wolffd@0: function net = gtm(dim_latent, nlatent, dim_data, ncentres, rbfunc, ...
wolffd@0:    prior)
wolffd@0: %GTM	Create a Generative Topographic Map.
wolffd@0: %
wolffd@0: %	Description
wolffd@0: %
wolffd@0: %	NET = GTM(DIMLATENT, NLATENT, DIMDATA, NCENTRES, RBFUNC), takes the
wolffd@0: %	dimension of the latent space DIMLATENT, the number of data points
wolffd@0: %	sampled in the latent space NLATENT, the dimension of the data space
wolffd@0: %	DIMDATA, the number of centres in the RBF model NCENTRES, the
wolffd@0: %	activation function for the RBF RBFUNC and returns a data structure
wolffd@0: %	NET. The parameters in the RBF and GMM sub-models are set by calls to
wolffd@0: %	the corresponding creation routines RBF and GMM.
wolffd@0: %
wolffd@0: %	The fields in NET are
wolffd@0: %	  type = 'gtm'
wolffd@0: %	  nin = dimension of data space
wolffd@0: %	  dimlatent = dimension of latent space
wolffd@0: %	  rbfnet = RBF network data structure
wolffd@0: %	  gmmnet = GMM data structure
wolffd@0: %	  X = sample of latent points
wolffd@0: %
wolffd@0: %	NET = GTM(DIMLATENT, NLATENT, DIMDATA, NCENTRES, RBFUNC, PRIOR),
wolffd@0: %	sets a Gaussian zero mean prior on the parameters of the RBF model.
wolffd@0: %	PRIOR must be a scalar and represents the inverse variance of the
wolffd@0: %	prior distribution.  This gives rise to a weight decay term in the
wolffd@0: %	error function.
wolffd@0: %
wolffd@0: %	See also
wolffd@0: %	GTMFWD, GTMPOST, RBF, GMM
wolffd@0: %
wolffd@0: 
wolffd@0: %	Copyright (c) Ian T Nabney (1996-2001)
wolffd@0: 
wolffd@0: net.type = 'gtm';
wolffd@0: % Input to functions is data
wolffd@0: net.nin = dim_data;
wolffd@0: net.dim_latent = dim_latent;
wolffd@0: 
wolffd@0: % Default is no regularisation
wolffd@0: if nargin == 5
wolffd@0:    prior = 0.0;
wolffd@0: end
wolffd@0: 
wolffd@0: % Only allow scalar prior
wolffd@0: if isstruct(prior) | size(prior) ~= [1 1]
wolffd@0:    error('Prior must be a scalar');
wolffd@0: end
wolffd@0: 
wolffd@0: % Create RBF network
wolffd@0: net.rbfnet = rbf(dim_latent, ncentres, dim_data, rbfunc, ...
wolffd@0:    'linear', prior);
wolffd@0: 
wolffd@0: % Mask all but output weights
wolffd@0: net.rbfnet.mask = rbfprior(rbfunc, dim_latent, ncentres, dim_data);
wolffd@0: 
wolffd@0: % Create field for GMM output model
wolffd@0: net.gmmnet = gmm(dim_data, nlatent, 'spherical');
wolffd@0: 
wolffd@0: % Create empty latent data sample
wolffd@0: net.X = [];