|
Daniel@0
|
1 <html>
|
|
Daniel@0
|
2 <head>
|
|
Daniel@0
|
3 <title>
|
|
Daniel@0
|
4 Netlab Reference Manual mlpprior
|
|
Daniel@0
|
5 </title>
|
|
Daniel@0
|
6 </head>
|
|
Daniel@0
|
7 <body>
|
|
Daniel@0
|
8 <H1> mlpprior
|
|
Daniel@0
|
9 </H1>
|
|
Daniel@0
|
10 <h2>
|
|
Daniel@0
|
11 Purpose
|
|
Daniel@0
|
12 </h2>
|
|
Daniel@0
|
13 Create Gaussian prior for mlp.
|
|
Daniel@0
|
14
|
|
Daniel@0
|
15 <p><h2>
|
|
Daniel@0
|
16 Synopsis
|
|
Daniel@0
|
17 </h2>
|
|
Daniel@0
|
18 <PRE>
|
|
Daniel@0
|
19 prior = mlpprior(nin, nhidden, nout, aw1, ab1, aw2, ab2)</PRE>
|
|
Daniel@0
|
20
|
|
Daniel@0
|
21
|
|
Daniel@0
|
22 <p><h2>
|
|
Daniel@0
|
23 Description
|
|
Daniel@0
|
24 </h2>
|
|
Daniel@0
|
25 <CODE>prior = mlpprior(nin, nhidden, nout, aw1, ab1, aw2, ab2)</CODE>
|
|
Daniel@0
|
26 generates a data structure
|
|
Daniel@0
|
27 <CODE>prior</CODE>, with fields <CODE>prior.alpha</CODE> and <CODE>prior.index</CODE>, which
|
|
Daniel@0
|
28 specifies a Gaussian prior distribution for the network weights in a
|
|
Daniel@0
|
29 two-layer feedforward network. Two different cases are possible. In
|
|
Daniel@0
|
30 the first case, <CODE>aw1</CODE>, <CODE>ab1</CODE>, <CODE>aw2</CODE> and <CODE>ab2</CODE> are all
|
|
Daniel@0
|
31 scalars and represent the regularization coefficients for four groups
|
|
Daniel@0
|
32 of parameters in the network corresponding to first-layer weights,
|
|
Daniel@0
|
33 first-layer biases, second-layer weights, and second-layer biases
|
|
Daniel@0
|
34 respectively. Then <CODE>prior.alpha</CODE> represents a column vector of
|
|
Daniel@0
|
35 length 4 containing the parameters, and <CODE>prior.index</CODE> is a matrix
|
|
Daniel@0
|
36 specifying which weights belong in each group. Each column has one
|
|
Daniel@0
|
37 element for each weight in the matrix, using the standard ordering as
|
|
Daniel@0
|
38 defined in <CODE>mlppak</CODE>, and each element is 1 or 0 according to
|
|
Daniel@0
|
39 whether the weight is a member of the corresponding group or not. In
|
|
Daniel@0
|
40 the second case the parameter <CODE>aw1</CODE> is a vector of length equal to
|
|
Daniel@0
|
41 the number of inputs in the network, and the corresponding matrix
|
|
Daniel@0
|
42 <CODE>prior.index</CODE> now partitions the first-layer weights into groups
|
|
Daniel@0
|
43 corresponding to the weights fanning out of each input unit. This
|
|
Daniel@0
|
44 prior is appropriate for the technique of automatic relevance
|
|
Daniel@0
|
45 determination.
|
|
Daniel@0
|
46
|
|
Daniel@0
|
47 <p><h2>
|
|
Daniel@0
|
48 See Also
|
|
Daniel@0
|
49 </h2>
|
|
Daniel@0
|
50 <CODE><a href="mlp.htm">mlp</a></CODE>, <CODE><a href="mlperr.htm">mlperr</a></CODE>, <CODE><a href="mlpgrad.htm">mlpgrad</a></CODE>, <CODE><a href="evidence.htm">evidence</a></CODE><hr>
|
|
Daniel@0
|
51 <b>Pages:</b>
|
|
Daniel@0
|
52 <a href="index.htm">Index</a>
|
|
Daniel@0
|
53 <hr>
|
|
Daniel@0
|
54 <p>Copyright (c) Ian T Nabney (1996-9)
|
|
Daniel@0
|
55
|
|
Daniel@0
|
56
|
|
Daniel@0
|
57 </body>
|
|
Daniel@0
|
58 </html> |
