%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% This is an implementation of an ARTMAP variant,
% as described in:
% Gaddam, C. S. (2007).
% Feature Selection via transmitter depletion in ARTMAP. Online Document, xx(x) xxx-xxx.
% Boston, MA: Boston University.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Programmed by Chaitanya Sai (August 2007-08)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% The notation follows the article's notation,
% as does the headers for each step of the
% algorithm
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% close all;
if ~exist('dataStruct_biasedART','var')
load fuzzyARTbenchmarks
load dataset_descriptions
disp('Finished loading datasets')
end
%%%%%%%%DATASETS%%%%%%%%%%%%%
%
% 1'six_point_test'
% 2'Stripes_Sparse'
% 3'Stripes_Dense'
% 4'CIS_Sparse'
% 5'CIS_Dense'
% 6'Checkerboard_Sparse'
% 7'Checkerboard_Dense'
% 8'Binary_five'
% 9'Test on Stripe 3 Capped'
% 10'Test on Stripe 2 Capped'
% 11'Test on Stripe 1 Capped'
% 12'Test on Stripe 0 Capped'
% 13'Genre prediction'
%%%%%%%%DATASETS%%%%%%%%%%%%%
disp({dataset_strings.string_val}')
%dataSetNum=input('Input dataset?');
if (init_vals.datasrc==1)
dataSetNum=4;
elseif (init_vals.datasrc==2)
dataSetNum=5;
elseif (init_vals.datasrc==3)
dataSetNum=2;
elseif (init_vals.datasrc==4)
dataSetNum=3;
elseif (init_vals.datasrc==5)
dataSetNum=6;
elseif (init_vals.datasrc==6)
dataSetNum=7;
elseif (init_vals.datasrc==7)
dataSetNum=9;
elseif (init_vals.datasrc==8)
dataSetNum=10;
elseif (init_vals.datasrc==9)
dataSetNum=11;
elseif (init_vals.datasrc==10)
dataSetNum=12;
elseif (init_vals.datasrc==11)
dataSetNum=13;
end
%lambda_Attention=input('lambda Value (0 -> fuzzy ARTMAP)?');
lambda_Attention=0;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
try
dataStruct_fuzzyART(dataSetNum).test_output=dataStruct_fuzzyART(dataSetNum).test_output(:);
catch
disp('Test Output Not Available')
end
inputNet = struct('M', {[]}, ... % Number of features
'TestOnly', {0}, ... % Train & Test or Test only? unless specified otherwise
'test_fast', {1}, ... % Use fast tester unless specified otherwise
'compute_overlap', {0}, ... % Indirectly measure overlap by looking at how many nodes overlap with a test_point
'C', {0}, ... % Number of committed coding nodes (paper says set to 0, but harder to Matlab)
'plotSteps',{0}, ... % % Iterate through individual learning steps
'NodeList', {[]}, ... % List of node associated with each input
'maxNumCategories', {100}, ...
'numClasses', {[]}, ... % The total number of classes
'w', {[]}, ... % coding node weight vector
'W', {[]}, ... % output class weight vector
'rho_bar', {0}, ... % rho_bar, the baseline vigilance. baseline=0 maximizes code compression
'rho', {0.0}, ... % vigilance
'alpha', {.1}, ... % Signal Rule Parameter
'alphaTrain',{[]},... % Store when reverting alpha value in TESTING
'beta', {1.0}, ... % Learning fraction
'gamma',{.000001},... % fraction additive in numerator and denominator of matching signal
'epsilon', {-.00001}, ... % Match Tracking (codes inconsistent cases)
'base_vigilance', {0.0}, ... % Baseline vigilance
'p', {1.0}, ... % CAM rule power
'dataSubsets', {1}, ... % 4-fold cross-validation
'votes', {5}, ... % Number of voting systems
'learningRate', {1.0}, ...
'win_sequence',{[]},...
'act_sequence',{[]},...
'search_cycles', {0},... % Total number of search cycles during training
'lambda_Attention',{[]},...
'dataStore',{[]},... % Store data if using structs as inputs
'numEpochs',{1},...%Number of times to present input
'model_num',{[]},... % Depletion Model Number
'type_depletion',{4},... %Options are 0 1 2 3 4
'deplete_fast',{[]},...
'learn_trail',{[]},...
'learn_description',{''},...
