input value.
/// It is very recommended to have input values normalized and standardized before
/// they are passed as an input.
/// </param>
public InputAnalogNeuron(int inputFieldIdx, Interval inputRange)
{
Placement = new NeuronPlacement(inputFieldIdx, -1, inputFieldIdx, inputFieldIdx, 0, 0);
_inputRange = inputRange.DeepClone();
StimuliStat = new BasicStat();
TransmissionFreqStat = new BasicStat();
Reset(false);
return;
}
//Methods
/// <summary>
/// Computes the Euclidean distance from another neuron within the pool
/// </summary>
/// <param name="party">Another neuron placement within the same pool.</param>
public double ComputeEuclideanDistance(NeuronPlacement party)
{
double sum = 0;
for (int i = 0; i < Coordinates.Length; i++)
{
sum += ((double)(Coordinates[i] - party.Coordinates[i])).Power(2);
}
return(Math.Sqrt(sum));
}
//Constructor
/// <summary>
/// Creates an initialized instance
/// </summary>
/// <param name="inputFieldIdx">Index of corresponding reservoir input field.</param>
/// <param name="inputRange">
/// Range of input value.
/// It is very recommended to have input values normalized and standardized before
/// they are passed to input neuron.
/// </param>
/// <param name="inputCodingFractions">Number of coding fractions (see SpikeTrainConverter to understand)</param>
public InputSpikingNeuron(int inputFieldIdx, Interval inputRange, int inputCodingFractions)
{
Placement = new NeuronPlacement(inputFieldIdx, -1, inputFieldIdx, inputFieldIdx, 0, 0);
_inputRange = new Interval(inputRange.Min.Bound(), inputRange.Max.Bound());
_signalConverter = new SignalConverter(_inputRange, inputCodingFractions);
StimuliStat = new BasicStat();
TransmissionSignalStat = new BasicStat();
Reset(false);
return;
}
//Constructor
/// <summary>
/// Creates an initialized instance
/// </summary>
/// <param name="placement">Home pool identificator and neuron placement within the pool.</param>
/// <param name="transmissionSignalType">Type of the neuron's signal (Excitatory/Inhibitory).</param>
/// <param name="activation">Instantiated activation function.</param>
/// <param name="bias">Constant bias.</param>
public ReservoirSpikingNeuron(NeuronPlacement placement,
CommonEnums.NeuronSignalType transmissionSignalType,
IActivationFunction activation,
double bias
)
{
_firingRate = new FiringRate();
Placement = placement;
TransmissionSignalType = transmissionSignalType;
Bias = bias;
//Check whether function is spiking
if (activation.OutputSignalType != ActivationFactory.FunctionOutputSignalType.Spike)
{
throw new ArgumentException("Activation function is not spiking.", "activation");
}
_activation = activation;
StimuliStat = new BasicStat();
StatesStat = new BasicStat();
TransmissionSignalStat = new BasicStat();
TransmissionFreqStat = new BasicStat();
Reset(false);
return;
}
//Constructor
/// <summary>
/// Creates an initialized instance
/// </summary>
/// <param name="placement">Home pool identificator and neuron placement within the pool.</param>
/// <param name="transmissionSignalType">Type of the neuron's signal (Excitatory/Inhibitory).</param>
/// <param name="activation">Instantiated activation function.</param>
/// <param name="bias">Constant bias.</param>
/// <param name="retainmentRatio">Retainment ratio.</param>
public ReservoirAnalogNeuron(NeuronPlacement placement,
CommonEnums.NeuronSignalType transmissionSignalType,
IActivationFunction activation,
double bias,
double retainmentRatio
)
{
Placement = placement;
TransmissionSignalType = transmissionSignalType;
Bias = bias;
//Check whether function is analog
if (activation.OutputSignalType != ActivationFactory.FunctionOutputSignalType.Analog)
{
throw new ArgumentException("Activation function is not analog.", "activation");
}
_activation = activation;
_retainmentRatio = retainmentRatio;
StimuliStat = new BasicStat();
StatesStat = new BasicStat();
TransmissionSignalStat = new BasicStat();
TransmissionFreqStat = new BasicStat();
Reset(false);
return;
}
/// <summary>
/// Instantiates the reservoir
/// </summary>
/// <param name="instanceName">The name of the reservoir instance</param>
/// <param name="numOfInputNodes">Number of reservoir inputs</param>
/// <param name="inputRange">Range of input values</param>
/// <param name="settings">Reservoir settings</param>
/// <param name="augmentedStates">Specifies whether this reservoir will add augmented states to output predictors</param>
/// <param name="randomizerSeek">
/// A value greater than or equal to 0 will always ensure the same initialization of the internal
/// random number generator and therefore the same reservoir structure, which is good for tuning purposes.
