2022-04-02
  • Recap
    • Convolution parameters: input size, stride, kernel size, padding
    • Key points: size, filter, feature map
  • Pooling
    • CNN components: convolution + nonlinear transform + pooling
    • pooling: aggregate multiple values into one
      • types of pooling: max pooling, average pooling
    • pooling as downsampling
  • Variants of CNN
    • LeNet-5
    • ImageNet challenge
    • AlexNet: 巨大的进步，kernel visualization
    • VGG16 (visual geometry group): 16 convolution layers
    • GoogleLeNet: 同一层不同size的filters
  • 模型与数据量要匹配