BookmarkSubscribeRSS Feed
WWD
Obsidian | Level 7 WWD
Obsidian | Level 7

I have a question that falls within the following area:

 

Course = AI and Machine Learning Professional

Module= Machine Learning Specialist

Lession = Lesson 5 Support Vector Machines

Subsection = Improving the SVM model

Video time = 21-second mark

In the video, the narrator mentions that fact that there is no iterating the SVM model.  The statistician may make changes to the hyperparameters, but the model in completely determined and that there is not a sequence of models (that statement is vague and probably why I'm asking this question.).

 

After running the SVM node, under the Assessment tab and looking at the Output table, the reader will notice that there is a report on the number of iterations used to build the model.

 

My question is "Does the number of iteration pertain to the number of steps that were required to find a solution to the minimization problem that must be solved to identify H, where H = {<w,x> + b = 0}?

2 REPLIES 2
JThompson
SAS Employee

There are "iterations" in the optimization process for SVMs, but what the video means, is that they are not the same as the iterations of models that are developed during the process for neural networks.  For neural networks the objective function is an "error function", which is minimized.  Each step during the process to find the minimum error is a model in its own right.  For SVMs, the objective function being maximized is based on the maximum margin hyperplane. The solution is done by solving an optimization problem given constraints (using the Lagrange approach). This is the key difference to neural nets; the constraints.  There are iterations in solving this optimization problem, but each step is not an actual model.  A model is not arrived at until the optimization problem is solved. So to answer your question, yes, the number of iterations describes the number of steps in this optimization process.  Hope this helps!

WWD
Obsidian | Level 7 WWD
Obsidian | Level 7

JThompson:

 

Thank you for helping me get this point straightened out.

 

Bill Donaldson