% backprop a per-period backpropagation training for a multilayer feedforward

% neural network.
% Network = backprop(Layers,N,M,SatisfactoryMSE,Input,Desired) returns
% Network, a two field structure of the form Network.structure = Layers
% and Network.weights where weights is a cell array specifying the final
% weight matrices computed by minimizing the mean squared error between
% the Desired output and the actual output of the network given a set of
% training samples: Input and the SatisfactoryMSE (satisfactory mean
% squared error)
%
% Input:
% Layers - a vector of integers specifying the number of nodes at each
% layer, i.e for all i, Layers(i) = number of nodes at layer i, there
% must be at least three layers and the input layer Layers(1) must
% equal the dimension of each vector in Input, likewise, Layers(end)
% must be equal to the dimension of each vector in Desired
% N - training rate for network learning (0.1 - 0.9)
% M - momentum for the weight update rule [0.1 - 0.9)
% SatisfactoryMSE - the mse at which to terminate computation
% Input - the training samples, a P-by-N matrix, where each Input[p] is
% a training vector
% Desired - the desired outputs, a P-by-M matrix where each Desired[p]
% is the desired output for the corresponding input Input[p]
%
% This algorithm uses the hyperbolic tangent node function
% 2/(1+e^(-net)) - 1, for use with bipolar data
%
% NOTE: due to its generality this algorithm is not as efficient as a
% one designed for a specific problem if the number of desired layers is
% known ahead of time, it is better to a) 'unfold' the loops inside the
% loop presenting the data. That is, calculate the input and output of each
% layer explicitly one by one and subsequently the modified error and weight
% matrix modifications b) remove momentum and training rate as parameters
% if they are known
%
% Author: Dale Patterson
% $Version: 2.2.1 $ $Date: 2.25.06 $
%
function Network = backprop(L,n,m,smse,X,D)
% determine number of input samples, desired output and their dimensions
[P,N] = size(X);
[Pd,M] = size(D);
% make user that each input vector has a corresponding desired output
if P ~= Pd
error('backprop:invalidTrainingAndDesired', ...
'The number of input vectors and desired ouput do not match');
end
% make sure that at least 3 layers have been specified and that the
% the dimensions of the specified input layer and output layer are
% equivalent to the dimensions of the input vectors and desired output
if length(L) < 3
error('backprop:invalidNetworkStructure','The network must have at least 3 l
ayers');
else
if N ~= L(1) || M ~= L(end)
e = sprintf('Dimensions of input (%d) does not match input layer (%d)',N
,L(1));
 error('backprop:invalidLayerSize', e);
elseif M ~= L(end)
e = sprintf('Dimensions of output (%d) does not match output layer (%d)'
,M,L(end));
error('backprop:invalidLayerSize', e);
end
end
% will use the number of layers often, so save the number here
nLayers = length(L);
% randomize the weight matrices (uniform random values in [-.5 .5], there
% is a weight matrix between each layer of nodes. Each layer (exclusive the
% output layer) has a bias node whose activation is always 1, that is, the
% node function is C(net) = 1. Furthermore, there is a link from each node
% in layer i to the bias node in layer j (the last row of each matrix)
% because this is less computationally expensive then the alternative.
% NOTE: below that the wieghts of all links to bias nodes are defined as
% zero
w = cell(nLayers-1,1); % a weight matrix between each layer
for i=1:nLayers-2
w{i} = [.5 - rand(L(i+1),L(i)+1) ; zeros(1,L(i)+1)];
end
w{end} = .5 - rand(L(end),L(end-1)+1);
X = [X ones(P,1)]; % affix the column of bias activations to the input layer
% preallocate activation,net vectors and delta weight matrices for faster
% computation
% activation vectors, all but output layer include bias activation
a = cell(nLayers,1);
for i=1:nLayers-1
a{i} = ones(L(i)+1,1);
end
a{end} = ones(L(end),1);
% net vectors, one for each node in that layer but there is
% no net for input layer
net = cell(nLayers-1,1);
for i=1:nLayers-2;
net{i} = ones(L(i+1)+1,1);
end
net{end} = ones(L(end),1);
% delta weight matrices
dw = cell(nLayers-1,1);
for i=1:nLayers-1
dw{i} = zeros(size(w{i}));
end
% initialize stopping conditions
mse = Inf; % assuming the intial weight matrices are bad
presentations = 0; % we'll measure by epoch instead of presentation
% loop until computational bounds are exceeded or the network has converged
% to a satisfactory condition. We allow for 30000 epochs, it may be
% necessary to reduce this if the number of training samples is large
while mse > smse && presentations < P * 10000
sse = 0; % running total of squared error
for p=1:P
 % get the current input vector and desired output
a{1} = X(p,:)';
Dp = D(p,:)';
% compute the inputs and outputs to each layer
for i=1:nLayers-1
% compute inputs to this layer
net{i} = w{i} * a{i};
% compute outputs of this layer
% for all layers but the output layer, the last node is the
% bias node and its activation is 1
if i < nLayers-1
a{i+1} = [2./(1+exp(-net{i}(1:end-1)))-1 ; 1];
else
a{i+1} = 2./(1+exp(-net{i})) - 1;
end
end
% accumlate the squared error
sse = sse + sum((Dp-a{end}).^2);
% calculate the modified error at each layer and update the weight
% matrices accordingly. first calculate delta, the modified error
% for the output nodes (S'(Output[net])*(Dp-Output[Activation])
% then for each weight matrix, add n * delta * activation and
% propagate delta to the previous layer
delta = (Dp-a{end}) .* (1+a{end}) .* (1-a{end});
for i=nLayers-1:-1:1
dw{i} = n * delta * a{i}' + (m .* dw{i});
w{i} = w{i} + dw{i};
if i > 1 % dont compute mod err for input layer
delta = (1+a{i}).*(1-a{i}).*(delta'*w{i})';
end
end
end
presentations = presentations + P;
mse = sse/(P*M); % mse = 1/P * 1/M * summed squared error
end
% return the trained network
Network.structure = L;
Network.weights = w;
Network.mse = mse;
Network.presentations = presentations;