net = mdn(nin, nhidden, ncentres, dimtarget) net = mdn(nin, nhidden, ncentres, dimtarget, mixtype, ... prior, beta)
net = mdn(nin, nhidden, ncentres, dimtarget) takes the number of
inputs,
hidden units for a 2-layer feed-forward
network and the number of centres and target dimension for the
mixture model whose parameters are set from the outputs of the neural network.
The fifth argument mixtype is used to define the type of mixture
model. (Currently there is only one type supported: a mixture of Gaussians with
a single covariance parameter for each component.) For this model,
the mixture coefficients are computed from a group of softmax outputs,
the centres are equal to a group of linear outputs, and the variances are
obtained by applying the exponential function to a third group of outputs.
The network is initialised by a call to mlp, and the arguments
prior, and beta have the same role as for that function.
Weight initialisation uses 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.
A specialised data structure (rather than gmm)
is used for the mixture model outputs to improve
the efficiency of error and gradient calculations in network training.
The fields are described in mdnfwd where they are set up.
The fields in net are
type = 'mdn' nin = number of input variables nout = dimension of target space (not number of network outputs) nwts = total number of weights and biases mdnmixes = data structure for mixture model output mlp = data structure for MLP network
net = mdn(2, 4, 3, 1, 'spherical');This creates a Mixture Density Network with 2 inputs and 4 hidden units. The mixture model has 3 components and the target space has dimension 1.
mdnfwd, mdnerr, mdn2gmm, mdngrad, mdnpak, mdnunpak, mlpCopyright (c) Ian T Nabney (1996-9)
David J Evans (1998)