matlab - What does the value of the gradient during NN training signify? -
the default value gradient descent approach 1e-5.
is small value generalization testing set? range should keep in?
does gradient signify error between targets , predicted class during training (i.e using training data)?
if you're not using regularization, should check several values learning rate , several values number of iterations. should on on hold out set (also called validation set). if you're using regularization should not , instead try several values weight of regularization term (usually c or lambda).
as values people try 2^-10 2^-1. in general useful if feature values in reasonable numerical range (from -1 1) or (0 1).
Comments
Post a Comment