'short_learn_descript',{''},...
'uncommitted_fail',{0},...
'NI_fail',{[]},...
'NI_fail_addNode',{[]},...
'node_created',{[]},...
'start_conservative',{[]},...
'ovlp_net',{[]},...
'A_dep_dist',{[]},...
'e_mod_dist',{[]},...
'e_realLow',{0},...
'e_zero',{0},...
'diffE_count',{[]},...
'perc_error',{[]},... % percentage error added to labels in training
'var_error_features',{[]},... % variance of error added to features in training
'testAccuracy',{[]}); % Attention gain parameter
% inputNet.lambda_Attention=0.5;
inputNet.lambda_Attention=lambda_Attention;
inputNet.deplete_fast = 0; %TURN OFF (=0) IF DEPLETING USING ITERATIVE VERSION
inputNet.start_conservative =0;
inputNet.DEFAULT_RETEST=1;
%inputNet.alpha=10^(-8);
inputNet.alpha=init_vals.alpha;
%inputNet.beta=1;
inputNet.beta=init_vals.beta;
%inputNet.gamma=0.0000001;
inputNet.gamma=init_vals.gamma;
%inputNet.rho_bar=0;
inputNet.rho_bar=init_vals.rho;
%inputNet.plotSteps=0;
inputNet.plotSteps=init_vals.plot_steps;
inputNet.model_num=32;
%Train for number of epochs
%inputNet.numEpochs=1;
inputNet.numEpochs=init_vals.epochs;
%%%Inject label errors into dataset | e.g perc_error = 10 implies 10% of
%%%training data have erroneous class labels during training
perc_error=0;
inputNet.perc_error=perc_error;
%%%Inject feature noise into dataset | e.g var_error_features = 10 implies 10% of
%%%training data have noise added to input features
var_error_features=0;
inputNet.var_error_features=var_error_features;
inputNet.compute_overlap=1;
if (dataSetNum==13)
%Matrix manipulation becomes too memory intensive for movie genres
%benchmark
inputNuet.test_fast=0;
end
if (inputNet.perc_error~=0)
disp('Adding error in training set!');
elseif (inputNet.var_error_features~=0)
disp('Adding feature error in training set!');
end
%%%%%%%%%%%%%%%%%%%
%Create uniform plot grid%%
%%%%%%%%%%%%%%%%%%%
if ismember(dataSetNum,[1 2 3 4 5 6 7])
if (perc_error>0)
orig_dataset_labels=dataStruct_fuzzyART(dataSetNum).training_output;
dataStruct_fuzzyART(dataSetNum).training_output=addClassError(dataStruct_fuzzyART(dataSetNum).training_output,inputNet.perc_error);
temp_labels_noise=dataStruct_fuzzyART(dataSetNum).training_output;
elseif (var_error_features>0)
orig_dataset_features=dataStruct_fuzzyART(dataSetNum).training_input;
dataStruct_fuzzyART(dataSetNum).training_input=max(min(dataStruct_fuzzyART(dataSetNum).training_input+randn(size(dataStruct_fuzzyART(dataSetNum).training_input))*var_error_features,1),0);
temp_features_noise=dataStruct_fuzzyART(dataSetNum).training_input;
end
if ~exist('dataTestUnif','var')
load ARTMAP_data_test dataTestUnif dataTestOutUnif test_x test_y
end
dataStruct_fuzzyART(dataSetNum).test_input=dataTestUnif;
if dataSetNum>1
dataStruct_fuzzyART(dataSetNum).test_output=dataTestOutUnif{floor(dataSetNum/2)}';
end
end
[a,b,c]=Default_ARTMAP(dataStruct_fuzzyART(dataSetNum),inputNet);
if ismember(dataSetNum,[1 2 3 4 5 6 7])
if (perc_error>0)
inputNet.lambda_Attention=0;
[a_zero,b_zero,c_zero]=Default_ARTMAP(dataStruct_fuzzyART(dataSetNum),inputNet);
disp('Same training order lambda = 0')
