function updated_artmap = DEFARTMAP_Learn(artmap, inputVectors, classes)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% This is an implementation of Default ARTMAP,
% as described in CAS/CNS Technical Report TR-2003-008
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Programmed by Ogi Ogas
%%%%%%%%%%%%%%%%%%%%%%%%%
% The notation follows the article's notation,
% as does the headers for each step of the
% algorithm
% 02/21/05 Added search cycle count to artmap structure
% 08.31.05 Added list of nodes activated for each training input (This is
% only used for certain priming simulations; otherwise, this is ignored.)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% WINNER-TAKE-ALL TRAINING
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
ON=1; OFF=0;
DISPLAY_WEIGHTBOXES=OFF;
DISPLAY_ONLYFINALWEIGHTBOX=OFF;
VISUALS=OFF;
DESCRIPT=OFF;
SHORT_DESCRIPT=OFF;
LEARN_TRAIL=ON;
disp('Using Original Bloated Version!!')
MT_nodes=0;
space_amt=' ';
% Make sure that the data is appropriate for the given network.
[numFeatures, n] = size(inputVectors);
if (numFeatures ~= 2*artmap.M)
error('The input vector does not contain the same number of features as the network.');
end
[artmap.learn_trail(1:n).A]=deal([]);
[artmap.learn_trail(1:n).K]=deal([]);
[artmap.learn_trail(1:n).winning_node]=deal([]);
[artmap.learn_trail(1:n).weight_before]=deal([]);
[artmap.learn_trail(1:n).weight_after]=deal([]);
[artmap.learn_trail(1:n).MT_nodes]=deal([]);
[artmap.learn_trail(1:n).MT_mismatch]=deal([]);
[artmap.learn_trail(1:n).MT_x_before]=deal([]);
[artmap.learn_trail(1:n).MT_x_after]=deal([]);
[artmap.learn_trail(1:n).MT_e_before]=deal([]);
[artmap.learn_trail(1:n).MT_e_after]=deal([]);
[artmap.learn_trail(1:n).MT_Adash_before]=deal([]);
[artmap.learn_trail(1:n).MT_Adash_after]=deal([]);
% Initialize the return artmap network
updated_artmap = {};
A=[]; K=[];
search_cycles = 0;
% Classify and learn on each sample.
for sampleNumber = 1:n
% Reset the activated node (NOTE: this is Moshe Bar addition)
activated_node=[];
% Get the next input vector.
% B.3 / B.6
A = inputVectors(:, sampleNumber);
MT_nodes=0;
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).A=A;
end
% Get the next class
% B.3 / B.6
K = classes(sampleNumber);
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).K=K;
end
% B.5
artmap.rho = artmap.rho_bar;
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Sample number: ',num2str(sampleNumber) ' =======================================================']);
artmap.learn_description=strvcat(artmap.learn_description,['Input presented: ',regexprep(num2str(A(:)',3),'\s*',' ')]);
artmap.learn_description=strvcat(artmap.learn_description,['Category of new input: ',num2str(K)]);
end
% Create a new category if this supervisory signal
% has never been seen before.
