Artificial Neural Networks are the learning algorithms which simulate biological networks of our human brain. They are one of the most effective learning algorithms currently known. They provide a robust approach for approximating :

• real-valued
• discrete-valued
• vector-valued target functions.

Biological Motivation :

Aritificial neural networks are inspired by biological learning systems which are built of very complex system of interconnected neurons. ANN’s might not mimic the biological neural networks in all the cases, but they are the motivation behind design of ANN’s. Although biological neural networks communicate with low speed(10^-3 seconds in case of biological neural networks and 10^-10 seconds in case of ANN’s ), but they are proved to be much efficient in terms of complex decision making than that of ANN’s. Therefore, ANN’s are run on parallel and distributed systems when they are designed for application specific purposes for more efficiency.

What exactly a neural network looks like?

Neural networks are made up of nodes called neurons which mimic the biological neurons in our human brain. Essentially, a neural network consists of an input layer, an output layer and ‘n’ hidden layers ( n is variable depending on the application ).

In the above image, the vector X [X1,X2,….Xn] corresponds to the input and the vector Y [ Y1,Y2,….,Yn ] corresponds to the output. The number of input nodes and output nodes vary with the application. Nodes in adjacent layers are connected through a link. A weight is associated with each link that determines the contribution of input value to that of the output. Each node is associated with an activation function that has a set of input values and an output value(s). We will learn more about weights and activation functions in much detail in upcoming posts.

When to use Neural Networks?

ANN’s are well suited for problems containing complex or huge data, sensory information etc. ANN’s are immune to noise in most of the cases as it involves huge amount of learning while training a neural network. So problems with noisy data would also suit ANN’s well. Some characteristics of problems which suit Neural Networks are discussed below :

• Instances of data containing many attributes or attribute-value pairs.
• Training samples containing errors or noise.
• Ability of humans to understand the target function is not important. Eg: Object Detection.

Basic types of Artificial Neural Networks

• Feed Forward Neural Networks.
• Recurrent Neural Networks.
• Convolution Neural Networks.

There are several such neural networks but these are some of the mainly used neural network architectures. We will discuss about ANN’s in much detail in my next post. Until then cheers..🤞