The learning rate in gradient descent algorithm is being decreased
during training. The decrement is done if the classification accuracy
stops growing (< 0.5%). The accuracy is evaluated on a cross-validation
set very often. But any set can be used for this. For example, it is
better to use a training set if the cross-validation set is
non-representative or too small, not to step out of the "global" minimum
due to a bad estimation of the step length (learning rate).
Petr
Dne 24.1.2013 15:39, Ana Montalvo napsal(a):