Crossentropy loss function - What is a good performance goal?

Good Afternoon,

Looking around ANSWER and exploring GOOGLE GROUPS i found this method by Dr. Greg Heath to define a valid training goal for the MSE performance function:

[I,N]=size(x);
[O,N]=size(t);
MSE00a=mean(var(t,0,2));                
Ntrn=floor(0.7*N);                     
Hub=floor((Ntrn-O)/(I+1+O));
MSEgoal=0.01*(Ndof/Ntrneq)*MSE00a;    

And i was wondering if there is a similar method to set a Crossentropy reference goal for neural net performance, since i want to experiment different type of loss functions in order to get the best results.

King Regards,

Answer 1

[Greg Heath]

These equations are not necessarily precise.

For example:

data = design + test

design = training + validation

In particular:

Test subset data should not be used to estimate design parameters.

However, since we typically let the training function randomly perform the trn/val/tst division, the separate train/val/tst subsets are not available before training.

That is why I typically design 10 nets for every trial value for the number of hidden nodes.

Hope this helps

Thank you for formally accepting my answer

Greg