THETA_NEURON_NETWORK constructs a theta neuron network object
Description:
This function constructs theta neuron network object. Inputs,
all of which are optional, follow a Name/Value format.
Any invalid inputs cause -1 to be returned.
Syntax:
ThNN = THETA_NEURON_NETWORK('PropertyName',PropertyValue,...);
Input Parameters:
o SimulationMethod: Possible values are 'Event-Driven' or 'Numerical'.
'Event-Driven' is the default value.
o NIPS2007Gradient: Boolean flag to indicate that the gradient should be
calculated by the approximation method in "Temporal Coding using the
Response Properties of Spiking Neurons" by T. Voegtlin. The default
value is zero.
o TimeStep: The time step to use is the simulation method is numerical
integration. The default value is 0.02.
o ReferenceTime: Assigning a number to the reference time and layer array
is used for the structure format, then a neuron with index 1 is
automatically connected to all neurons in the network and made to fire
at time equals reference time. Reference time may be an array. The
default value is 1. If reference time is assigned to a negative
number, then no reference neuron is assumed to exist.
o InitialWeightMethod: A cell array indicating the method by which initial
weights are determined. Possible values are:
{'Fixed', val} - val can either be a scalar or matrix
{'RandomNormal', mean, variance}
{'RandomUniform', minimum, maximum}
The default value is {'RandomNormal', 0.01, 0.001}.
o InitialDelayMethod: A cell array indentical to the initial weight
method, but which initializes the weights. The default is {'Fixed',
0}.
o StructureFormat: A cell array that determines the structure of the
network. Possible values are:
{'LayerArray', [NumNeuronsLayer1 NumNeuronsLayer2 ...]}
{'ConnectionMatrix', ConnectionMatrix}
{'GivenStructure', Weights, Delays}
For the connection matrix, a 1 at index (i,j) indicates that neuron i
has a directed connection to neuron j. For given structure, the weight
and delay initialization methods are overwritten and instead the
weights and delays are directly assigned. The default value is
{'LayerArray' [1 1]}.
o Alpha: Scalar scaling constant for all neurons. Alpha may be a scalar or
an array. An array of size NumNeurons, enables different values to be
assigned to each neuron. The default value is 1.
o Io: Scalar baseline current for all neurons. Io may be a scalar or an
array. An array of size NumNeurons, enables different values to be
assigned to each neuron. The default value is -0.005.
o Phases: A cell array in the case where are neurons are assigned the same
initial phase, or a cell array of cells in the case where each neuron
is assigned a different initial phase. Each member of the low-level
array is a cell of size 1x2 or 1x1 that indicates the phase
initialization method. Possible Values are:
{'PositiveFixedPoint'}
{'NegativeFixedPoint'}
{'PositiveFixedPointDelta' Delta}
{'NegativeFixedPointDelta' Delta}
{'GivenValue' Value}
When Io>0, the default is a GivenValue of 0. When Io<0, the default
is a PositiveFixedPointDelta of 1e-4.
Output Parameter:
o ThNN: A theta neuron network object. This object can be run or trained.
Class Functions:
Type "methods('theta_neuron_network')"
Examples:
>> ThNN1=theta_neuron_network;
>> ThNN2=theta_neuron_network('ReferenceTime',-1, ...
'StructureFormat', {'LayerArray', [1 1 1]}, 'Phases', ...
{{'PositiveFixedPoint'},{'NegativeFixedPoint'},{'GivenValue' 1}}, ...
