Convolutional Neural Networks

Intuition. We aim to take advantage of the features of human vision to inspire the architection of computer vision model. Notably human vision has:

• Local connectivity, i.e. near things are related
• Invariant in many transformation, i.e. rotating, stretching, flipping objects doesn’t make them different objects
• the RGB channels of an image are not natural, it’s for human eyes only.

Components §

Convolution Layer §

Intuition. The term comes from a function convolution, where two functions are “combined.” 3b1b from 4:45 shows taking a moving average of a function $f$, which is a type of convolution. The moving average (function $g$) filters the original function $f$. This is denoted as $f\ast g$. In computer vision, we only consider discrete convolution, which is also called a kernel. How Blurs & Filters Work - Computerphile - YouTube visualizes the convolution well. Also see: where the left $x_i$ is the original “image” and the right is the filtered image. $k_{ij}$ is the kernel applied that results in cell $(i,j)$ in the output. If you don’t want the image to get to smaller, you can also zero-pad the image’s border:

Pooling Layer §

Pooling layers don’t have special parameters. They are just a simple kernel that takes either the maximum or the average of the input: Normally, CNNs alternate between convolutional and pooling layers.