disp(sum(a_zero==dataStruct_fuzzyART(dataSetNum).test_output)*100/length(a_zero));
labels_switch_positions=find(orig_dataset_labels~=temp_labels_noise);
disp(intersect(labels_switch_positions,c.NI_fail))
dataStruct_fuzzyART(dataSetNum).training_output=orig_dataset_labels;
elseif (var_error_features>0)
dataStruct_fuzzyART(dataSetNum).training_input=orig_dataset_features;
end
end
if (ismember(inputNet.model_num,[31 2^c.M]) && (inputNet.deplete_fast==1) && (inputNet.lambda_Attention==1))
c.lambda_Attention='\infty';
end
if (c.M==2) && (inputNet.plotSteps==0)
a=a(:);
if dataSetNum~=1
colormap([1 .4 .4; 1 .8 .8;.8 .8 1; .4 .4 1])
imagesc(test_x(:),test_y(:),(reshape(a+(a~=dataStruct_fuzzyART(dataSetNum).test_output).*((a==2)-(a==1)),length(test_x),length(test_x))));
set(gca,'ydir','normal')
hold on
else
colormap([1 .7 .7;.7 .7 1])
imagesc(test_x(:),test_y(:),(reshape(a,length(test_x),length(test_x))));
set(gca,'ydir','normal')
hold on
end
drawRectsArtmap;
axis equal
axis tight
axis([-.1 1.1 -.1 1.1]);
% title(['fuzzy ARTMAP: feature selection in ART search cycle; Type (One Epoch)'],'fontsize',12);
% ylabel(['\lambda = ' num2str(c.lambda_Attention,3)],'fontsize',12)
if dataSetNum~=1
xlabel(['Accuracy : ', num2str(sum(dataStruct_fuzzyART(dataSetNum).test_output==a)*100/length(dataStruct_fuzzyART(dataSetNum).test_output),3),'%, Committed Nodes: ',num2str(c.C), ......
],'fontsize',12);
end
if (perc_error>0)
figure(2)
if dataSetNum~=1
colormap([1 .4 .4; 1 .8 .8;.8 .8 1; .4 .4 1])
a_test=dataStruct_fuzzyART(dataSetNum).test_output;
colormap([1 .8 .8;.8 .8 1])
imagesc(test_x(:),test_y(:),(reshape(a_test+(a_test~=dataStruct_fuzzyART(dataSetNum).test_output).*((a_test==2)-(a_test==1)),length(test_x),length(test_x))));
set(gca,'ydir','normal')
hold on
else
colormap([1 .7 .7;.7 .7 1])
imagesc(test_x(:),test_y(:),(reshape(a,length(test_x),length(test_x))));
set(gca,'ydir','normal')
hold on
end
for i=1:size(c_zero.w,2)
if (sum(c_zero.w(:,i))==2)
rect_created=[(min(c_zero.w(1,i),1-c_zero.w(3,i))) (min(c_zero.w(2,i),1-c_zero.w(4,i))) abs(1-c_zero.w(3,i)-c_zero.w(1,i))+.01 abs(1-c_zero.w(4,i)-c_zero.w(2,i))+.01];
%IF RESCALING
%rect_created=[(min((9*c.w(1,i)-4),1-(9*c.w(3,i)-4))) (min((9*c.w(2,i)-4),1-(9*c.w(4,i)-4))) abs(1-(9*c.w(3,i)-4)-(9*c.w(1,i)-4))+.01 abs(1-(9*c.w(4,i)-4)-(9*c.w(2,i)-4))+.01];
if c_zero.W(i,1)==0
rectangle('Position',rect_created,'EdgeColor',[0 0 1],'LineWidth',1.5)
hold on
%text((rect_created(1)+rect_created(3)/2),(rect_created(2)+rect_created(4)/2),'\beta')
else
rectangle('Position',rect_created,'EdgeColor',[1 0 0],'LineWidth',1.5)
hold on
%text((rect_created(1)+rect_created(3)/2),(rect_created(2)+rect_created(4)/2),'\alpha')
end
end
end
axis equal
axis tight
axis([-.1 1.1 -.1 1.1]);
% title(['fuzzy ARTMAP: feature selection in ART search cycle; Type ' num2str(c_zero.model_num) ' (One Epoch)'],'fontsize',12);
% ylabel(['\lambda = ' num2str(c_zero.lambda_Attention,3)],'fontsize',12)
% if dataSetNum~=1
% xlabel(['Accuracy% : ', num2str(sum(dataStruct_fuzzyART(dataSetNum).test_output==a_zero)*100/length(dataStruct_fuzzyART(dataSetNum).test_output),3),', C Nodes: ',num2str(c_zero.C), ...
