diff toolboxes/FullBNT-1.0.7/netlab3.3/rbf.m @ 0:e9a9cd732c1e tip

first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
line wrap: on
line diff
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/toolboxes/FullBNT-1.0.7/netlab3.3/rbf.m	Tue Feb 10 15:05:51 2015 +0000
@@ -0,0 +1,123 @@
+function net = rbf(nin, nhidden, nout, rbfunc, outfunc, prior, beta)
+%RBF	Creates an RBF network with specified architecture
+%
+%	Description
+%	NET = RBF(NIN, NHIDDEN, NOUT, RBFUNC) constructs and initialises a
+%	radial basis function network returning a data structure NET. The
+%	weights are all initialised with a zero mean, unit variance normal
+%	distribution, with the exception of the variances, which are set to
+%	one. This makes use of the Matlab function RANDN and so the seed for
+%	the random weight initialization can be  set using RANDN('STATE', S)
+%	where S is the seed value. The activation functions are defined in
+%	terms of the distance between the data point and the corresponding
+%	centre.  Note that the functions are computed to a convenient
+%	constant multiple: for example, the Gaussian is not normalised.
+%	(Normalisation is not needed as the function outputs are linearly
+%	combined in the next layer.)
+%
+%	The fields in NET are
+%	  type = 'rbf'
+%	  nin = number of inputs
+%	  nhidden = number of hidden units
+%	  nout = number of outputs
+%	  nwts = total number of weights and biases
+%	  actfn = string defining hidden unit activation function:
+%	    'gaussian' for a radially symmetric Gaussian function.
+%	    'tps' for r^2 log r, the thin plate spline function.
+%	    'r4logr' for r^4 log r.
+%	  outfn = string defining output error function:
+%	    'linear' for linear outputs (default) and SoS error.
+%	    'neuroscale' for Sammon stress measure.
+%	  c = centres
+%	  wi = squared widths (null for rlogr and tps)
+%	  w2 = second layer weight matrix
+%	  b2 = second layer bias vector
+%
+%	NET = RBF(NIN, NHIDDEN, NOUT, RBFUND, OUTFUNC) allows the user to
+%	specify the type of error function to be used.  The field OUTFN is
+%	set to the value of this string.  Linear outputs (for regression
+%	problems) and Neuroscale outputs (for topographic mappings) are
+%	supported.
+%
+%	NET = RBF(NIN, NHIDDEN, NOUT, RBFUNC, OUTFUNC, PRIOR, BETA), in which
+%	PRIOR is a scalar, allows the field NET.ALPHA in the data structure
+%	NET to be set, corresponding to a zero-mean isotropic Gaussian prior
+%	with inverse variance with value PRIOR. Alternatively, PRIOR can
+%	consist of a data structure with fields ALPHA and INDEX, allowing
+%	individual Gaussian priors to be set over groups of weights in the
+%	network. Here ALPHA is a column vector in which each element
+%	corresponds to a separate group of weights, which need not be
+%	mutually exclusive.  The membership of the groups is defined by the
+%	matrix INDX in which the columns correspond to the elements of ALPHA.
+%	Each column has one element for each weight in the matrix, in the
+%	order defined by the function RBFPAK, and each element is 1 or 0
+%	according to whether the weight is a member of the corresponding
+%	group or not. A utility function RBFPRIOR is provided to help in
+%	setting up the PRIOR data structure.
+%
+%	NET = RBF(NIN, NHIDDEN, NOUT, FUNC, PRIOR, BETA) also sets the
+%	additional field NET.BETA in the data structure NET, where beta
+%	corresponds to the inverse noise variance.
+%
+%	See also
+%	RBFERR, RBFFWD, RBFGRAD, RBFPAK, RBFTRAIN, RBFUNPAK
+%
+
+%	Copyright (c) Ian T Nabney (1996-2001)
+
+net.type = 'rbf';
+net.nin = nin;
+net.nhidden = nhidden;
+net.nout = nout;
+
+% Check that function is an allowed type
+actfns = {'gaussian', 'tps', 'r4logr'};
+outfns = {'linear', 'neuroscale'};
+if (strcmp(rbfunc, actfns)) == 0
+  error('Undefined activation function.')
+else
+  net.actfn = rbfunc;
+end
+if nargin <= 4
+   net.outfn = outfns{1};
+elseif (strcmp(outfunc, outfns) == 0)
+   error('Undefined output function.')
+else
+   net.outfn = outfunc;
+ end
+
+% Assume each function has a centre and a single width parameter, and that
+% hidden layer to output weights include a bias.  Only the Gaussian function
+% requires a width
+net.nwts = nin*nhidden + (nhidden + 1)*nout;
+if strcmp(rbfunc, 'gaussian')
+  % Extra weights for width parameters
+  net.nwts = net.nwts + nhidden;
+end
+
+if nargin > 5
+  if isstruct(prior)
+    net.alpha = prior.alpha;
+    net.index = prior.index;
+  elseif size(prior) == [1 1]
+    net.alpha = prior;
+  else
+    error('prior must be a scalar or a structure');
+  end  
+  if nargin > 6
+    net.beta = beta;
+  end
+end
+
+w = randn(1, net.nwts);
+net = rbfunpak(net, w);
+
+% Make widths equal to one
+if strcmp(rbfunc, 'gaussian')
+  net.wi = ones(1, nhidden);
+end
+
+if strcmp(net.outfn, 'neuroscale')
+  net.mask = rbfprior(rbfunc, nin, nhidden, nout);
+end
+