[Google_Bootcamp_Day19]

Convolutional Neural Network 3

1 * 1 convolutions

• Network in network
• Use 1 * 1 convolution when you want to shrink the size of channel
• similar as fully-connected layer

Inception Network

• Idea : let the network learn whatever parameters it wants to use, whatever the combinations of these filter sizes it wants

• Problem : Computational cost
• Solution : using 1 * 1 convolution

Inception Module

Practical Advice

• Use open-source implementations
• Use Transer Learning
  • Freeze pre-trained model and train last few layers that you want to target
  • If dataset = large, freeze few layers then train others
  • If dataset = very large, initialize weights with pre-trained weights, then train the whole pre-trained model
• Data Augmentation
  • Mirroring on the vertical axis
  • Random cropping
  • Color Shifting
  • Rotation
  • Shearing
  • Local Warping

[source] https://www.coursera.org/learn/convolutional-neural-networks