/// A value less than 0 causes a fully random initialization each time creating a reservoir instance.
/// </param>
public Reservoir(string instanceName, int numOfInputNodes, Interval inputRange, ReservoirSettings settings, bool augmentedStates, int randomizerSeek = -1)
{
//Set instance name
_instanceName = instanceName;
//Copy settings
_settings = settings.DeepClone();
//Random generator initialization
if (randomizerSeek < 0)
{
_rand = new Random();
}
else
{
_rand = new Random(randomizerSeek);
}
//Prepare neuron buffers
//Input neurons
_inputNeurons = new INeuron[numOfInputNodes];
for (int i = 0; i < numOfInputNodes; i++)
{
if (_settings.InputCoding == CommonEnums.InputCodingType.Analog)
{
//Analog input
_inputNeurons[i] = new InputAnalogNeuron(i, inputRange);
}
else
{
//Spiking input
_inputNeurons[i] = new InputSpikingNeuron(i, inputRange, _settings.InputDuration);
}
}
//Pools
_numOfPredictors = 0;
List <INeuron> allNeurons = new List <INeuron>();
int neuronGlobalFlatIdx = 0;
int totalNumOfNeurons = 0;
_poolNeuronsCollection = new List <INeuron[]>(_settings.PoolSettingsCollection.Count);
for (int poolID = 0; poolID < _settings.PoolSettingsCollection.Count; poolID++)
{
PoolSettings poolSettings = _settings.PoolSettingsCollection[poolID];
totalNumOfNeurons += poolSettings.Dim.Size;
_numOfPredictors += poolSettings.RouteToReadout ? poolSettings.Dim.Size : 0;
INeuron[] poolNeurons = new INeuron[poolSettings.Dim.Size];
//Retainment rates
double[] retRates = new double[poolSettings.Dim.Size];
retRates.Populate(0);
if (poolSettings.RetainmentNeuronsFeature)
{
int[] indices = new int[poolSettings.Dim.Size];
indices.ShuffledIndices(_rand);
int numOfRetNeurons = (int)Math.Round(poolSettings.RetainmentNeuronsDensity * poolSettings.Dim.Size, 0);
for (int i = 0; i < numOfRetNeurons; i++)
{
retRates[indices[i]] = _rand.NextDouble(poolSettings.RetainmentMinRate, poolSettings.RetainmentMaxRate, false, RandomClassExtensions.DistributionType.Uniform);
}
}
//Neuron signal types distribution
CommonEnums.NeuronSignalType[] sigTypes = new CommonEnums.NeuronSignalType[poolSettings.Dim.Size];
sigTypes.Populate(CommonEnums.NeuronSignalType.Excitatory);
int[] inhibitoryIndices = new int[poolSettings.Dim.Size];
inhibitoryIndices.ShuffledIndices(_rand);
int numOfInhibitoryNeurons = (int)Math.Round(poolSettings.InhibitoryNeuronsDensity * poolSettings.Dim.Size, 0);
for (int i = 0; i < numOfInhibitoryNeurons; i++)
{
sigTypes[inhibitoryIndices[i]] = CommonEnums.NeuronSignalType.Inhibitory;
}
//Instantiate neurons
int neuronPoolIdx = 0;
for (int x = 0; x < poolSettings.Dim.X; x++)
{
for (int y = 0; y < poolSettings.Dim.Y; y++)
{
for (int z = 0; z < poolSettings.Dim.Z; z++)
{
NeuronPlacement placement = new NeuronPlacement(neuronGlobalFlatIdx, poolID, neuronPoolIdx, x, y, z);
IActivationFunction activation = null;
double bias = 0;