if (artmap.C==0)
% B.4 Set initial weights for newly committed coding node j=C=1
artmap.C=1;
j=1;
artmap.w(:,1) = A;
artmap.W = zeros(1,artmap.numClasses);
artmap.W(1,K)=1;
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).winning_node=artmap.C;
artmap.learn_trail(sampleNumber).weight_before = A';
artmap.learn_trail(sampleNumber).weight_after = A';
end
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['This is the first category box. Weights are set to: ', regexprep(num2str(artmap.w(:,1)',3) ,'\s*',' ')]);
artmap.learn_description=strvcat(artmap.learn_description,'First Point--no category boxes!');
artmap.learn_description=strvcat(artmap.learn_description,['Adding New Node']);
artmap.learn_description=strvcat(artmap.learn_description,['Input: ', regexprep(num2str(A',3),'\s*',space_amt)]);
artmap.learn_description=strvcat(artmap.learn_description,['Weight Node Number: ', num2str(artmap.C)]);
artmap.learn_description=strvcat(artmap.learn_description,['Newly Added Weight: ', regexprep(num2str(artmap.w(:,artmap.C)',3),'\s*',space_amt)]);
end
if (SHORT_DESCRIPT==ON)
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Adding New Node']);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Input: ', regexprep(num2str(A',3),'\s*',space_amt)]);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Weight Node Number: ', num2str(artmap.C)]);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Newly Added Weight: ', regexprep(num2str(artmap.w(:,artmap.C)',3),'\s*',space_amt)]);
end
y = zeros(1,artmap.C); % the code at F2. At each node j, activity reset to 0.
activated_node=1; % The first and only node is activated.
else
% B.7 Calculate signals to committed coding nodes
T = DEFARTMAP_choice_function(artmap,A);
% B.8 Sort committed coding node signals (A passing through
% w_ij)
% This next line sorts Tj in decreasing order
[T_sorted, T_j] = sort(-T);
% This next line orders the coding nodes j with Tj>alpha*M, still in decreasing
% order. We will need j itself (remember, j is an integer
% from 1 to C) for next step.
T_j = T_j(find(-T_sorted>artmap.alpha.*artmap.M));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% B.9 Find appropriate coding node (you're only going to pick one
% coding node j, the best j for the input vector A).
% Go down coding nodes, as ordered by the sorted T_j
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% B.9a Take next coding node j=J (from sorted T_j), and see if
% there is a suitable match.
MATCH=0;
%These additions are specific to ARTMAP Attention
%=========ARTMAP Attention=======
%attention parameter resets to zero at every input presentation
e_Attention=zeros(size(A));
A_dash=A;
%================================
for j=T_j
if (artmap.use_type_three_tracking==1)
%%%%%%%%%%Uncommitted nodes will always pass this test%%%%%%%%%%%%%%
comparing_to_rho = sum(min(A_dash,artmap.w(:,j)))./(sum(A_dash)+eps);
artmap.num_type_three_used=artmap.num_type_three_used+(sum(A_dash)~=sum(A));
else
comparing_to_rho = sum(min(A_dash,artmap.w(:,j)))./artmap.M;
end
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Computing match between depleted input A and weight ' num2str(j) ': ' regexprep( num2str(artmap.w(:,j)',3) ,'\s*',space_amt)]);
artmap.learn_description=strvcat(artmap.learn_description,['Original input A: ', regexprep(num2str(A',3),'\s*',space_amt)]);
artmap.learn_description=strvcat(artmap.learn_description,['Depleted input A: ', regexprep(num2str(A_dash',3),'\s*',space_amt)]);
artmap.learn_description=strvcat(artmap.learn_description,['Match value sum(min(A_depleted , w_j)/M) rho: ',num2str(comparing_to_rho)]);
end
%%%%%%%%%%%Match Tracking: Change for Type III%%%%%%%%%%%%%%
if (artmap.use_type_three_tracking==1)
MATCH_TRACK_TEST=(sum(min(A_dash,artmap.w(:,j)))./(sum(A_dash)+eps) >= artmap.rho);
else
MATCH_TRACK_TEST=(sum(min(A_dash,artmap.w(:,j)))./artmap.M >= artmap.rho);
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).rho_calc=sum(min(A_dash,artmap.w(:,j)))./(sum(A_dash)+eps);
artmap.learn_trail(sampleNumber).rho = artmap.rho;
end
if MATCH_TRACK_TEST
search_cycles = search_cycles+1;
% Then set Yj=1: Winner-take-all
% B.9b (Not necessary with WTA learning)
% B.9c CORRECT PREDICTION?
