|
Which machine learning algorithms (besides SVM's) use the structural risk minimization principle? |
|
If I'm not mistaken, any of the lp (e.g. l2 or l1 or l0) regularization algorithms constitute SRM, because they nest the hypothesis space based upon the model complexity. +1 Could you elaborate a bit more on these regularization algorithms and the differences between l2, l1 and l0?
(May 23 '12 at 07:18)
classifire
1
@John: L_p spaces dictate how errors are penalized. The higher p is, the more outliers contribute to the error.
(May 23 '12 at 18:39)
Emre S
1
l2 is easy to optimize. It gives good generalization. l1 is trickier to optimize. It gives good generalization and sparse models. l0 is very hard to optimize. It gives sparse models. I am not sure how well it generalizes in practice.
(May 28 '12 at 17:47)
Joseph Turian ♦♦
|
