The MLP is the most powerful feedforward neural network. It consists of several layers, where each neuron receives its copy of all the output from the previous layer of neurons. This model is ideal for certain types of tasks, for example, training on a limited number of more or less unstructured parameters.

Nevertheless, let's see what happens to the number of parameters (weights) in such a model when raw data is used as input. For example, the CIFAR-10 dataset contains 32 x 32 x 3 color images, and if we consider each channel of each pixel as an independent input parameter for MLP, each neuron in the first hidden layer adds about 3,000 new parameters to the model! With the increase in image size, the situation quickly gets out of hand, producing images that users can't use for real applications.

One popular solution is to lower the resolution of the images so that MLP becomes applicable. Nevertheless, when we lower the resolution, we risk losing a large amount of information. It would be great if it were possible to process the information before applying a decrease in quality so that we don't cause an explosive increase in the number of model parameters. There is a very effective way to solve this problem, which is based on the convolution operation.