What Is Artificial Neural Networks?

In this tutorial you will learn:

• What is neural networks?
• How do neural networks work?
• Artificial neural network models
• Single-layer neural networks
• Multi layer neural networks
• Feedforward neural networks
• Feedback neural networks
• Activation functions

What is neural networks?

Artificial neural networks are just one of the artificial intelligence techniques developed by imitating the working structure of the human brain. The human brain can process huge amounts of information easily that get from its senses. This information are processed by the neurons in the human brain. These neurons use electrical signals to passed information. The following image shows the structure of a neuron:

The neurons have three main components.

• Dendrites : These are Entry points that take input from other neurons in the network in form of electrical impulses.
• Cell Body : It processes the inputs from other neurons and decides whether the information will pass or not.
• Axon terminals: They transmit the output sent by the cell body to the next neuron. The transmits become in form of electrical signals.

How do neural networks work?

Artificial neural networks work similarly to biological neurons. ANNs perform mathematical operations to generate output from takes inputs. The neurons multiply the inputs with the input’s weights and sum with bias. Each neuron performs a simple operation called threshold to decide if the output will fire or not. If the output value over the threshold value it will fire but if it below the threshold value it will not fire. The purpose of ANN is to obtain the most appropriate weights to be applied to all inputs. The following image shows how to apply mathematical operations.

ANN’s architecture consists of the following concepts.

• Input layer
• Hidden layer
• Output layer
• Weights
• Bias
• Activation functions

Features of Artificial Neural Networks

The artificial neural networks have the following features.

• Nonlinearity
• Parallel Operation
• Learning
• Generalization
• Fault tolerance and flexibility
• Working with missing data

Artificial Neural Network Models

Artificial neural network models can be analyzed in four groups as single-layer networks, multilayer networks, feed-forward artificial neural networks, and feed-backward artificial neural networks.

Single Layer Neural Networks

Single-layer neural networks have only input and output layers. In single-layer neural networks, the output is linear and takes values of 1 or -1. If the output is 1 it is accepted, if it is -1 it is not accepted.

Multi Layer Neural Networks

As the name suggests multi-layer neural networks have many layers. An input layer, an output layer, and a few hidden layers. Each layer neurons is fully connected with the next layer neurons. The following image shows the multilayer neural networks.

Feedforward Artificial Neural Networks

In feedforward neural networks, neurons are in the form of regular layers from entry to exit. There is only a bond from one layer to the layers after it. The information coming to the entrance of the artificial neural network is transmitted to the middle point, in other words to the cells in the hidden layer, without being changed. It then passes through the output layer, respectively, and transferred to the external environment.

Feedback Artificial Neural Networks

In feedback artificial neural networks, unlike feedforward networks, the output of a neuron is not only input to the next neuron layer. It can be connected as an input to any neuron in its previous layer or its own layer. With this structure, feedback artificial neural networks display a nonlinear dynamic behavior. According to the connection type of the connections that give the feedback feature; Feedback artificial neural networks with different behavior and structure can be obtained with the same artificial neural network.

Activation Functions

In the activation function, the artificial neural cell works on the input data and obtains the corresponding net output results. This function is generally not linear. It is important to choose the function correctly. Because the result will affect performance. The function can be in single and double polarity.

Some activation functions: Sigmoid, Tanh, Relu, Leaky Relu, Softmax, Swish