wolffd@0: function net = netinit(net, prior)
wolffd@0: %NETINIT Initialise the weights in a network.
wolffd@0: %
wolffd@0: %	Description
wolffd@0: %
wolffd@0: %	NET = NETINIT(NET, PRIOR) takes a network data structure NET and sets
wolffd@0: %	the weights and biases by sampling from a Gaussian distribution. If
wolffd@0: %	PRIOR is a scalar, then all of the parameters (weights and biases)
wolffd@0: %	are sampled from a single isotropic Gaussian with inverse variance
wolffd@0: %	equal to PRIOR. If PRIOR is a data structure of the kind generated by
wolffd@0: %	MLPPRIOR, then the parameters are sampled from multiple Gaussians
wolffd@0: %	according to their groupings (defined by the INDEX field) with
wolffd@0: %	corresponding variances (defined by the ALPHA field).
wolffd@0: %
wolffd@0: %	See also
wolffd@0: %	MLPPRIOR, NETUNPAK, RBFPRIOR
wolffd@0: %
wolffd@0: 
wolffd@0: %	Copyright (c) Ian T Nabney (1996-2001)
wolffd@0: 
wolffd@0: if isstruct(prior)
wolffd@0:     if (isfield(net, 'mask'))
wolffd@0: 	if find(sum(prior.index, 2)) ~= find(net.mask)
wolffd@0: 	    error('Index does not match mask');
wolffd@0: 	end
wolffd@0: 	sig = sqrt(prior.index*prior.alpha);
wolffd@0: 	% Weights corresponding to zeros in mask will not be used anyway
wolffd@0: 	% Set their priors to one to avoid division by zero
wolffd@0: 	sig = sig + (sig == 0);  
wolffd@0: 	sig = 1./sqrt(sig);
wolffd@0:     else
wolffd@0: 	sig = 1./sqrt(prior.index*prior.alpha);
wolffd@0:     end
wolffd@0:     w = sig'.*randn(1, net.nwts); 
wolffd@0: elseif size(prior) == [1 1]
wolffd@0:   w = randn(1, net.nwts).*sqrt(1/prior);
wolffd@0: else
wolffd@0:   error('prior must be a scalar or a structure');
wolffd@0: end  
wolffd@0: 
wolffd@0: if (isfield(net, 'mask'))
wolffd@0:     w = w(logical(net.mask));
wolffd@0: end
wolffd@0: net = netunpak(net, w);
wolffd@0: