CS231n

Course notes of CS231n, Convolutional Neural Networks for Visual Recognition.

2. Image Classification pipeline

• Challenges
  • Viewpoint variation 多视角
  • Illumination 光线
  • Deformation 变形
  • Occlusion 遮挡
  • Background Clutter 杂乱的背景
  • Intraclass variation 内类多样性
• Nearest Neighbor
  • Compute the distances of testing images and training images
  • \(L_n\) norm
    • \((\sum_{i}{|X^i|^n})^{\frac{1}{n}}\)
  • \(L_n\) distance
    • \((\sum_{i}{|X_1^i - X_2^i|^n})^{\frac{1}{n}}\)
  • method
    • For each testing image, find the nearest training image
    • Use the label of finding training image as the prediction
  • complexity
    • Training: \(O(1)\)
    • Testing: \(O(n)\)

• K-Nearest Neighbor
  • Find k nearest t raining images to the testing image
  • Use the most voted label
  • K-Nearest Neighbors Demo
  • Never used on images
    • very slow when testing
    • Distance metrics on pixels are not infomative
• Dataset spliting
  • train, test
  • train, validation, test
  • Cross-Validation
    • folds
    • each fold as validation and average the results
    • Useful for small datasets, but not used too frequently in deep learning
  • errorbar, violin plot https://matplotlib.org/gallery/index.html#statistics
• Linear Classifier
  • \(f(\mathbf x, W) = W\mathbf x+b\)
  • have hard cases
    • xor
    • circles