12. PyTorch Tutorial---Convolutional Neural Network (CNN)

Deep learning is a branch of machine learning and is considered an important step taken by researchers in recent decades. Examples of deep learning implementations include applications such as image recognition and speech recognition.

Two important types of deep neural networks are as follows −

Convolutional Neural Network
Recurrent Neural Network
In this chapter, we will focus on the first type, the convolutional neural network (CNN).

Convolutional Neural Networks
Convolutional neural networks are designed to process data through multiple layers of arrays. This type of neural network is used in applications such as image recognition or face recognition.

The main difference between CNN and any other ordinary neural network is that CNN takes the input as a two-dimensional array and operates directly on the image instead of focusing on feature extraction that other neural networks focus on.

The main method of CNN involves solving recognition problems. Top companies like Google and Facebook have invested in research and development of recognition projects to complete tasks faster.

Every convolutional neural network consists of three basic concepts −

1. Local response field
2. Convolution
3. Pooling
Let us understand these terms in detail.

Local response field
CNN exploits the spatial correlation present in the input data. Each connection in a parallel layer of a neural network consists of certain input neurons. This specific area is called the local response field. It only focuses on hidden neurons. Hidden neurons will process input data within a specific field without being aware of changes outside specific boundaries.

The diagram for generating local response fields is shown below −

Convolution
In the image above, we observe that each connection learns the weight associated with the hidden neuron and is connected accordingly as it moves from one layer to another. Here, individual neurons are constantly on the move. This process is called "convolution".

The connection map from the input layer to the hidden feature map is defined as the "shared weight", and the included bias is called the "shared bias".

Pooling
Convolutional neural network uses pooling