Plotting MSE of Validation and Testing Data during training of ANN

Hi,

Is it possible to plot the MSE of the Validation data and Testing Data during the actual training of an ANN? I think I have a problem of overtraining and want to know the best time to manually stop the training. Alternatively is there a method to load the weights etc from an old epoch obtained during training?

Many Thanks

Answer 1

[Greg Heath]

What version of the NN toolbox are you using? The latest versions have

fitnet for regression or curvefitting

patternnet for classification and pattern recognition

The previous versions were newfit and newpr, respectively.

All versions have a default trn/val/tst (0.7/0.15/0.15) random division option for validation stopping(val) and unbiased prediction(tst). Various other combinations of data division types and percent ratios are available via overwriting the defaults.

lookfor divide

Another way to avoid overtraining an overfit net is to make sure the number of output training equations Neq = prod(size(targets)) = O*N is significantly larger than the number of unknown weights Nw = net.numWeightElements = (I+1)*H+(H+1)*O for an I-H-O FFMLP.

Neq >> Nw ==> H << (Neq-O)/(I+O+1)

Hope this helps.

Thank you for officially accepting my answer(;>)

[Sam harris]

Thanks for you help Greg, I'll check the number of output training equations and see how it goes.

thanks,

Sam