comparison toolboxes/FullBNT-1.0.7/netlab3.3/rbfjacob.m @ 0:e9a9cd732c1e tip

first hg version after svn

+function jac = rbfjacob(net, x)
+%RBFJACOB Evaluate derivatives of RBF network outputs with respect to inputs.
+%
+%	Description
+%	G = RBFJACOB(NET, X) takes a network data structure NET and a matrix
+%	of input vectors X and returns a three-index matrix G whose I, J, K
+%	element contains the derivative of network output K with respect to
+%	input parameter J for input pattern I.
+%
+%	See also
+%	RBF, RBFGRAD, RBFBKP
+%
+
+%	Copyright (c) Ian T Nabney (1996-2001)
+
+% Check arguments for consistency
+errstring = consist(net, 'rbf', x);
+if ~isempty(errstring);
+  error(errstring);
+end
+
+if ~strcmp(net.outfn, 'linear')
+  error('Function only implemented for linear outputs')
+end
+
+[y, z, n2] = rbffwd(net, x);
+
+ndata = size(x, 1);
+jac = zeros(ndata, net.nin, net.nout);
+Psi = zeros(net.nin, net.nhidden);
+% Calculate derivative of activations wrt n2
+switch net.actfn
+case 'gaussian'
+  dz = -z./(ones(ndata, 1)*net.wi);
+case 'tps'
+  dz = 2*(1 + log(n2+(n2==0)));
+case 'r4logr'
+  dz = 2*(n2.*(1+2.*log(n2+(n2==0))));
+otherwise
+   error(['Unknown activation function ', net.actfn]);
+end
+
+% Ignore biases as they cannot affect Jacobian
+for n = 1:ndata
+  Psi = (ones(net.nin, 1)*dz(n, :)).* ...
+    (x(n, :)'*ones(1, net.nhidden) - net.c');
+  % Now compute the Jacobian
+  jac(n, :, :) =  Psi * net.w2;
+end