Machine Learning Team

1
Machine LearningTeam 4

• T.S. Kwon
• K.M. Park
• E.Y. Jang
• Do, D.M. Quan

2
Handwritten numeric recognition

• Introduction
• We used support vector machine
• N-class SVM is needed
• Because there are 10 classes 0, 1, 2,, 9
• Feature extraction
• Aspect ratio adaptive normalization

3
N-class SVM

• One-to-one voting
• Independent SVMs for every pair of classes
• We need number of SVMs.
• e.g. SVM(number 1 vs number 2),
• SVM(number 1 vs number 3),
• The most winner is selected

4
Aspect ratio adaptive normalization

• C.-L. Liu, M. Koga, H. Sato, H. Fujisawa, Aspect
  ratio adaptive normalization for handwritten
  character recognition, ICMI 2000

Original aspect ratio is preserved Normalized
ignoring aspect ratio Normalized adaptively
using aspect ratio
5
Aspect ratio adaptive normalization

• An aspect ratio mapping function

6
Results

• Trained SVM using 10000 data for each number
• Feature vector size is 6x8 for each input image
• Parameters
• Svmtrain -s 1 -n 0.1 -t 2 -g 0.5 -e 0.00001
• Accuracy
• 96.81

7
Face image compression

• Design
• One 8x8 hidden layer
• The output of each hidden node will be quantized
  into 2bit
• 8x8x2128
• Idea
• Give very good initial solution to NN
• Gaussian resizing kernel as NN weight

8
Results
2Bit quantized
Hidden layer
Input images
Output images
9
Thank you

• Any question?