matlab - Regularization in Feed-forward Neural Network -
i have gone through on-line open course lectures andrew ng in coursera. @ end of lectures regarding neural networks, explained reguralization afraid missed something. reguralization, value of cost function calculated follows:
j(theta) = -1/m * jvalmain + lambda/(2*m)*jvalreg
jvalmain
set of sums on y
, , output of nn. second component jvalreg
apply reguralization , looks this:
jvalreg = lambda/(2*m)*sum( sum( sum( theta(j)(i)(k)^2 ) ) )
theta
set of weights, m
number of elements/cases in database , lambda
. lambda? scalar or vector or matrix? how apply reguralization via lambda? lambda regulate particular jth , ith weight lth layer or regulate weights 1 number. somehow confuses me. if familiar concept, grateful help. cheers!
lambda
regularization parameter in estimation. think of means control bias in estimate. scalar , used prevent on fitting of data. here few lines taken notes of coursera assignments.
... value of lambda can affect results of regularized polynomial regression on training , cross validation set. in particular, model without regularization (lambda = 0) fits training set well, not generalize. conversely, model regularization (lambda = 100) not fit training set , testing set well. choice of lambda (e.g., lambda = 1) can provide fit data.
Comments
Post a Comment