All the machine learning/deep learning algorithms you have learned about imply that the type of input data is one-dimensional. When you look at a real-world application, however, data is not necessarily one-dimensional. A typical case is an image. Though we can still convert two-dimensional (or higher-dimensional) data into a one-dimensional array from the standpoint of implementation, it would be better to build a model that can handle two-dimensional data as it is. Otherwise, some information embedded in the data, such as positional relationships, might be lost when flattened to one dimension.
To solve this problem, an algorithm called Convolutional Neural Networks (CNN) was proposed. In CNN, features are extracted from two-dimensional input data through convolutional layers and pooling layers (this will be explained later), and then these features are put into general multi-layer perceptrons. This preprocessing for MLP is inspired by human visual areas and...