Daniel@0: function [mask, prior] = rbfprior(rbfunc, nin, nhidden, nout, aw2, ab2)
Daniel@0: %RBFPRIOR Create Gaussian prior and output layer mask for RBF.
Daniel@0: %
Daniel@0: %	Description
Daniel@0: %	[MASK, PRIOR] = RBFPRIOR(RBFUNC, NIN, NHIDDEN, NOUT, AW2, AB2)
Daniel@0: %	generates a vector MASK  that selects only the output layer weights.
Daniel@0: %	This is because most uses of RBF networks in a Bayesian context have
Daniel@0: %	fixed basis functions with the output layer as the only adjustable
Daniel@0: %	parameters.  In particular, the Neuroscale output error function is
Daniel@0: %	designed to work only with this mask.
Daniel@0: %
Daniel@0: %	The return value PRIOR is a data structure,  with fields PRIOR.ALPHA
Daniel@0: %	and PRIOR.INDEX, which specifies a Gaussian prior distribution for
Daniel@0: %	the network weights in an RBF network. The parameters AW2 and AB2 are
Daniel@0: %	all scalars and represent the regularization coefficients for two
Daniel@0: %	groups of parameters in the network corresponding to  second-layer
Daniel@0: %	weights, and second-layer biases respectively. Then PRIOR.ALPHA
Daniel@0: %	represents a column vector of length 2 containing the parameters, and
Daniel@0: %	PRIOR.INDEX is a matrix specifying which weights belong in each
Daniel@0: %	group. Each column has one element for each weight in the matrix,
Daniel@0: %	using the standard ordering as defined in RBFPAK, and each element is
Daniel@0: %	1 or 0 according to whether the weight is a member of the
Daniel@0: %	corresponding group or not.
Daniel@0: %
Daniel@0: %	See also
Daniel@0: %	RBF, RBFERR, RBFGRAD, EVIDENCE
Daniel@0: %
Daniel@0: 
Daniel@0: %	Copyright (c) Ian T Nabney (1996-2001)
Daniel@0: 
Daniel@0: nwts_layer2 = nout + (nhidden *nout);
Daniel@0: switch rbfunc
Daniel@0: case 'gaussian'
Daniel@0:    nwts_layer1 = nin*nhidden + nhidden;
Daniel@0: case {'tps', 'r4logr'}
Daniel@0:    nwts_layer1 = nin*nhidden;
Daniel@0: otherwise
Daniel@0:    error('Undefined activation function');
Daniel@0: end  
Daniel@0: nwts = nwts_layer1 + nwts_layer2;
Daniel@0: 
Daniel@0: % Make a mask only for output layer
Daniel@0: mask = [zeros(nwts_layer1, 1); ones(nwts_layer2, 1)];
Daniel@0: 
Daniel@0: if nargout > 1
Daniel@0:   % Construct prior
Daniel@0:   indx = zeros(nwts, 2);
Daniel@0:   mark2 = nwts_layer1 + (nhidden * nout);
Daniel@0:   indx(nwts_layer1 + 1:mark2, 1) = ones(nhidden * nout, 1);
Daniel@0:   indx(mark2 + 1:nwts, 2) = ones(nout, 1);
Daniel@0: 
Daniel@0:   prior.index = indx;
Daniel@0:   prior.alpha = [aw2, ab2]';
Daniel@0: end