1 Statistical Learning Theory Project Statistical Learning Theory SVM not TOY It’s now 3D!!!...

Post on 17-Jan-2016

220 views 0 download

transcript

1Statistical Learning Theory Project

Statistical Learning Statistical Learning TheoryTheory

Statistical Learning Statistical Learning TheoryTheory

SVM not TOY

It’s now 3D!!!

Members: R93922108 R93546015

R93922140 P92922007 D93922011

2Statistical Learning Theory Project

IntroductionIntroductionIntroductionIntroduction

The work has done by another group last yearhttp://arbor.ee.ntu.edu.tw/~kidd/svm_project/3D_SVM/svm.htm

3Statistical Learning Theory Project

Last YearLast YearLast YearLast Year

4Statistical Learning Theory Project

Project GoalProject GoalProject GoalProject Goal

1. Recover to code.

2. Make it more user friendly.

3. Release to community.

5Statistical Learning Theory Project

JAVA 3D PlatformJAVA 3D PlatformJAVA 3D PlatformJAVA 3D Platform

1. Choose the Java3D platform as the engine of 3D SVMTOY.

2. The same reasons mentioned in previous work.

Easy to combine the LIBSVM java code. Using Java 3D is easier than OpenGL and

DirectX.

6Statistical Learning Theory Project

Functions of 3D SVMTOYFunctions of 3D SVMTOYFunctions of 3D SVMTOYFunctions of 3D SVMTOY

1. Input ( by mouse or keyboard)

2. Change

3. Run

4. Clear

5. View

6. Parameter

7Statistical Learning Theory Project

MethodMethodMethodMethod

We use a brute-force way to testing all points in a 3-D space, and get the approximated bounded hyperplane.

Each dimension has 100 points, we only draw about 20,000 points, much less than 1,000,000 points last year.If the predicted label of the point is different from that of the neighbor points ((i-1,j,k),(i,j-1,k),(i,j,k-1)), then the point will be drawn in the display plane as a point of the bounded hyperplane.

8Statistical Learning Theory Project

DemoDemoDemoDemo

http://www.csie.ntu.edu.tw/~r93546015/blogtoy/blosxom.cgi

9Statistical Learning Theory Project

Running TimeRunning TimeRunning TimeRunning Time

We have tried 30 points on the notebook (Pentium 1.5 G) . 30 points takes 5 seconds.(25 secs last year)100 points takes 12 seconds.(1.5 mins last year)

100 points is probably the maximum number of points that a user would click.The running time is ideal.

10Statistical Learning Theory Project

ConclusionConclusionConclusionConclusion

In the foundation of preliminary work, we have successfully implemented 3D SVMToy

With a friendly input interface.

Show clearly bounded hyperplane of the model.

Runs much faster and more smoothly than last year’s work.