if (sigTypes[neuronPoolIdx] == CommonEnums.NeuronSignalType.Excitatory)
{
//Activation and bias for Excitatory neuron
activation = ActivationFactory.Create(poolSettings.ExcitatoryActivation);
bias = _rand.NextDouble(poolSettings.ExcitatoryBias.Min, poolSettings.ExcitatoryBias.Max, poolSettings.ExcitatoryBias.RandomSign, poolSettings.ExcitatoryBias.DistrType);
}
else
{
//Activation and bias for Inhibitory neuron
activation = ActivationFactory.Create(poolSettings.InhibitoryActivation);
bias = _rand.NextDouble(poolSettings.InhibitoryBias.Min, poolSettings.InhibitoryBias.Max, poolSettings.InhibitoryBias.RandomSign, poolSettings.InhibitoryBias.DistrType);
}
//Neuron instance
if (activation.OutputSignalType == ActivationFactory.FunctionOutputSignalType.Spike)
{
//Spiking neuron
poolNeurons[neuronPoolIdx] = new ReservoirSpikingNeuron(placement,
sigTypes[neuronPoolIdx],
activation,
bias
);
}
else
{
//Analog neuron
poolNeurons[neuronPoolIdx] = new ReservoirAnalogNeuron(placement,
sigTypes[neuronPoolIdx],
activation,
bias,
retRates[neuronPoolIdx]
);
}
allNeurons.Add(poolNeurons[neuronPoolIdx]);
++neuronPoolIdx;
++neuronGlobalFlatIdx;
}
}
}
_poolNeuronsCollection.Add(poolNeurons);
}
//All neurons flat structure
_neurons = allNeurons.ToArray();
//Interconnections
//Banks allocations
_neuronInputConnectionsCollection = new List <ISynapse> [totalNumOfNeurons];
_neuronNeuronConnectionsCollection = new List <ISynapse> [totalNumOfNeurons];
for (int n = 0; n < totalNumOfNeurons; n++)
{
_neuronInputConnectionsCollection[n] = new List <ISynapse>();
_neuronNeuronConnectionsCollection[n] = new List <ISynapse>();
}
//Wiring setup
//Pools internal connections
for (int poolID = 0; poolID < _poolNeuronsCollection.Count; poolID++)
{
//Input connection
SetPoolInputConnections(poolID, _settings.PoolSettingsCollection[poolID]);
//Pool interconnection
if (_settings.PoolSettingsCollection[poolID].InterconnectionAvgDistance > 0)
{
//Required to keep average distance
SetPoolDistInterconnections(poolID, _settings.PoolSettingsCollection[poolID]);
}
else
{
//Not required to keep average distance
SetPoolRandInterconnections(poolID, _settings.PoolSettingsCollection[poolID]);
}
}
//Add pool to pool connections
foreach (ReservoirSettings.PoolsInterconnection poolsInterConn in _settings.PoolsInterconnectionCollection)
{
SetPool2PoolInterconnections(poolsInterConn);
}
//Spectral radius
if (_settings.SpectralRadius > 0)
{
double maxEigenValue = ComputeMaxEigenValue();
if (maxEigenValue == 0)
{
throw new Exception("Invalid reservoir weights. Max eigenvalue is 0.");
}
double scale = _settings.SpectralRadius / maxEigenValue;
//Scale internal weights
foreach (List <ISynapse> connCollection in _neuronNeuronConnectionsCollection)
{
foreach (ISynapse conn in connCollection)
{
conn.Weight *= scale;
}
}
}
//Augmented states
_augmentedStatesFeature = augmentedStates;
return;
}