The structure of a convolutional network

The following figure shows the structure of a basic classification CNN:

Figure 4.10 – A basic convolutional network with convolutional, FC, and pooling layers

Most CNNs share basic properties. Here are some of them:

• We would typically alternate one or more convolutional layers with one pooling layer (or a stride convolution). In this way, the convolutional layers can detect features at every level of the receptive field size. The aggregated receptive field size of deeper layers is larger than the ones at the beginning of the network. This allows them to capture more complex features from larger input regions. Let’s illustrate this with an example. Imagine that the network uses 3×3 convolutions with stride = 1 and 2×2 pooling with stride = 2:
  • The units of the first convolutional layer will receive input from 3×3 pixels of the image.
  • A group of 2×2 output units of the first...