Daniel@0: function [e, edata, eprior] = mlperr_weighted(net, x, t, eso_w)
Daniel@0: %MLPERR	Evaluate error function for 2-layer network.
Daniel@0: %
Daniel@0: %	Description
Daniel@0: %	E = MLPERR(NET, X, T) takes a network data structure NET together
Daniel@0: %	with a matrix X of input vectors and a matrix T of target vectors,
Daniel@0: %	and evaluates the error function E. The choice of error function
Daniel@0: %	corresponds to the output unit activation function. Each row of X
Daniel@0: %	corresponds to one input vector and each row of T corresponds to one
Daniel@0: %	target vector.
Daniel@0: %
Daniel@0: %	[E, EDATA, EPRIOR] = MLPERR(NET, X, T) additionally returns the data
Daniel@0: %	and prior components of the error, assuming a zero mean Gaussian
Daniel@0: %	prior on the weights with inverse variance parameters ALPHA and BETA
Daniel@0: %	taken from the network data structure NET.
Daniel@0: %
Daniel@0: %	See also
Daniel@0: %	MLP, MLPPAK, MLPUNPAK, MLPFWD, MLPBKP, MLPGRAD
Daniel@0: %
Daniel@0: 
Daniel@0: %	Copyright (c) Ian T Nabney (1996-9)
Daniel@0: 
Daniel@0: % Check arguments for consistency
Daniel@0: errstring = consist(net, 'mlp', x, t);
Daniel@0: if ~isempty(errstring);
Daniel@0:   error(errstring);
Daniel@0: end
Daniel@0: [y, z, a] = mlpfwd(net, x);
Daniel@0: 
Daniel@0: switch net.actfn
Daniel@0: 
Daniel@0:   case 'linear'        %Linear outputs
Daniel@0: 
Daniel@0:     edata = 0.5*sum(sum((y - t).^2));
Daniel@0: 
Daniel@0:   case 'logistic'      % Logistic outputs
Daniel@0: 
Daniel@0:     % Ensure that log(1-y) is computable: need exp(a) > eps
Daniel@0:     maxcut = -log(eps);
Daniel@0:     % Ensure that log(y) is computable
Daniel@0:     mincut = -log(1/realmin - 1);
Daniel@0:     a = min(a, maxcut);
Daniel@0:     a = max(a, mincut);
Daniel@0:     y = 1./(1 + exp(-a));
Daniel@0:     edata = - sum(sum(t.*log(y) + (1 - t).*log(1 - y)));
Daniel@0: 
Daniel@0:   case 'softmax'       % Softmax outputs
Daniel@0:   
Daniel@0:     nout = size(a,2);
Daniel@0:     % Ensure that sum(exp(a), 2) does not overflow
Daniel@0:     maxcut = log(realmax) - log(nout);
Daniel@0:     % Ensure that exp(a) > 0
Daniel@0:     mincut = log(realmin);
Daniel@0:     a = min(a, maxcut);
Daniel@0:     a = max(a, mincut);
Daniel@0:     temp = exp(a);
Daniel@0:     y = temp./(sum(temp, 2)*ones(1,nout));
Daniel@0:     % Ensure that log(y) is computable
Daniel@0:     y(y<realmin) = realmin;
Daniel@0:     e_app=sum(t.*log(y),2);
Daniel@0:     edata = - eso_w'*e_app;
Daniel@0:     clear e_app;
Daniel@0:     
Daniel@0:   otherwise
Daniel@0:     error(['Unknown activation function ', net.actfn]);  
Daniel@0: end
Daniel@0: 
Daniel@0: [e, edata, eprior] = errbayes(net, edata);