lttr2.m

书籍“Regularization tools for training large feed-forward neural networks using Automatic Differentiat

function [w1,b1,w2,b2,tr,rq] = lttr2(w1,b1,f1,w2,b2,f2,...
					xc,P,T,VA,VAT,TE,TET,TP)
%
%LTTR2 Trains a large feed-forward network with one hidden layer
%using a truncated Gauss-Newton method on a Tikhonov regularized problem.
%
%The general design of LTTR2 is much the same as the design of the 
%functions in the Neural Network (NN) Toolbox from MathWorks even though the
%calling sequences differ. See Neural Network Toolbox User's Guide from the
%MathWorks Inc.
%
%LTTR2 is intended for problems so large that explicit storing of
%the corresponding Jacobian matrix is impossible or inconvenient.
%For smaller problems where the Jacobian can be stored in main memory, there
%is a function TTR2 that uses the Gauss-Newton method (not truncated)
%on a Tikhonov regularized problem. However, even for smaller problems LTTR2
%is often to be prefered to TTR2 (not to mention the Levenberg-Marquardt
%implementations in the NN Toolbox) since it is normally much faster.
%
%You can find more details in two papers by Eriksson J., Gulliksson M.,
%Lindstr鰉 P. and Wedin P-