Artificial Neural Network (ANN) is a computational technology based on a biological neural model and tries to simulate the behavior and work of the neural model on various inputs. As a computational technology, ANN is an information processing technique that uses a biological neural quantitative model that inspires the creation of a computational process that is identical to the work of neurons in the human nervous system. Like biological neural networks, mathematical models of ANN connect a number of adaptive inputs and outputs of a system that are organized in the processing element layer as well as the relationships between biological neurons.
One of the ANN architecture that is often used to recognize the pattern is the back propagation network (JPB). The back propagation network is a multi-screen network made up of non-linear units. Like a linear unit, the nonlinear unit computes its activation level by summing up all the weighted activations. However, unlike a linear unit, a non-linear unit transforms its activation through a non-linear transfer function. The purpose of the JPB is to study the inconsistency of mapping between input-output pattern pairs. JPBs are usually used as pattern classifiers, or are generally used to solve nonlinear problems.
ANN learned from a sample called training set. Because learning from the sample, ANN has the potential to build a computational system as a result of mapping the input and output relationships that exist in the system. The training set is known as a training pattern in the form of a vector and is derived from sources such as images, voice signals, and other information.
The learning process of ANN is classified into two:
Learning with supervision (supervised learning).
Unattended learning (unsupervised learning).
Learning with supervision (supervised learning), ie the network responds by getting a specific target. Before the network changes its own weight to achieve the target, the interconnected weights are initialized. The process of learning ANN with supervision is a learning process by providing training to achieve a specified target output. ANN gets training to get to know certain patterns. By providing output targets, the input changes will be adapted by output by changing the weights of the interconnection following the specified learning algorithm. The training set is selected from the maximum output function of each state of the modified parameter. By initializing the weight of each cell, the ANN will search for the smallest error, so the output function form closes the desired target.
The learning process is done by doing a set of training. In putting together a set of training, there are a few things to note:
The order of the target pattern
Criteria for error calculation
Criteria of learning process
Number of iterations to go through
Initialize weights and initial parameters
Training is done by pairing input and output patterns. For surveillance purposes, pattern pairs do not need to follow certain formulas. ANN should be able to adapt random inputs to obtain a set of outputs that still follow the target. One of the most popularly studied learning processes used is the learning process using back-propagation network algorithms.
Unattached learning (unsupervised learning), in a networkless learning process does not get a target, so ANN regulates its own interconnection weight. Unattended learning is sometimes referred to as self-organizing learning, or learning to classify a pattern without the need for a training set. In an unattended learning process ANN will classify samples of available input patterns into different groups. Example of ANN with unattended learning one of them is kohonen network.
Thanks @jalora