LFD Book Forum (http://book.caltech.edu/bookforum/index.php)
-   Chapter 4 - Overfitting (http://book.caltech.edu/bookforum/forumdisplay.php?f=111)
-   -   Is It Safe For Neophytes To Go Out And Use Regularization? (http://book.caltech.edu/bookforum/showthread.php?t=1559)

 munchkin 09-16-2012 07:26 PM

Is It Safe For Neophytes To Go Out And Use Regularization?

Careful reading of the book and working on the related final problems has left me concerned about this. If regularization is more art than science and what works depends on the data and the application of the data (which sounds like domain expertise to me) how can a neophyte go out and know that weight decay is the right regularization technique to use for their field? The book has already made it clear that it doesn't work for everything. How will we know?

Also, are the three techniques presented ( hard-order constraint, soft-order constraint and augmented error) mutually exclusive for a given learning task? What I mean by this is that cross validation is supposed to provide the optimum lamda which corresponds to the optimum C value for soft-order constraints. If I have a set of optimum regularized weights delivered by CV can I still switch between different regularizers (low-order, uniform, etc.) when comparing the augmented error of that lamda value? Was this augmented error versus lamda comparison supposed to be part of the cross validation process?

 magdon 09-17-2012 06:20 PM

Re: Is It Safe For Neophytes To Go Out And Use Regularization?

In general, you never know what the optimal regularizer is. The lesson of the story of regularization is that you need some regularization to dampen the effects of noise (deterministic and stochastic).

So you are usually safe using the uniform regularizer. If you don't have any regularization, it can be a disaster. If you use some regularization (choosing the reg. parameter with CV) you will be ok, but not necessarily optimal. The main goal of regularization is to avoid disaster of heavy overfitting.

Quote:
 Originally Posted by munchkin (Post 5371) Careful reading of the book and working on the related final problems has left me concerned about this. If regularization is more art than science and what works depends on the data and the application of the data (which sounds like domain expertise to me) how can a neophyte go out and know that weight decay is the right regularization technique to use for their field? The book has already made it clear that it doesn't work for everything. How will we know? Also, are the three techniques presented ( hard-order constraint, soft-order constraint and augmented error) mutually exclusive for a given learning task? What I mean by this is that cross validation is supposed to provide the optimum lamda which corresponds to the optimum C value for soft-order constraints. If I have a set of optimum regularized weights delivered by CV can I still switch between different regularizers (low-order, uniform, etc.) when comparing the augmented error of that lamda value? Was this augmented error versus lamda comparison supposed to be part of the cross validation process? Thanks for your attention.

 munchkin 09-18-2012 09:29 AM

Re: Is It Safe For Neophytes To Go Out And Use Regularization?