|
wolffd@0
|
1 function [mask, prior] = rbfprior(rbfunc, nin, nhidden, nout, aw2, ab2)
|
|
wolffd@0
|
2 %RBFPRIOR Create Gaussian prior and output layer mask for RBF.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 % [MASK, PRIOR] = RBFPRIOR(RBFUNC, NIN, NHIDDEN, NOUT, AW2, AB2)
|
|
wolffd@0
|
6 % generates a vector MASK that selects only the output layer weights.
|
|
wolffd@0
|
7 % This is because most uses of RBF networks in a Bayesian context have
|
|
wolffd@0
|
8 % fixed basis functions with the output layer as the only adjustable
|
|
wolffd@0
|
9 % parameters. In particular, the Neuroscale output error function is
|
|
wolffd@0
|
10 % designed to work only with this mask.
|
|
wolffd@0
|
11 %
|
|
wolffd@0
|
12 % The return value PRIOR is a data structure, with fields PRIOR.ALPHA
|
|
wolffd@0
|
13 % and PRIOR.INDEX, which specifies a Gaussian prior distribution for
|
|
wolffd@0
|
14 % the network weights in an RBF network. The parameters AW2 and AB2 are
|
|
wolffd@0
|
15 % all scalars and represent the regularization coefficients for two
|
|
wolffd@0
|
16 % groups of parameters in the network corresponding to second-layer
|
|
wolffd@0
|
17 % weights, and second-layer biases respectively. Then PRIOR.ALPHA
|
|
wolffd@0
|
18 % represents a column vector of length 2 containing the parameters, and
|
|
wolffd@0
|
19 % PRIOR.INDEX is a matrix specifying which weights belong in each
|
|
wolffd@0
|
20 % group. Each column has one element for each weight in the matrix,
|
|
wolffd@0
|
21 % using the standard ordering as defined in RBFPAK, and each element is
|
|
wolffd@0
|
22 % 1 or 0 according to whether the weight is a member of the
|
|
wolffd@0
|
23 % corresponding group or not.
|
|
wolffd@0
|
24 %
|
|
wolffd@0
|
25 % See also
|
|
wolffd@0
|
26 % RBF, RBFERR, RBFGRAD, EVIDENCE
|
|
wolffd@0
|
27 %
|
|
wolffd@0
|
28
|
|
wolffd@0
|
29 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
30
|
|
wolffd@0
|
31 nwts_layer2 = nout + (nhidden *nout);
|
|
wolffd@0
|
32 switch rbfunc
|
|
wolffd@0
|
33 case 'gaussian'
|
|
wolffd@0
|
34 nwts_layer1 = nin*nhidden + nhidden;
|
|
wolffd@0
|
35 case {'tps', 'r4logr'}
|
|
wolffd@0
|
36 nwts_layer1 = nin*nhidden;
|
|
wolffd@0
|
37 otherwise
|
|
wolffd@0
|
38 error('Undefined activation function');
|
|
wolffd@0
|
39 end
|
|
wolffd@0
|
40 nwts = nwts_layer1 + nwts_layer2;
|
|
wolffd@0
|
41
|
|
wolffd@0
|
42 % Make a mask only for output layer
|
|
wolffd@0
|
43 mask = [zeros(nwts_layer1, 1); ones(nwts_layer2, 1)];
|
|
wolffd@0
|
44
|
|
wolffd@0
|
45 if nargout > 1
|
|
wolffd@0
|
46 % Construct prior
|
|
wolffd@0
|
47 indx = zeros(nwts, 2);
|
|
wolffd@0
|
48 mark2 = nwts_layer1 + (nhidden * nout);
|
|
wolffd@0
|
49 indx(nwts_layer1 + 1:mark2, 1) = ones(nhidden * nout, 1);
|
|
wolffd@0
|
50 indx(mark2 + 1:nwts, 2) = ones(nout, 1);
|
|
wolffd@0
|
51
|
|
wolffd@0
|
52 prior.index = indx;
|
|
wolffd@0
|
53 prior.alpha = [aw2, ab2]';
|
|
wolffd@0
|
54 end |
