createnet_alpha.m

Nonlinear dynamical factor analysis Matlab package

function net = createnet_alpha(n_in, n_hid, n_out, nonlin, ...
			 w1std, b1std, w2std, b2std, malpha, valpha)
% CREATENET_ALPHA  Create an MLP network with one hidden layer and alphas
%                  for all variables
%
%    Usage:
%      net = createnet(n_in, n_hid, n_out, nonlin,
%                      w1std, b1std, w2std, b2std, malpha, valpha)
%
%      where idim, hdim and odim are numbers of input, hidden
%      and output neurons, respectively.  Nonlin specifies the
%      name of activation function for the hidden layer (default: tanh).
%      The network is initialized randomly and -std parameters
%      specify standard deviations for all values (defaults: 2,4,1,1).
%      Malpha and valpha specify initial values for mean and variance
%      alphas, respectively.

% Copyright (C) 2002 Harri Valpola and Antti Honkela.
%
% This package comes with ABSOLUTELY NO WARRANTY; for details
% see License.txt in the program package.  This is free software,
% and you are welcome to redistribute it under certain conditions;
% see License.txt for details.

if nargin < 5
  w1std = 2;
  b1std = 4;
  w2std = 1;
  b2std = 1;
end
if nargin < 9
  malpha = .1;
  valpha = .1;
end
varcoef = .001;

net.w1 = probdist_alpha(w1std * randn(n_hid, n_in), ...
			varcoef * ones(n_hid, n_in), ...
			malpha * ones(n_hid, n_in), ...
			valpha * ones(n_hid, n_in));
net.b1 = probdist_alpha(b1std * randn(n_hid, 1), ...
			varcoef * ones(n_hid, 1), ...
			malpha * ones(n_hid, 1), ...
			valpha * ones(n_hid, 1));
net.w2 = probdist_alpha(w2std * randn(n_out, n_hid), ...
			varcoef * ones(n_out, n_hid), ...
			malpha * ones(n_out, n_hid), ...
			valpha * ones(n_out, n_hid));
net.b2 = probdist_alpha(b2std * randn(n_out, 1), ...
			varcoef * ones(n_out, 1), ...
			malpha * ones(n_out, 1), ...
			valpha * ones(n_out, 1));

if nargin > 3
  net.nonlin = nonlin;
else
  net.nonlin = 'tanh';
end