J=find(artmap.W(j,:));
if J==K
% if sigma_k = W_JK = 1
MATCH=1;
% B.11 LEARNING: UPDATING WEIGHTS OF EXISTING
% CODING NODE IN F2
artmap.win_sequence(length(artmap.win_sequence)+1)=j;
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,'Correct match between input and the (next) nearest category box.');
artmap.learn_description=strvcat(artmap.learn_description,['Changing weight Node Number: ', num2str(j)]);
artmap.learn_description=strvcat(artmap.learn_description,['Weight for this category box before modification: ', regexprep(num2str( artmap.w(:,j)',3),'\s*',' ')]);
end
if (SHORT_DESCRIPT==ON)
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Changing weight']);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Input: ', regexprep(num2str(A',3),'\s*',space_amt)]);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Weight node number: ', num2str(j)]);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Weight before modification: ', regexprep( num2str(artmap.w(:,j)',3) ,'\s*',space_amt)]);
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).winning_node=j;
artmap.learn_trail(sampleNumber).weight_before = artmap.w(:,j);
end
artmap.w=DEFARTMAP_Update_Weights(artmap,j,A);
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).weight_after = artmap.w(:,j);
end
if (SHORT_DESCRIPT==ON)
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Weight after modification: ', regexprep( num2str(artmap.w(:,j)',3) ,'\s*',space_amt)]);
end
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Weight after modification: ', regexprep( num2str(artmap.w(:,j)',3) ,'\s*',space_amt)]);
end
if (DESCRIPT==ON)
%disp('the updated weights for this category box are now:')
%artmap.w(:,j)
%disp('Weight boxes: one should be stretched, or holds new point')
artmap.learn_description=strvcat(artmap.learn_description,'Finished Learning on this input; moving to next input');
end
if (DISPLAY_WEIGHTBOXES==ON)
GRAPH_WEIGHTS(artmap,A,K)
end
activated_node=j;
break; % Don't check any more T_j
else % if J~=K
% if sigma_k = 0
% B.9d MATCH TRACKING: RAISE VIGILANCE
%%%%%%%%%%%Match Tracking: Change for Type III%%%%%%%%%%%%%%
if LEARN_TRAIL==ON
MT_nodes=MT_nodes+1;
artmap.learn_trail(sampleNumber).MT_mismatch{MT_nodes}=1;
artmap.learn_trail(sampleNumber).MT_nodes{MT_nodes}=j;
artmap.learn_trail(sampleNumber).MT_rho_before{MT_nodes} =artmap.rho;
end
if (artmap.use_type_three_tracking==1)
artmap.rho = sum(min(A_dash,artmap.w(:,j)))./(sum(A_dash)+eps) + artmap.epsilon;
else
artmap.rho = sum(min(A_dash,artmap.w(:,j)))./artmap.M + artmap.epsilon;
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_rho_after{MT_nodes} =artmap.rho;
end
if (DESCRIPT==ON)
%disp('INCORRECT match. Raise vigilance. Go to next closest category box (This is called match tracking, not really vigilance!')
