wolffd@0: function prior = mlpprior(nin, nhidden, nout, aw1, ab1, aw2, ab2)
wolffd@0: %MLPPRIOR Create Gaussian prior for mlp.
wolffd@0: %
wolffd@0: %	Description
wolffd@0: %	PRIOR = MLPPRIOR(NIN, NHIDDEN, NOUT, AW1, AB1, AW2, AB2)  generates a
wolffd@0: %	data structure PRIOR, with fields PRIOR.ALPHA and PRIOR.INDEX, which
wolffd@0: %	specifies a Gaussian prior distribution for the network weights in a
wolffd@0: %	two-layer feedforward network. Two different cases are possible. In
wolffd@0: %	the first case, AW1, AB1, AW2 and AB2 are all scalars and represent
wolffd@0: %	the regularization coefficients for four groups of parameters in the
wolffd@0: %	network corresponding to first-layer weights, first-layer biases,
wolffd@0: %	second-layer weights, and second-layer biases respectively. Then
wolffd@0: %	PRIOR.ALPHA represents a column vector of length 4 containing the
wolffd@0: %	parameters, and PRIOR.INDEX is a matrix specifying which weights
wolffd@0: %	belong in each group. Each column has one element for each weight in
wolffd@0: %	the matrix, using the standard ordering as defined in MLPPAK, and
wolffd@0: %	each element is 1 or 0 according to whether the weight is a member of
wolffd@0: %	the corresponding group or not.  In the second case the parameter AW1
wolffd@0: %	is a vector of length equal to the number of inputs in the network,
wolffd@0: %	and the corresponding matrix PRIOR.INDEX now partitions the first-
wolffd@0: %	layer weights into groups corresponding to the weights fanning out of
wolffd@0: %	each input unit. This  prior is appropriate for the technique of
wolffd@0: %	automatic relevance determination.
wolffd@0: %
wolffd@0: %	See also
wolffd@0: %	MLP, MLPERR, MLPGRAD, EVIDENCE
wolffd@0: %
wolffd@0: 
wolffd@0: %	Copyright (c) Ian T Nabney (1996-2001)
wolffd@0: 
wolffd@0: nextra = nhidden + (nhidden + 1)*nout;
wolffd@0: nwts = nin*nhidden + nextra;
wolffd@0: 
wolffd@0: if size(aw1) == [1,1] 
wolffd@0: 
wolffd@0:     indx = [ones(1, nin*nhidden), zeros(1, nextra)]';
wolffd@0:   
wolffd@0: elseif size(aw1) == [1, nin]
wolffd@0:   
wolffd@0:     indx = kron(ones(nhidden, 1), eye(nin));
wolffd@0:     indx = [indx; zeros(nextra, nin)];
wolffd@0: 
wolffd@0: else
wolffd@0:   
wolffd@0:     error('Parameter aw1 of invalid dimensions');
wolffd@0:     
wolffd@0: end
wolffd@0: 
wolffd@0: extra = zeros(nwts, 3);
wolffd@0: 
wolffd@0: mark1 = nin*nhidden;
wolffd@0: mark2 = mark1 + nhidden;
wolffd@0: extra(mark1 + 1:mark2, 1) = ones(nhidden,1);
wolffd@0: mark3 = mark2 + nhidden*nout;
wolffd@0: extra(mark2 + 1:mark3, 2) = ones(nhidden*nout,1);
wolffd@0: mark4 = mark3 + nout;
wolffd@0: extra(mark3 + 1:mark4, 3) = ones(nout,1);
wolffd@0: 
wolffd@0: indx = [indx, extra];
wolffd@0: 
wolffd@0: prior.index = indx;
wolffd@0: prior.alpha = [aw1, ab1, aw2, ab2]';