'Alpha',[1 2 3]);
See also
theta_neuron
theta_neuron_network/apply_gradient
theta_neuron_network/calculate_gradient
theta_neuron_network/code_data
theta_neuron_network/display
theta_neuron_network/get_error
theta_neuron_network/get_input_neurons
theta_neuron_network/get_output_neurons
theta_neuron_network/numerical_gradient
theta_neuron_network/relate_q_to_k
theta_neuron_network/run_network
theta_neuron_network/run_networked_neuron
theta_neuron_network/subsasgn
theta_neuron_network/subsref
theta_neuron_network/train
theta_neuron_network/train_dataset
theta_neuron_network/train_epoch0001 function ThNN = theta_neuron_network(varargin) 0002 %THETA_NEURON_NETWORK constructs a theta neuron network object 0003 % 0004 %Description: 0005 %This function constructs theta neuron network object. Inputs, 0006 %all of which are optional, follow a Name/Value format. 0007 %Any invalid inputs cause -1 to be returned. 0008 % 0009 %Syntax: 0010 %ThNN = THETA_NEURON_NETWORK('PropertyName',PropertyValue,...); 0011 % 0012 %Input Parameters: 0013 %o SimulationMethod: Possible values are 'Event-Driven' or 'Numerical'. 0014 % 'Event-Driven' is the default value. 0015 %o NIPS2007Gradient: Boolean flag to indicate that the gradient should be 0016 % calculated by the approximation method in "Temporal Coding using the 0017 % Response Properties of Spiking Neurons" by T. Voegtlin. The default 0018 % value is zero. 0019 %o TimeStep: The time step to use is the simulation method is numerical 0020 % integration. The default value is 0.02. 0021 %o ReferenceTime: Assigning a number to the reference time and layer array 0022 % is used for the structure format, then a neuron with index 1 is 0023 % automatically connected to all neurons in the network and made to fire 0024 % at time equals reference time. Reference time may be an array. The 0025 % default value is 1. If reference time is assigned to a negative 0026 % number, then no reference neuron is assumed to exist. 0027 %o InitialWeightMethod: A cell array indicating the method by which initial 0028 % weights are determined. Possible values are: 0029 % {'Fixed', val} - val can either be a scalar or matrix 0030 % {'RandomNormal', mean, variance} 0031 % {'RandomUniform', minimum, maximum} 0032 % The default value is {'RandomNormal', 0.01, 0.001}. 0033 %o InitialDelayMethod: A cell array indentical to the initial weight 0034 % method, but which initializes the weights. The default is {'Fixed', 0035 % 0}. 0036 %o StructureFormat: A cell array that determines the structure of the 0037 % network. Possible values are: 0038 % {'LayerArray', [NumNeuronsLayer1 NumNeuronsLayer2 ...]} 0039 % {'ConnectionMatrix', ConnectionMatrix} 0040 % {'GivenStructure', Weights, Delays} 0041 % For the connection matrix, a 1 at index (i,j) indicates that neuron i 0042 % has a directed connection to neuron j. For given structure, the weight 0043 % and delay initialization methods are overwritten and instead the 0044 % weights and delays are directly assigned. The default value is 0045 % {'LayerArray' [1 1]}. 0046 %o Alpha: Scalar scaling constant for all neurons. Alpha may be a scalar or 0047 % an array. An array of size NumNeurons, enables different values to be 0048 % assigned to each neuron. The default value is 1. 0049 %o Io: Scalar baseline current for all neurons. Io may be a scalar or an 0050 % array. An array of size NumNeurons, enables different values to be 0051 % assigned to each neuron. The default value is -0.005. 0052 %o Phases: A cell array in the case where are neurons are assigned the same 0053 % initial phase, or a cell array of cells in the case where each neuron 0054 % is assigned a different initial phase. Each member of the low-level 0055 % array is a cell of size 1x2 or 1x1 that indicates the phase 0056 % initialization method. Possible Values are: 0057 % {'PositiveFixedPoint'} 0058 % {'NegativeFixedPoint'} 0059 % {'PositiveFixedPointDelta' Delta} 0060 % {'NegativeFixedPointDelta' Delta} 0061 % {'GivenValue' Value} 0062 % When Io>0, the default is a GivenValue of 0. When Io<0, the default 0063 % is a PositiveFixedPointDelta of 1e-4. 