clear all;

nX = 5;                 % number of inputs
nH = 5;                 % number of hidden units
nY = 3;                 % number of outputs
nD = 100;               % number of data points
Trn = 4;                % transfer function type for all units

N = nX+nH+nY;           % compute total number of units

Con = tril(ones(N));    % make fully-connected feedforward network
Con = Con - eye(N);

X  = randn(nX, nD);     % generate random data and weights
Y  = randn(nX, nD);
W0 = randn(N) .* Con;

[err, grad, out] = backprop( X, Y, W0, Con, Trn);
Y = out;                % set desired outputs to actual outputs

W =  randn(N).*Con;     % choose different random weights

for k=1:500,            % gradient descent with fixed learning rate
   [err(k), grad] = backprop( X, Y, W, Con, Trn);
   W = W - 0.001*grad;
end

figure(1);              % plot learning curve
clf;
subplot(1,2,1);
plot(err);
xlabel('Iteration');
ylabel('Error');
subplot(1,2,2);         % compare learned and correct weights
plot(W0(:), W(:), '.');
xlabel('Correct weight');
ylabel('Learned weight');