comparison toolboxes/FullBNT-1.0.7/netlab3.3/mlpprior.m @ 0:e9a9cd732c1e tip

first hg version after svn

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