How to choose the right hyperparameters?

Previous : Purpose of the hyperparameters Next : Choice of the estimator

How to choose the right hyperparameters?

We can sometimes choose nearly correct

and

in an empirical way,
to resore edges avoiding to obtain a too noisy image.


Original image	Corrupted image		=0.3 =5 best result (empirical choice)

The following pictures illustrate the difficulty of a manual hyperparameter choice :


=2 =1 oversmoothed		=0.02 =5 insufficient regularization

=0.3 =0.5 too small threshold		=0.3 =50 too high threshold

An automatic choice of the and hyperparameters is necessary!

André Jalobeanu - 24 Aug 1998