0064 % 0065 %Output Parameter: 0066 %o ThNN: A theta neuron network object. This object can be run or trained. 0067 % 0068 %Class Functions: 0069 %Type "methods('theta_neuron_network')" 0070 % 0071 %Examples: 0072 %>> ThNN1=theta_neuron_network; 0073 % 0074 %>> ThNN2=theta_neuron_network('ReferenceTime',-1, ... 0075 % 'StructureFormat', {'LayerArray', [1 1 1]}, 'Phases', ... 0076 % {{'PositiveFixedPoint'},{'NegativeFixedPoint'},{'GivenValue' 1}}, ... 0077 % 'Alpha',[1 2 3]); 0078 % 0079 %See also 0080 % theta_neuron 0081 % theta_neuron_network/apply_gradient 0082 % theta_neuron_network/calculate_gradient 0083 % theta_neuron_network/code_data 0084 % theta_neuron_network/display 0085 % theta_neuron_network/get_error 0086 % theta_neuron_network/get_input_neurons 0087 % theta_neuron_network/get_output_neurons 0088 % theta_neuron_network/numerical_gradient 0089 % theta_neuron_network/relate_q_to_k 0090 % theta_neuron_network/run_network 0091 % theta_neuron_network/run_networked_neuron 0092 % theta_neuron_network/subsasgn 0093 % theta_neuron_network/subsref 0094 % theta_neuron_network/train 0095 % theta_neuron_network/train_dataset 0096 % theta_neuron_network/train_epoch 0097 0098 0099 %Copyright (C) 2008 Sam McKennoch <Samuel.McKennoch@loria.fr> 0100 0101 %Check to see if formatting is correct 0102 if mod(nargin,2)~=0 0103 disp('Error in theta neuron network constructor: There must be an even number of inputs'); 0104 ThNN=-1; 0105 return; 0106 end 0107 0108 %Assign Default Values, Alpha, Io, Phases left unassigned for default values 0109 ThNN.SimulationMethod='Event-Driven'; 0110 ThNN.NIPS2007Gradient=0; 0111 ThNN.TimeStep=0.02; 0112 ThNN.ReferenceTime=1; %Assign this to a negative value if don't want to use a reference time and are using layers 0113 ThNN.InitialWeightMethod = {'RandomNormal', 0.01, 0.001}; 0114 ThNN.InitialDelayMethod = {'Fixed', 0}; 0115 ThNN.StructureFormat = {'LayerArray', [1 1]}; 0116 0117 %Cycle through inputs and assign naively 0118 for j=1:(nargin/2) 0119 switch varargin{2*j-1} 0120 case {'Alpha','Io'} 0121 eval([varargin{2*j-1},'=[',num2str(varargin{2*j}),'];']); 0122 case 'Phases' 0123 if ~iscell(varargin{2*j}{1}) 0124 if length(varargin{2*j})==1 0125 eval([varargin{2*j-1},'={{''',varargin{2*j}{1},'''}};']); 0126 else 0127 eval([varargin{2*j-1},'={{''',varargin{2*j}{1},''',', num2str(varargin{2*j}{2}),'}};']); 0128 end 0129 else 0130 Phases=varargin{2*j}; 0131 end 0132 case fields(ThNN) 0133 ThNN.(varargin{2*j-1})=varargin{2*j}; 0134 otherwise 0135 disp(['Warning: Value ', varargin{2*j-1}, ' is unknown; ignoring...']); 0136 end 0137 end 0138 0139 switch ThNN.StructureFormat{1} 0140 0141 %Given Connections By Stating Numbers of Neurons in Each Layer 0142 %Default values used for parameters plus weight and delay values 0143 %'Structure Format',{'LayerArray',LayerArray,['Alpha',Alpha,'Io',Io,'Phases',Phases]} 0144 case 'LayerArray' 0145 Layers = ThNN.StructureFormat{2}; 0146 if ThNN.ReferenceTime<0 0147 Layers(1)=Layers(1)-1; 0148 end 0149 ConnectionMatrix=make_connection_matrix(Layers); 0150 if ThNN.ReferenceTime<0 && length(Layers)>2 0151 ConnectionMatrix(1,(Layers(1)+Layers(2)+2):(sum(Layers)+1))=0; 0152 end 0153 ThNN.Weights = initialize_connections(ConnectionMatrix,ThNN.InitialWeightMethod); 0154 ThNN.Delays = initialize_connections(ConnectionMatrix,ThNN.InitialDelayMethod); 0155 0156 %Given Connections Through a Connection Matrix 0157 %Default values used for parameters plus weight and delay values 0158 %'Structure Format',{'ConnectionMatrix',ConnectionMatrix} 0159 case 'ConnectionMatrix' 0160 ConnectionMatrix=ThNN.StructureFormat{2}; 0161 ThNN.Weights = initialize_connections(ConnectionMatrix,ThNN.InitialWeightMethod); 0162 ThNN.Delays = initialize_connections(ConnectionMatrix,ThNN.InitialDelayMethod); 0163 0164 %Given Connections