The Convolution Layer

There are some CNN-specific terms, such as padding and stride. The data that flows through each layer in a CNN is data with shape (such as three-dimensional data), unlike in previous fully connected networks. Therefore, you may feel that CNNs are difficult when you learn about them for the first time. Here, we will look at the mechanism of the convolution layer used in CNNs.

Issues with the Fully Connected Layer

The fully connected neural networks that we have seen so far used fully connected layers (Affine layers). In a fully connected layer, all the neurons in the adjacent layer are connected, and the number of outputs can be determined arbitrarily.

The issue with a fully connected layer, though, is that the shape of the data is ignored. For example, when the input data is an image, it usually has a three-dimensional shape, determined by the height, the width, and the channel dimension. However, three-dimensional data must be converted into one...