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