artmap.learn_description=strvcat(artmap.learn_description,['INCORRECT match; match tracking and updating rho: ', num2str(artmap.rho)]);
end
% %=========ARTMAP Attention=======
%This is without normalization of any sort
x_Attention=min(A_dash,artmap.w(:,j));
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_x_before{MT_nodes} =x_Attention';
end
%This is with normalization
%======================================================
%=============
if artmap.type_depletion==4
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Match Tracking triggers depletion; dyad before normalization: ', regexprep(num2str(x_Attention(:)',3),'\s*',' ')]);
end
if (artmap.deplete_in_X_denom==1)
x_Attention=max(x_Attention+[x_Attention(artmap.M+1:end,1) ; x_Attention(1:artmap.M,1)]-sum(x_Attention)/(sum(A_dash)+eps),0);
else
x_Attention=max(x_Attention+[x_Attention(artmap.M+1:end,1) ; x_Attention(1:artmap.M,1)]-sum(x_Attention)/artmap.M,0);
end
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Match Tracking triggers depletion; dyad after normalization: ', regexprep(num2str(x_Attention(:)',3),'\s*',' ')]);
end
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_x_after{MT_nodes} =x_Attention';
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_e_before{MT_nodes} =e_Attention';
end
e_Attention=max(e_Attention,artmap.lambda_Attention*x_Attention);
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_e_after{MT_nodes} =e_Attention';
end
%Attention can be too harsh if it is depleting along
%many dimension. To offset this, we could scale
%according to
if artmap.type_depletion==1
%A) the number of dimensions that are non-zero
%==========
e_Attention=e_Attention/max(sum(sign02(e_Attention)),1);
%B) The sum total of attention, implying that attention
%is a constant
elseif artmap.type_depletion==2
e_Attention=e_Attention/(sum(e_Attention)+eps)*artmap.lambda_Attention;
%C) The attention graph is used to comparatively scale things implying that attention
%is a constant
elseif artmap.type_depletion==3
e_Attention=e_Attention/(sum(e_Attention)+eps).*e_Attention;
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_Adash_before{MT_nodes} =A_dash';
end
A_dash=rect(A-e_Attention,0);
% %================================
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_Adash_after{MT_nodes} =A_dash';
end
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Match Tracking triggers depletion. e_Attention: ', regexprep(num2str(e_Attention(:)',3),'\s*',' ')]);
artmap.learn_description=strvcat(artmap.learn_description,['Depleted input: ', regexprep(num2str(A_dash(:)',3),'\s*',' ')]);
end
end % if J==K b
else % if (sum(min(A,artmap.w(:,j)))./artmap.M >= artmap.rho)
if (DESCRIPT==ON)
%disp('Did not pass vigilance test. Do not change vigilance. Moving to next category box.')
artmap.learn_description=strvcat(artmap.learn_description,'Did not pass vigilance test. Do not change vigilance. Moving to next category box.');
end
if (artmap.deplete_in_mismatch==1)
if LEARN_TRAIL==ON
MT_nodes=MT_nodes+1;
artmap.learn_trail(sampleNumber).MT_nodes{MT_nodes}=j;
end
if (DESCRIPT==ON)
%disp('INCORRECT match. Raise vigilance. Go to next closest category box (This is called match tracking, not really vigilance!')
artmap.learn_description=strvcat(artmap.learn_description,['INCORRECT match; match tracking and updating rho: ', num2str(artmap.rho)]);
end
% %=========ARTMAP Attention=======
%This is without normalization of any sort
x_Attention=min(A_dash,artmap.w(:,j));
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_x_before{MT_nodes} =x_Attention';
end
%This is with normalization
%======================================================
%=============
if artmap.type_depletion==4
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Match Tracking triggers depletion; dyad before normalization: ', regexprep(num2str(x_Attention(:)',3),'\s*',' ')]);
end
if (artmap.deplete_in_X_denom==1)
x_Attention=max(x_Attention+[x_Attention(artmap.M+1:end,1) ; x_Attention(1:artmap.M,1)]-sum(x_Attention)/(sum(A_dash)+eps),0);
else