% ', \alpha: ',num2str(c_zero.alpha)......
% ', Training Pts : ',num2str(length(dataStruct_fuzzyART(dataSetNum).training_input))...
% ', U Fail : ',num2str(c_zero.uncommitted_fail) ', NIT fail : ',num2str(length(c_zero.NI_fail))],'fontsize',12);
% end
end
try
depleted_count=0;
for i=1:length(c.learn_trail)
if ~isempty(c.learn_trail(dataSetNum).MT_nodes)
depleted_count=depleted_count+1;
depleted_sample(depleted_count)=i;
depleted_num(depleted_count)=length(c.learn_trail(dataSetNum).MT_nodes);
depleted_A(depleted_count,:)=c.learn_trail(dataSetNum).MT_Adash_after{end};
end
end
depleted_sample=depleted_sample(:);
depleted_num=depleted_num(:);
min(sum(depleted_A,2))
catch
disp('Trail Not Available');
end
elseif ismember(c.M,6)
inputNet.lambda_Attention=0;
num_mapping=c.W*[1 2]';
max_vals=(b==repmat(max(b,[],2),1,3));
ties=(mod(max_vals*(2*num_mapping+1),2)==0);
inputNet.gamma=0.01;
[a_nodep,b_nodep,c_nodep]=Default_ARTMAP(dataStruct_fuzzyART(dataSetNum),inputNet);
test_out=num2str(dataStruct_fuzzyART(dataSetNum).test_input');
test_out=test_out(:,1:3:size(test_out,2));
num_mapping=c_nodep.W*[1 2]';
max_vals=(b_nodep==repmat(max(b_nodep,[],2),1,3));
ties_nodep=(mod(max_vals*(2*num_mapping+1),2)==0);
show_val={'+','-','T'};
a_with_ties=a.*(~ties)+3*ties;
a_nodep_with_ties=a_nodep.*(~ties_nodep)+3*ties_nodep;
[strvcat(num2str(round(b_nodep*1000)/1000,'%1.3f\t')) ...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
[show_val{a_nodep_with_ties}]'...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
[show_val{a_with_ties}]'...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
strvcat(num2str(round(b*1000)/1000,'%1.3f\t')) ...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
repmat(' ',2^(c.M),1)...
strvcat(test_out)]
disp('Not Displaying Outputs: More than 2 input dimensions')
disp('Total number of differences: Depletion versus no depletion')
sum(a~=a_nodep)
elseif c.M==41
disp('Finished running Boston dataset simulation')
if dataSetNum<13
title_boston=['Boston dataset ' num2str(dataSetNum-8) ' (uncapped) \lambda=' num2str(c.lambda_Attention,2)];
end
[conf_nonnorm,conf_norm,avg_acc]= confMatrixPlot(dataStruct_fuzzyART(dataSetNum).test_output(:),a(:),c,dataStruct_fuzzyART(dataSetNum).labels,title_boston);
elseif strcmp(dataStruct_fuzzyART(dataSetNum).description,'movie_genres')
b_mean=mean(b,1);
b_std=std(b,1);
b_new=(b-repmat(b_mean,length(dataStruct_fuzzyART(dataSetNum).test_output),1))./repmat(b_std,length(dataStruct_fuzzyART(dataSetNum).test_output),1);
top_node_num=round(c.C*.02);
dum=sort(b_new,2,'descend');
y_rect=b_new.*(b_new>repmat(mean(b_new,2),1,c.C)).*(b_new>repmat(dum(:,top_node_num),1,c.C));
sigma = c.W'*y_rect';
sigma_mean=mean(sigma,2);
sigma_std=std(sigma',1)';
sigma_new=(sigma-repmat(sigma_mean,1,length(dataStruct_fuzzyART(dataSetNum).test_output)))./repmat(sigma_std,1,length(dataStruct_fuzzyART(dataSetNum).test_output));
[dum,classes]=max(sigma,[],1);
[conf_nonnorm,conf_norm,avg_acc]= confMatrixPlot(dataStruct_fuzzyART(dataSetNum).test_output(:),classes,c,dataStruct_fuzzyART(dataSetNum).labels,'');
end
disp('Output class predictions are stored in variable ''a''.');
disp('Distributed output predictions are stored in ''b''.');
disp('The ARTMAP network details are stored in ''c''');
