Appendix

Go to a chapter.

Every chapter in the book. Pick a chapter or jump straight to one of its modules.

• Chapter 1Why not a dense layer?Flatten an image and watch the parameter count explode.

  1. 1.1Images are numbers
  2. 1.2Dense layers
  3. 1.3Counting parameters
  4. 1.4Lost neighbors
• Chapter 2Convolution by handSlide a kernel and compute each output as a dot product.

  1. 2.1One output
  2. 2.2Sliding the kernel
  3. 2.3What kernels find
  4. 2.4Moving the duck
• Chapter 3Stride, padding and dilationThree settings that decide where the kernel lands and how big the output is.

  1. 3.1Stride
  2. 3.2Padding
  3. 3.3Dilation
  4. 3.4Putting it together
• Chapter 4ChannelsWhy kernels are really three-dimensional.

  1. 4.1Color images
  2. 4.2Kernels with depth
  3. 4.3Many kernels
  4. 4.41×1 convolutions
• Chapter 5PoolingMax and average pooling, and which input wins.

  1. 5.1Max pooling
  2. 5.2Average pooling
  3. 5.3Small shifts
  4. 5.4Pool or stride?
• Chapter 6Receptive fieldsStack convolutions and pooling layers, and watch how much of the image one output can see.

  1. 6.1One layer
  2. 6.2Stacking layers
  3. 6.3Stride and pooling
  4. 6.4Not every pixel counts
• Chapter 7Backprop through a convolutionWhere the gradients go.

  1. 7.1Loss and gradients
  2. 7.2Gradients for the weights
  3. 7.3Gradients for the input
  4. 7.4Learning a kernel
• Chapter 8Upsampling and transposed convolutionSpreading values back out, and checkerboard artifacts.

  1. 8.1Making maps bigger
  2. 8.2Transposed convolution
  3. 8.3Why “transposed”
  4. 8.4Checkerboard artifacts
• Chapter 9What the network learnedA real CNN running in the browser, layer by layer.

  1. 9.1A trained network
  2. 9.2The first layer
  3. 9.3The second layer
  4. 9.4Where it fails