Mercurial > hg > camir-aes2014

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

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 function net = rbf(nin, nhidden, nout, rbfunc, outfunc, prior, beta)
wolffd@0 2 %RBF Creates an RBF network with specified architecture
wolffd@0 3 %
wolffd@0 4 % Description
wolffd@0 5 % NET = RBF(NIN, NHIDDEN, NOUT, RBFUNC) constructs and initialises a
wolffd@0 6 % radial basis function network returning a data structure NET. The
wolffd@0 7 % weights are all initialised with a zero mean, unit variance normal
wolffd@0 8 % distribution, with the exception of the variances, which are set to
wolffd@0 9 % one. This makes use of the Matlab function RANDN and so the seed for
wolffd@0 10 % the random weight initialization can be set using RANDN('STATE', S)
wolffd@0 11 % where S is the seed value. The activation functions are defined in
wolffd@0 12 % terms of the distance between the data point and the corresponding
wolffd@0 13 % centre. Note that the functions are computed to a convenient
wolffd@0 14 % constant multiple: for example, the Gaussian is not normalised.
wolffd@0 15 % (Normalisation is not needed as the function outputs are linearly
wolffd@0 16 % combined in the next layer.)
wolffd@0 17 %
wolffd@0 18 % The fields in NET are
wolffd@0 19 % type = 'rbf'
wolffd@0 20 % nin = number of inputs
wolffd@0 21 % nhidden = number of hidden units
wolffd@0 22 % nout = number of outputs
wolffd@0 23 % nwts = total number of weights and biases
wolffd@0 24 % actfn = string defining hidden unit activation function:
wolffd@0 25 % 'gaussian' for a radially symmetric Gaussian function.
wolffd@0 26 % 'tps' for r^2 log r, the thin plate spline function.
wolffd@0 27 % 'r4logr' for r^4 log r.
wolffd@0 28 % outfn = string defining output error function:
wolffd@0 29 % 'linear' for linear outputs (default) and SoS error.
wolffd@0 30 % 'neuroscale' for Sammon stress measure.
wolffd@0 31 % c = centres
wolffd@0 32 % wi = squared widths (null for rlogr and tps)
wolffd@0 33 % w2 = second layer weight matrix
wolffd@0 34 % b2 = second layer bias vector
wolffd@0 35 %
wolffd@0 36 % NET = RBF(NIN, NHIDDEN, NOUT, RBFUND, OUTFUNC) allows the user to
wolffd@0 37 % specify the type of error function to be used. The field OUTFN is
wolffd@0 38 % set to the value of this string. Linear outputs (for regression
wolffd@0 39 % problems) and Neuroscale outputs (for topographic mappings) are
wolffd@0 40 % supported.
wolffd@0 41 %
wolffd@0 42 % NET = RBF(NIN, NHIDDEN, NOUT, RBFUNC, OUTFUNC, PRIOR, BETA), in which
wolffd@0 43 % PRIOR is a scalar, allows the field NET.ALPHA in the data structure
wolffd@0 44 % NET to be set, corresponding to a zero-mean isotropic Gaussian prior
wolffd@0 45 % with inverse variance with value PRIOR. Alternatively, PRIOR can
wolffd@0 46 % consist of a data structure with fields ALPHA and INDEX, allowing
wolffd@0 47 % individual Gaussian priors to be set over groups of weights in the
wolffd@0 48 % network. Here ALPHA is a column vector in which each element
wolffd@0 49 % corresponds to a separate group of weights, which need not be
wolffd@0 50 % mutually exclusive. The membership of the groups is defined by the
wolffd@0 51 % matrix INDX in which the columns correspond to the elements of ALPHA.
wolffd@0 52 % Each column has one element for each weight in the matrix, in the
wolffd@0 53 % order defined by the function RBFPAK, and each element is 1 or 0
wolffd@0 54 % according to whether the weight is a member of the corresponding
wolffd@0 55 % group or not. A utility function RBFPRIOR is provided to help in
wolffd@0 56 % setting up the PRIOR data structure.
wolffd@0 57 %
wolffd@0 58 % NET = RBF(NIN, NHIDDEN, NOUT, FUNC, PRIOR, BETA) also sets the
wolffd@0 59 % additional field NET.BETA in the data structure NET, where beta
wolffd@0 60 % corresponds to the inverse noise variance.
wolffd@0 61 %
wolffd@0 62 % See also
wolffd@0 63 % RBFERR, RBFFWD, RBFGRAD, RBFPAK, RBFTRAIN, RBFUNPAK
wolffd@0 64 %
wolffd@0 65
wolffd@0 66 % Copyright (c) Ian T Nabney (1996-2001)
wolffd@0 67
wolffd@0 68 net.type = 'rbf';
wolffd@0 69 net.nin = nin;
wolffd@0 70 net.nhidden = nhidden;
wolffd@0 71 net.nout = nout;
wolffd@0 72
wolffd@0 73 % Check that function is an allowed type
wolffd@0 74 actfns = {'gaussian', 'tps', 'r4logr'};
wolffd@0 75 outfns = {'linear', 'neuroscale'};
wolffd@0 76 if (strcmp(rbfunc, actfns)) == 0
wolffd@0 77 error('Undefined activation function.')
wolffd@0 78 else
wolffd@0 79 net.actfn = rbfunc;
wolffd@0 80 end
wolffd@0 81 if nargin <= 4
wolffd@0 82 net.outfn = outfns{1};
wolffd@0 83 elseif (strcmp(outfunc, outfns) == 0)
wolffd@0 84 error('Undefined output function.')
wolffd@0 85 else
wolffd@0 86 net.outfn = outfunc;
wolffd@0 87 end
wolffd@0 88
wolffd@0 89 % Assume each function has a centre and a single width parameter, and that
wolffd@0 90 % hidden layer to output weights include a bias. Only the Gaussian function
wolffd@0 91 % requires a width
wolffd@0 92 net.nwts = nin*nhidden + (nhidden + 1)*nout;
wolffd@0 93 if strcmp(rbfunc, 'gaussian')
wolffd@0 94 % Extra weights for width parameters
wolffd@0 95 net.nwts = net.nwts + nhidden;
wolffd@0 96 end
wolffd@0 97
wolffd@0 98 if nargin > 5
wolffd@0 99 if isstruct(prior)
wolffd@0 100 net.alpha = prior.alpha;
wolffd@0 101 net.index = prior.index;
wolffd@0 102 elseif size(prior) == [1 1]
wolffd@0 103 net.alpha = prior;
wolffd@0 104 else
wolffd@0 105 error('prior must be a scalar or a structure');
wolffd@0 106 end
wolffd@0 107 if nargin > 6
wolffd@0 108 net.beta = beta;
wolffd@0 109 end
wolffd@0 110 end
wolffd@0 111
wolffd@0 112 w = randn(1, net.nwts);
wolffd@0 113 net = rbfunpak(net, w);
wolffd@0 114
wolffd@0 115 % Make widths equal to one
wolffd@0 116 if strcmp(rbfunc, 'gaussian')
wolffd@0 117 net.wi = ones(1, nhidden);
wolffd@0 118 end
wolffd@0 119
wolffd@0 120 if strcmp(net.outfn, 'neuroscale')
wolffd@0 121 net.mask = rbfprior(rbfunc, nin, nhidden, nout);
wolffd@0 122 end
wolffd@0 123