mlpjacob.m

来自「递归贝叶斯估计的工具包」· M 代码 · 共 39 行

function g = mlpjacob(net, x)
%MLPJACOB Backpropagate gradient of error function for 2-layer network.
%
%   Description
%   G = MLPJACOB(NET, X) takes a network data structure NET and an
%   input vector X and returns a matrix G whose J, K
%   element contains the derivative of network output K with respect to
%   input parameter J.
%
%   See also
%   MLP, MLPGRAD, MLPBKP
%

%   Copyright (c) Christopher M Bishop, Ian T Nabney (1996, 1997)
%                 This function coded by Rudolph van der Merwe (2002)

[y, z, a] = mlpfwd(net, x);

switch net.outfn
case 'linear'
  deltas = eye(net.nout);
  delhid = deltas*net.w2';
  delhid = delhid.*repmat((1.0 - z.*z),net.nout,1);
case {'logistic'}
  deltas = diag(y.*(1-y));
  delhid = deltas*net.w2';
  delhid = delhid.*repmat((1.0 - z.*z),net.nout,1);
case {'softmax'}
  deltas = diag(y)-y'*y;
  delhid = deltas*net.w2';
  delhid = delhid.*repmat((1.0 - z.*z),net.nout,1);
otherwise
  error(['Unknown activation function ', net.actfn]);
end

% Finally, evaluate the first-layer gradients.
g = (delhid*net.w1')';