Visualizing the learning process with activation-based methods
Before we get into discussing activations, layers, filters, neurons, gradients, convolutions, kernels, and all the fantastic elements that make up a CNN, let's first briefly revisit the mechanics of a CNN and this one in particular.
The convolution layer is the essential building block of a CNN. It convolves the input with learnable filters, which are relatively small but are applied across the entire width, height, and depth at specific distances or strides. See Figure 8.10. In the fruit CNN case, the first convolutional layer has 16 filters with a 2 × 2 kernel, the default 1 × 1 stride, and no zero padding (valid). Each filter produces a two-dimensional activation map (also known as a feature map). It's called an activation map because it denotes positions of activations in the images – in other words, where specific "features" are located. In this context, a feature is an abstract...