Through a Connection Matrix 0165 %Default values used for parameters plus weight and delay values 0166 %'Structure Format',{'ConnectionMatrix',ConnectionMatrix} 0167 case 'ConnectionProbability' 0168 ConnectionProbability=ThNN.StructureFormat{2}; 0169 NumNeurons=ThNN.StructureFormat{3}; 0170 ConnectionMatrix=rand(NumNeurons)<ConnectionProbability; 0171 ThNN.Weights = initialize_connections(ConnectionMatrix,ThNN.InitialWeightMethod); 0172 ThNN.Delays = initialize_connections(ConnectionMatrix,ThNN.InitialDelayMethod); 0173 0174 %Explicitly Given Weights and Delays 0175 %Default values used for parameters 0176 %'StructureFormat',{'GivenStructure',Weights,Delays} 0177 case 'GivenStructure' %Explicitly Given Weights and Delays 0178 ThNN.Weights=ThNN.StructureFormat{2}; 0179 ThNN.Delays=ThNN.StructureFormat{3}; 0180 ConnectionMatrix = ((ThNN.Weights~=0) | (ThNN.Delays~=0)); 0181 0182 otherwise 0183 disp('Error in theta neuron network constructor: StructureFormat not Recognized'); 0184 ThNN=-1; 0185 return; 0186 end 0187 0188 %Placeholder Fields to Be updated by update_structure function (to save on computation later) 0189 ThNN.InputNeurons=0; 0190 ThNN.OutputNeurons=0; 0191 ThNN.RecursionFlag=0; 0192 ThNN.NeuronQueue=[]; 0193 ThNN.RelativeInputNeurons={}; 0194 ThNN.RelativeOutputNeurons={}; 0195 0196 %Use Neuron Defaults For Alpha, Io, InitialPhase unless specified 0197 ThNN.Neurons=theta_neuron('Size',max(size(ConnectionMatrix,1))); 0198 if exist('Alpha')==1 || exist('Io')==1 || exist('Phases')==1 0199 for j=1:max(size(ConnectionMatrix)) 0200 ParamString=''; 0201 if exist('Alpha')==1 0202 if isscalar(Alpha) %Matrix or Scalar? 0203 CurrentAlpha=Alpha; 0204 ParamString=[ParamString, '''Alpha'', CurrentAlpha']; 0205 else 0206 CurrentAlpha=Alpha(j); 0207 ParamString=[ParamString, '''Alpha'', CurrentAlpha']; 0208 end 0209 if exist('Io')==1 || exist('Phases')==1 0210 ParamString=[ParamString, ',']; 0211 end 0212 end 0213 if exist('Io')==1 0214 if isscalar(Io) %Matrix or Scalar? 0215 CurrentIo=Io; 0216 ParamString=[ParamString, '''Io'', CurrentIo']; 0217 else 0218 CurrentIo=Io(j); 0219 ParamString=[ParamString, '''Io'', CurrentIo']; 0220 end 0221 if exist('Phases')==1 0222 ParamString=[ParamString, ',']; 0223 end 0224 end 0225 if exist('Phases')==1 0226 if length(Phases)==1 %Matrix or Scalar? 0227 CurrentPhase=Phases{1}; 0228 ParamString=[ParamString, '''InitialPhaseMethod'', CurrentPhase']; 0229 else 0230 CurrentPhase=Phases{j}; 0231 ParamString=[ParamString, '''InitialPhaseMethod'', CurrentPhase']; 0232 end 0233 end 0234 eval(['ThNN.Neurons(j)=theta_neuron(',ParamString,');']); 0235 end 0236 end 0237 ThNN = class(ThNN,'theta_neuron_network'); 0238 0239 ThNN=update_structure(ThNN); 0240 0241 function ConnectionMatrix=make_connection_matrix(Layers) 0242 0243 ConnectionMatrix=zeros(sum(Layers)+1); %+1 for the reference neuron 0244 ConnectionMatrix(1,((Layers(1)+2):(sum(Layers)+1)))=1; 0245 for m=1:(length(Layers)-1) 0246 % CurrentLayer=[2 Layers(1)+1]; 0247 % NextLayer=[Layers(1)+2 sum(Layers(1:2))+1]; 0248 CurrentLayer=[sum(Layers(1:(m-1)))+2 sum(Layers(1:m))+1]; 0249 NextLayer=[sum(Layers(1:m))+2 sum(Layers(1:(m+1)))+1]; 0250 for j=CurrentLayer(1):CurrentLayer(2) 0251 for k=NextLayer(1):NextLayer(2) 0252 ConnectionMatrix(j,k)=1; 0253 end 0254 end 0255 end 0256 0257 function ValMatrix = initialize_connections(ConnectionMatrix,InitialMethod); 0258 %InitialMethods 0259 %'Fixed' 'val' 0260 %'RandomNormal' 'mean' 'std' %This is what I used before with 0.01 and maybe 0.1? 0261 %'RandomUniform' 'min' 'max' 0262 0263 0264 switch InitialMethod{1} 0265 case 'Fixed' %Can handle scalar or matrix 0266 ValMatrix = ConnectionMatrix.*InitialMethod{2}; 0267 case 'RandomNormal' 0268 ValMatrix = ConnectionMatrix.*((InitialMethod{3}*randn(size(ConnectionMatrix)))+InitialMethod{2}); 0269 case 'RandomUniform' 0270 ValMatrix = ConnectionMatrix.*((rand(size(ConnectionMatrix))*(InitialMethod{3}-InitialMethod{2}))+InitialMethod{2}); 0271 end