Artificial Neural Network is computers architected after human brains. It is a mathematical model to all the computational model, basically an information processing paradigm which works more like our biological nervous system. These are electronic models which are based on our brain neural structure.

ANN consists of interconnected artificial neurons and simple processing units, wired together in a complex network of communication.

Also, this system inherits the properties of the biological neurons by interconnected artificial neurons. Each unit or node represents a simplified version of the real neuron model. These units send off a new signal or fire if it recieves an input signal.

The engineers have designed the configuration of the ANN in such a way that it can solve artificial intelligence problems without creating a model of a real biological information system.

Working Of Artificial Neural Network

Neural Networks are constructed in a 3-D world from microscopic components, biologically but integrated circuits using current technology are 2-D devices with various layers for interconnection. Therefore, the types and scope of ANN are restrained to be implemented in silicon. Neural networks are clusters of original artificial neurons. These clusters are the layers which are interconnected artfully.

Artificial Neural network
A simple neural network diagram

Topology of artificial neural network mainly consists of three different layers namely –

Input layer

Under this layer, neurons interface with the real world to receive its inputs. External sources like CSV file or a web service help in loading the input from the world.

At least one input layer should be present in a network. Basically, the input layer receives the input, performs the calculations using its neurons and then transmits the output onto the next layers.

In an input layer, the number of neurons depends on the shape of your training set. Traditionally, the number of neurons is equal to the number of training set feature + 1. The additional node captures the bias term.

Output layer

This consists of neurons which provide the real world with the output of the network. It produces the final result. At least one output layer should be present in a neural network. The output layer receives the input from the previous layer, performs the calculations using its neurons and then computes the output.

In output layer, the number of neurons depends upon the kind of problem you are working on. If the neural network is a regressor or a classifier, the output layer has only one node. If a neural network is using probabilistic activation function then the output layer has one node for every class label in the model.

Hidden layer

This layer comprises of rest of the hidden neurons. Hidden layer makes the neural network stand out of the league in the field of machine learning.

Hidden layers are present between the input layers and the output layers. This is the part of the neural network which consists of private information. Mostly, each hidden layer consists of the same number of neurons. The larger the number of hidden layers, more time the neural network will take to compute the output.

In a hidden layer, the system calculates the optimum number of neurons with the formula given below.

Number of neurons = Trading Data Samples / (Factor*(Input neurons + Output neurons))

Factor prevents over fitting and lies between 1 and 10.

ANN is easily designed by creating these layers. So, basically ANN comprises of the connection between the layers which consists of neurons, their summation and the transfer functions.

The lines connecting each layer are the glue to the entire network as these connections provide a variable strength to the input. These connections are of two types – The first connection excites the neuron and the second one inhibits the neurons.

Layer Inhibition & Feedback

artificial neural network
feedback
competetion
A Simple ANN with Feedback & Competition

In a network, Lateral Inhibition is when a neuron inhibits the other neurons in the same layer. This is mainly used in the output layer. This concept is also known to be competition. Feedback is another type of connection in which the output of one layer goes back to a previous layer.

The connection between neurons impacts the operation of the network significantly.

Training an Artificial Neural Network

After having a proper structure of a neural network, the structure needs to be trained. The training begins after choosing the initial weights.

Two Approaches to Training

  • Supervised

In this system a supervisor feds input and output to the network. Then, after processing of input, we get an output. This output compared with the desired output.

In case, the comparing of the output gives an error, it is fed into the system. After this, system adjusts the weight accordingly to control the whole network.

The data set during the training period is known to be a training set. And, the system processes the same training set several times.

Nowadays, various tools are available in the market. These tools monitor how well an ANN is converging the ability to come to the right conclusion. The training processes go on for more than three days. And, tool monitors the ANN until system obtains the proper statistics.

At times, when the network lacks data, it isn’t able to learn anything and therefore designer has to make modifications in the inputs and outputs, the number of elements per layer, the summation and the transfer functions.   

  • Unsupervised

Under this type of training, their is no supervisor. The system gives only input and no output. Therefore, the system needs to decide the feature to use to predict the right answer. This is known as self-adaption. Unsupervised training is not easy to understand and therefore it use is rare.

Advantages of ANN

  • Self-Organisation

During the learning or training period, an ANN is capable of creating its own organisation of data.

  • High Performance

ANN is better than classical statistical modelling. Also, it is capable of building models which more reflective of the information structure in very less time.

  • Flexibility

Neural network adapts the environment conditions easily. Even though it takes an indignantly large time to learn a drastic change, they are good at adapting frequently changing data.

  • Adaptive Learning

Based on the data given during the training period, ANN can learn to do various tasks.

  • Real Time Operation

The system carries out ANN in parallel which makes them capable of operating in real-time.

Ekta Singh

Ekta Singh

Ekta Singh is an Aerospace Engineer. She is the Founder of The Enigmatic Creation. She loves to read and talk about books. She loves to write and aspires to be an author and an entrepreneur.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Close

Connect With Us


Subscribe to our newsletter today to receive updates of book reviews, technology and wake-up calls.

You have successfully subscribed to the newsletter

There was an error while trying to send your request. Please try again.

The Enigmatic Creation will use the information you provide on this form to be in touch with you and to provide updates and marketing.
%d bloggers like this: