|
wolffd@0
|
1 function prior = mlpprior(nin, nhidden, nout, aw1, ab1, aw2, ab2)
|
|
wolffd@0
|
2 %MLPPRIOR Create Gaussian prior for mlp.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 % PRIOR = MLPPRIOR(NIN, NHIDDEN, NOUT, AW1, AB1, AW2, AB2) generates a
|
|
wolffd@0
|
6 % data structure PRIOR, with fields PRIOR.ALPHA and PRIOR.INDEX, which
|
|
wolffd@0
|
7 % specifies a Gaussian prior distribution for the network weights in a
|
|
wolffd@0
|
8 % two-layer feedforward network. Two different cases are possible. In
|
|
wolffd@0
|
9 % the first case, AW1, AB1, AW2 and AB2 are all scalars and represent
|
|
wolffd@0
|
10 % the regularization coefficients for four groups of parameters in the
|
|
wolffd@0
|
11 % network corresponding to first-layer weights, first-layer biases,
|
|
wolffd@0
|
12 % second-layer weights, and second-layer biases respectively. Then
|
|
wolffd@0
|
13 % PRIOR.ALPHA represents a column vector of length 4 containing the
|
|
wolffd@0
|
14 % parameters, and PRIOR.INDEX is a matrix specifying which weights
|
|
wolffd@0
|
15 % belong in each group. Each column has one element for each weight in
|
|
wolffd@0
|
16 % the matrix, using the standard ordering as defined in MLPPAK, and
|
|
wolffd@0
|
17 % each element is 1 or 0 according to whether the weight is a member of
|
|
wolffd@0
|
18 % the corresponding group or not. In the second case the parameter AW1
|
|
wolffd@0
|
19 % is a vector of length equal to the number of inputs in the network,
|
|
wolffd@0
|
20 % and the corresponding matrix PRIOR.INDEX now partitions the first-
|
|
wolffd@0
|
21 % layer weights into groups corresponding to the weights fanning out of
|
|
wolffd@0
|
22 % each input unit. This prior is appropriate for the technique of
|
|
wolffd@0
|
23 % automatic relevance determination.
|
|
wolffd@0
|
24 %
|
|
wolffd@0
|
25 % See also
|
|
wolffd@0
|
26 % MLP, MLPERR, MLPGRAD, EVIDENCE
|
|
wolffd@0
|
27 %
|
|
wolffd@0
|
28
|
|
wolffd@0
|
29 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
30
|
|
wolffd@0
|
31 nextra = nhidden + (nhidden + 1)*nout;
|
|
wolffd@0
|
32 nwts = nin*nhidden + nextra;
|
|
wolffd@0
|
33
|
|
wolffd@0
|
34 if size(aw1) == [1,1]
|
|
wolffd@0
|
35
|
|
wolffd@0
|
36 indx = [ones(1, nin*nhidden), zeros(1, nextra)]';
|
|
wolffd@0
|
37
|
|
wolffd@0
|
38 elseif size(aw1) == [1, nin]
|
|
wolffd@0
|
39
|
|
wolffd@0
|
40 indx = kron(ones(nhidden, 1), eye(nin));
|
|
wolffd@0
|
41 indx = [indx; zeros(nextra, nin)];
|
|
wolffd@0
|
42
|
|
wolffd@0
|
43 else
|
|
wolffd@0
|
44
|
|
wolffd@0
|
45 error('Parameter aw1 of invalid dimensions');
|
|
wolffd@0
|
46
|
|
wolffd@0
|
47 end
|
|
wolffd@0
|
48
|
|
wolffd@0
|
49 extra = zeros(nwts, 3);
|
|
wolffd@0
|
50
|
|
wolffd@0
|
51 mark1 = nin*nhidden;
|
|
wolffd@0
|
52 mark2 = mark1 + nhidden;
|
|
wolffd@0
|
53 extra(mark1 + 1:mark2, 1) = ones(nhidden,1);
|
|
wolffd@0
|
54 mark3 = mark2 + nhidden*nout;
|
|
wolffd@0
|
55 extra(mark2 + 1:mark3, 2) = ones(nhidden*nout,1);
|
|
wolffd@0
|
56 mark4 = mark3 + nout;
|
|
wolffd@0
|
57 extra(mark3 + 1:mark4, 3) = ones(nout,1);
|
|
wolffd@0
|
58
|
|
wolffd@0
|
59 indx = [indx, extra];
|
|
wolffd@0
|
60
|
|
wolffd@0
|
61 prior.index = indx;
|
|
wolffd@0
|
62 prior.alpha = [aw1, ab1, aw2, ab2]';
|
