problem 5 polynomial kernel

[zhuoc]

Hi,

I'm working on problem 5. _All_ results look fine except using polynomial kernel in dual form, where I can only get around 40% accuracy...

Here, I set C to 1. If I change it to 0.1, I then get 94% accuracy. If I change C to 100, then both accuracy of polynomial and linear kernel (only in dual form) drop. 

Does anyone have an idea what I might get wrong? 

Thanks,

Zhuo

[Junier Oliva]

The point of this exercise is not to get a low training error. And yes, different values of C will result in different classifiers, thus different accuracies. Also, recall that a high training accuracy is not always a good thing since you may be over fitting.

Best,

Junier

--
http://alex.smola.org/teaching/cmu2013-10-701 (course website)
http://www.youtube.com/playlist?list=PLZSO_6-bSqHQmMKwWVvYwKreGu4b4kMU9 (YouTube playlist)
---
You received this message because you are subscribed to the Google Groups "10-701 Spring 2013 CMU" group.
To unsubscribe from this group and stop receiving emails from it, send an email to 10-701-spring-201...@googlegroups.com.
To post to this group, send email to 10-701-spri...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.
 
 

[zhuoc]

Hi Junier,

Thanks for your reply. I know that low training error is not necessarily good. However, my concern, and question, is that I get different results when solving the same problem in primal and dual form. I also set C to be equal. So I guess there should be some implementation problem?

Thanks,

Zhuo

On Tuesday, February 19, 2013 9:08:21 PM UTC-5, Junier wrote:

The point of this exercise is not to get a low training error. And yes, different values of C will result in different classifiers, thus different accuracies. Also, recall that a high training accuracy is not always a good thing since you may be over fitting.

Best,

Junier

On Tue, Feb 19, 2013 at 7:27 PM, zhuoc <czx...@gmail.com> wrote:

Hi,

I'm working on problem 5. _All_ results look fine except using polynomial kernel in dual form, where I can only get around 40% accuracy...

Here, I set C to 1. If I change it to 0.1, I then get 94% accuracy. If I change C to 100, then both accuracy of polynomial and linear kernel (only in dual form) drop. 

Does anyone have an idea what I might get wrong? 

Thanks,

Zhuo

--
http://alex.smola.org/teaching/cmu2013-10-701 (course website)
http://www.youtube.com/playlist?list=PLZSO_6-bSqHQmMKwWVvYwKreGu4b4kMU9 (YouTube playlist)
---
You received this message because you are subscribed to the Google Groups "10-701 Spring 2013 CMU" group.

To unsubscribe from this group and stop receiving emails from it, send an email to 10-701-spring-2013-cmu+unsub...@googlegroups.com.

[zhuoc]

Hi Junier,

Also, the weird thing is that only with polynomial kernel do I see the difference between primal and dual form solution. And the difference changes with different Cs. Do you have any idea about this?

Thanks,

Zhuo

On Tuesday, February 19, 2013 9:08:21 PM UTC-5, Junier wrote:

The point of this exercise is not to get a low training error. And yes, different values of C will result in different classifiers, thus different accuracies. Also, recall that a high training accuracy is not always a good thing since you may be over fitting.

Best,

Junier

On Tue, Feb 19, 2013 at 7:27 PM, zhuoc <czx...@gmail.com> wrote:

Hi,

I'm working on problem 5. _All_ results look fine except using polynomial kernel in dual form, where I can only get around 40% accuracy...

Here, I set C to 1. If I change it to 0.1, I then get 94% accuracy. If I change C to 100, then both accuracy of polynomial and linear kernel (only in dual form) drop. 

Does anyone have an idea what I might get wrong? 

Thanks,

Zhuo

--
http://alex.smola.org/teaching/cmu2013-10-701 (course website)
http://www.youtube.com/playlist?list=PLZSO_6-bSqHQmMKwWVvYwKreGu4b4kMU9 (YouTube playlist)
---
You received this message because you are subscribed to the Google Groups "10-701 Spring 2013 CMU" group.

To unsubscribe from this group and stop receiving emails from it, send an email to 10-701-spring-2013-cmu+unsub...@googlegroups.com.

[Helen]

Hi,
I came across similiar problem before.

It's mainly due to the precision in optimization.

Different optimization toolbox brought significantly different result.
Also, the precision in matlab might also bring problem.

After I tune the iteration control and precision, I can get around 94% accuracy using polynominal kernel in dual problem.