dnetEstep.m

function model = dnetEstep(model, Ypred)
  
% DNETESTEP Do an E-step (update importance weights) on an Density Network model.
% FORMAT
% DESC updates the importance weights (or posterior responsibilities) for
% a Density Network model.
% ARG model : the model which is to be updated.
% RETURN model : the model with updated weights.
%
% FORMAT
% DESC updates the importance weights (or posterior responsibilities) for
% a Density Network model.
% ARG model : the model which is to be updated.
% ARG ypred : model predictions at the mixture component centres.
% RETURN model : the model with updated weights.
%
% SEEALSO : dnetCreate, dnetMstep
%
% COPYRIGHT : Neil D. Lawrence, 2008

% MLTOOLS

  diffVal = zeros(model.N, model.M);
  if nargin < 2
    Ypred = dnetOut(model, model.X_u);
  end
  if model.N > model.M
    for k = 1:model.M
      diffY = model.y - repmat(Ypred(k, :), model.N, 1);
      diffVal(:, k) = -0.5*sum(diffY.*diffY, 2)*model.beta;
    end
  else
    for i = 1:model.N
      diffY = repmat(model.y(i, :), model.M, 1) - Ypred;
      diffVal(i, :) = -0.5*sum(diffY.*diffY, 2)'*model.beta;
    end
  end
  diffVal = diffVal - repmat(max(diffVal, [], 2), 1, model.M);
  w = exp(diffVal);
  w = w./repmat(sum(w, 2), 1, model.M);
  model.w = sparse(w);
  model.X = model.w*model.X_u;
end