Artificial Neural Networks (ANNs) are simplified models of the biological nervous system. They basically mimic the working of a human brain. An ANN, in general, a highly interconnected network of a large number of processing elements, i.e. Neurons.
An ANN can be consider as a highly parallel network. Distributed processing is a typical feature of a neural network.
Neural networks work on the principle of learning by examples. They are presented with known examples of a problem called ‘Training Set‘, to acquire knowledge about the problem. After training, the network can be effectively employe in solving instances of the problem previously unknown to the network.
Features of Artificial Neural Network :
(1) Adaptive Learning : An ANN possess the ability to learn from the existing environment and also adapt to the new environment.
(2) Self-Organization : An ANN can create its own organization or representation of the information it receives during learning time.
(3) Real-Time Operation : Since ANN computations may be carry out in parallel, they can be use for real time applications. Special hardware devices are being design and manufacture to take advantage of this capability of Artificial Neural Networks and to reduce the response time.
(4) Fault Tolerance : Neural networks are fault tolerant in the sense even if some portions of the neural net is removed (For example some connections are removed), there will be only a small degradations in neural network performance.
(5) Generalization : After learning from the available inputs and their relationships, ANN has capability to infer unseen relationships on unseen data, thus making the model generalize.
(6) Non-Linearity : ANNs have the ability to learn and model non-linear and complex relationships. In most of the Real-life problems, many of the relationships between inputs and outputs are non-linear as well as complex.
(7) Parallel Distributed Processing : ANNs have massive parallelism which makes them very efficient.
Applications of ANN :
(1) Forecasting : Neural network can be use very efficiently in forecasting exchange rates, predicting stock values, inflation and cash forecasting, forecasting weather conditions, etc. Researchers have prove that the forecasting accuracy of Neural Network Systems tend to excel over that of the linear regression model.
(2) Image Compression : Image Compression is a technique that removes some of the redundant information present in the image without affecting its perceptibility. Thus reducing the storage size requires to store the image. Neural Network can be effectively use to compress the image.
(3) Industrial Process Control : Neural Networks have been applied successfully in industrial process control of dynamic systems. Neural Networks are the best choice for modelling non-linear systems and implementing general purpose non-linear controllers. It becomes very useful due to their universal approximation capabilities.
(4) Optical Character Recognition : Well known application using image recognition is the Optical Character Recognition (OCR) tools that are available with the standard scanning software for the home computer. Scansoft has a great success in combining Neural Network for correctly recognizing both characters and words.
(5) Medical Science : Artificial Neural Network (ANN) is currently the next promising area of interest in medical science. It is believe that neural networks will have extensive application to biomedical problems in the next few years. ANN has already been successfully applied in medical applications. Diagnostic systems, bio-chemical analysis, disease detection and image analysis are the Medical fields where the Artificial Neural Networks are Needful.
Disadvantages of ANN :
(1) The best known disadvantage of Neural Networks is their “Black box” nature. This means that you don’t know how and why your neural network came up with a certain output. For example : when you put in an image of a cat into a neural network and it predicts it to be a car. It is very hard to understand what causes it to came up with this prediction.
(2) Neural Networks usually require much more data than traditional Machine Learning Algorithms, as in at least thousands if not millions of labeled samples.
(3) Computationally Expensive : Usually, Neural Networks are also more computationally expensive that traditional algorithms. State of the art deep learning algorithms can take several weeks to train completely from scratch. Most traditional Machine Learning Algorithms take much less time to train.