清華大學資訊科學系

CS4601: 人工智慧導論 (Introduction to AI)

Pattern Recognition


Major Steps:
  1. Feature extraction
      In general, feature extraction is a difficult problem. For instance, how do you find the distance between two eyes in a person's portrait (as an digital image)? If feature extraction can be done correctly, then the following process of pattern recognition can be easy.
  2. Data preprocessing
    • Data normalization
      • Gaussian Normalization
      • Unit Normalization
    • Data reduction
      • Editing: To remove noisy (inconsistent) data
          Randomly select a point and find its nearest neighbor. If these two points belong to different classes, then they are inconsistent data pairs and one of them has to be removed.
      • Condensing: To remove redundant (deeply embedded) data
          Randomly select a point and find its nearest neighbor. If these two points belong to the same class, then they are redundant data points and one of them has to be removed.
  3. Data classification
    • K nearest neighbor rule (KNNR)
      • When K is 1, the decision boundary can be obtained via the Voronoi diagram.
      • A large K leads to a more robust classifier.
      • K is usually odd (3 or 5, etc) for two-class problems to avoid the situation of a tie.
      • Distance metrics:
        Minkowski distance is a generalized distance metrics. It has the following metrics as special cases.
        • Maximum distance
        • Euclidean distance
        • City block distance (also known as Manhattan metric or taxicab distance)
    • Kernel and window estimators
    • Adaptive decision boundaries
    • Adaptive discriminant functions
    • One-layer perceptron approach
    • Minimum squared error discriminant functions


This page is maintained by Jyh-Shing Roger Jang. Comments and suggestions are welcome: jang@cs.nthu.edu.tw.