aiacademy: 深度學習 CNN 老師來啦!!!!

嚴格來說，我們對人腦視覺領域，還是不知道怎麼運作的… Lol

老師講古!~~~~

• 代表性架構
  • AlexNet
    • 8 layers: 5 cnn, 3 dnn
    • ReLu (Rectified Linear Units)
    • AlexNet: Data augmentation and dropout
    • 沒有公開 source code
  • VGG-Net
    • 有公開 source code
    • 16 layers or 19 layers
    • multiple scale training imporves the performance
  • googleNet
    • 22 layers
    • significantly more accurate than AlexNet
    • 12 times lesser parameters than AlexNet
    • inception module

    • Auxiliary classifier

    • GoogleNer: Inception module

      • Choose filter sizes of 1 * 1, 3 * 3, and 5 * 5
      • Concatenate all features maps
  • ResNet(2016)
    • 太多層: gradient exploding/vanishing
    • Residulal learning: Convergence

• Object detection: YOLO

• Semantic Segmentation

  • FCN (Fully Convolutional Netwotks for Semantic Segmentation)

Approach

• Super resolution

• Style Transfer

• Action and gestrue recognition

  • UCF Sports 101 Dataset