x_Attention=max(x_Attention+[x_Attention(artmap.M+1:end,1) ; x_Attention(1:artmap.M,1)]-sum(x_Attention)/artmap.M,0);
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_x_after{MT_nodes} =x_Attention';
end
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Match Tracking triggers depletion; dyad after normalization: ', regexprep(num2str(x_Attention(:)',3),'\s*',' ')]);
end
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_e_before{MT_nodes} =e_Attention';
end
e_Attention=max(e_Attention,artmap.lambda_Attention*x_Attention);
%Attention can be too harsh if it is depleting along
%many dimension. To offset this, we could scale
%according to
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_e_after{MT_nodes} =e_Attention';
end
if artmap.type_depletion==1
%A) the number of dimensions that are non-zero
%==========
e_Attention=e_Attention/max(sum(sign02(e_Attention)),1);
%B) The sum total of attention, implying that attention
%is a constant
elseif artmap.type_depletion==2
e_Attention=e_Attention/(sum(e_Attention)+eps)*artmap.lambda_Attention;
%C) The attention graph is used to comparatively scale things implying that attention
%is a constant
elseif artmap.type_depletion==3
e_Attention=e_Attention/(sum(e_Attention)+eps).*e_Attention;
end
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_Adash_before{MT_nodes} =A_dash';
end
A_dash=rect(A-e_Attention,0);
% %================================
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).MT_Adash_after{MT_nodes} =A_dash';
end
if (DESCRIPT==ON)
artmap.learn_description=strvcat(artmap.learn_description,['Match Tracking triggers depletion. e_Attention: ', regexprep(num2str(e_Attention(:)',3),'\s*',' ')]);
artmap.learn_description=strvcat(artmap.learn_description,['Depleted input: ', regexprep(num2str(A_dash(:)',3),'\s*',' ')]);
end
end
%=========ARTMAP Attention=======
%Two Options: a) Reset e_Attention
% b)continue with depletion
%================================
%=========ARTMAP Attention=======
% e_Attention=max(e_Attention,artmap.lamda_Attention*min(A_dash,artmap.w(:,j)));
% A_dash=rect(A-e_Attention,0);
%================================
end % if (sum(min(A,artmap.w(:,j)))./artmap.M >= artmap.rho)
end % for J=T_j
% B.10: ADD A NEW CODING NODE IN F2
% w(i,C) = A
if (MATCH==0) %
%sampleNumber
artmap.C=artmap.C+1;
artmap.w(:,artmap.C)=A;
if LEARN_TRAIL==ON
artmap.learn_trail(sampleNumber).winning_node=artmap.C;
artmap.learn_trail(sampleNumber).weight_before = A';
artmap.learn_trail(sampleNumber).weight_after = A';
end
if (DESCRIPT==ON)
%sprintf('None of the closest boxes are the right match. So create a new category box %d, with weights:,',artmap.C)
artmap.learn_description=strvcat(artmap.learn_description,['None of the closest boxes are the right match; create a new category box ',...
num2str(artmap.C), ' with weights: ', regexprep(num2str(artmap.w(:,artmap.C)',3),'\s*',space_amt)]);
%artmap.w(:,artmap.C)
%input('We will graph current weights, which should have a new point box, and then we can move to the next input.')
artmap.learn_description=strvcat(artmap.learn_description,'Finished Learning on this input; moving to next input');
end
if (SHORT_DESCRIPT==ON)
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Adding New Node']);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Input: ', regexprep(num2str(A',3),'\s*',space_amt)]);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Weight Node Number: ', num2str(artmap.C)]);
artmap.short_learn_descript=strvcat(artmap.short_learn_descript,['Newly Added Weight: ', regexprep(num2str(artmap.w(:,artmap.C)',3),'\s*',space_amt)]);
end
artmap.W=[artmap.W; zeros(1,artmap.numClasses)];
artmap.W(artmap.C,K)=1;
if (DISPLAY_WEIGHTBOXES==ON)
GRAPH_WEIGHTS(artmap,A,K)
end
activated_node=artmap.C; % Activated node is new node
end % (MATCH==0)
end % if (artmap.C==0)
artmap.NodeList = [artmap.NodeList, activated_node];
end % for sampleNumber = 1:numSamples
if (DISPLAY_ONLYFINALWEIGHTBOX==ON)
GRAPH_WEIGHTS(artmap,A,K)
end
% GRAPH_WEIGHTS(artmap,A,1)
% Fill the new network with the appropriate values.
artmap.search_cycles = search_cycles;
updated_artmap = artmap;
return
