LFD Book Forum

LFD Book Forum (http://book.caltech.edu/bookforum/index.php)
-   Homework 8 (http://book.caltech.edu/bookforum/forumdisplay.php?f=137)
-   -   SVMs versus NNs (http://book.caltech.edu/bookforum/showthread.php?t=1226)

Willem 08-30-2012 07:11 AM

SVMs versus NNs
 
Hi, are there any guidelines to decide when to use a neural network in stead of a SVM? I guess neural networks are better for regression problems, but how about classification? I read somewhere on this forum that logistic regression can sometimes perform better than SVMs when the underlying problem is intrinsically probabilistic. Are there similar rules of thumb for when using neural networks? Or do SVMs generally perform better (or similar)? It would be nice to have some guidelines of when to try which model first (and why). Thanks!

htlin 09-02-2012 04:13 PM

Re: SVMs versus NNs
 
Quote:

Originally Posted by Willem (Post 4633)
Hi, are there any guidelines to decide when to use a neural network in stead of a SVM? I guess neural networks are better for regression problems, but how about classification? I read somewhere on this forum that logistic regression can sometimes perform better than SVMs when the underlying problem is intrinsically probabilistic. Are there similar rules of thumb for when using neural networks? Or do SVMs generally perform better (or similar)? It would be nice to have some guidelines of when to try which model first (and why). Thanks!

In my experience, most of the well-tuned modern machine learning algorithms reach similar performance. The difference is how hard it is to make those algorithms "well-tuned" (without overfitting). I am more familiar with SVMs in my own experience, and I can tune them more easily than NNets --- but Yaser may tell you the other way around because of you-know-what reason. :p So in terms of performance, it is hard to give a guideline.

There are other issues, though. SVMs are notably harder to scale up to large data sets (except for the linear case); NNets can serve the need of "feature extraction" through the neurons (while the feature transformation of SVMs are just too implicit in the dual problem). Those issues often affect my own choice of algorithms.

And yes, SVMs are strong with classification but not so strong with regression. For regression, SVMs usually are not my first-hand choice.

Hope this helps.

SeanV 02-27-2013 06:11 AM

Re: SVMs versus NNs
 
1) the issue with logistic regression is that it is giving you a probability of class A rather than a classification...if you need that then it is (clearly) better . if all you are going to do is threshold the probability at 0.5 and classify then not.

2) I think SVMS are the method of choice unless you have large number of points...ie try that first....the issue is that NNETS basically have no optimisation procedure. You have to find a learning rate that jumps over the troughs and lands in a deep vallley... [imagine trying to descend a mountain range with your eyes closed]

3) I view SVMs as a linear classifier with a nice regularisation.... this is very important in classification, because as you saw in the lectures, there are many different linear classifiers that give you the same classifcation on the training set. Regression is different .. changing the straight line fit changes your MSE. I would say linear regression would be the natural analogue of SVMs ...[ and then you have to choose the regularisation term and nonlinear transformations]


All times are GMT -7. The time now is 08:28 PM.

Powered by vBulletin® Version 3.8.3
Copyright ©2000 - 2020, Jelsoft Enterprises Ltd.
The contents of this forum are to be used ONLY by readers of the Learning From Data book by Yaser S. Abu-Mostafa, Malik Magdon-Ismail, and Hsuan-Tien Lin, and participants in the Learning From Data MOOC by Yaser S. Abu-Mostafa. No part of these contents is to be communicated or made accessible to ANY other person or entity.