Artificial neural networks - are those which consist of singular elements - neurons.They are mathematical models of biological neurons, i.e. cells that make up the nervous system.
first started talking about neural networks in 1943, but after the invention of Perceptron Rosenblatt came the golden era, and networks have become very popular.However, after the publication of Minsk in 1969, in which a scientist has proved the inefficiency of Perceptron, under certain conditions, the interest in this sector fell sharply.But the story does not end with artificial networks.In 1985, George. Hopfield presented their studies and proved that the neural network - a great tool for machine learning.
was borrowed from biology several concepts and principles.Neuron - a kind of switch that receives and transmits impulses (signals).If the neuron receives a sufficiently powerful momentum, it is believed that it is activated and transmits pulses remaining neurons associated with it.Neuron same that was not activated, remains in a state of rest, the pulse does not pass.Neuron consists of several main components: synapses that connect neurons to each other and receive pulses, axon, whose task is to transmit pulses and the dendrite, which receives signals from various sources.When a neuron receives an impulse above a certain threshold, it immediately sends a signal to the next neurons.
A mathematical model is a little different.Log mathematical model of neuron - a vector which is composed of a large number of components.Each of the component - is one of the pulses which are received by the neuron.The output of the model is a single number.That is, in the model input vector is converted into a scalar, later transferred to other neurons.
Neural networks can be trained in two ways: with and without a teacher.The learning process consists of several steps.Please input the network served a stimulus from the outside.Then, in accordance with the rules of changing the free parameters of the neural network, then the network responds to input stimuli have differently.The process should be repeated as long as the network does not solve the problem.Supervised learning algorithm is that during training the network already has the correct answer.This method has been successfully used for many applications, but it is often criticized for the fact that he is biologically implausible.Neural networks are trained without a teacher in the case where the only known input signals.On this basis, the network gradually learns to give the best value outputs.
Application of neural networks is really diverse.Often they are used for the automation of recognition, forecasting, creation of various expert systems functional approximation.With such a network can perform audio detection or optical signal indicators predict exchange, to create a system capable of self-learning, which can, for example, to synthesize speech from a given text or car park.Neural networks used in the West increasingly, unfortunately, domestic firms have not yet have adopted this technique.
Despite the advantages of ANN on conventional calculations in some areas, the existing neural networks - not a perfect solution.Since they are able to learn, they can be wrong.In addition, it is impossible accurately to ensure that the neural network is designed to be optimal.The developer is required to understand the nature of the problem being solved, have a lot of information that describes the problem, to obtain data for testing and training network, choose the right method of training, the transfer function and the function of the adder.
