Artificial neural networks are those that consist of special elements - neurons. They are a mathematical model of biological neurons, that is, cells that make up the human nervous system.
For the first time, they started talking about neural networks in 1943, and after the invention of the Rosenblatt perceptron, a golden era began, and networks became very popular. However, after the publication of the work of Minsky in 1969, in which the scientist proved the inefficiency of the perceptron under certain conditions, interest in this industry fell sharply. But the story of artificial networks does not end there. In 1985, J. Hopfield presented his research and proved that neural networks are a great tool for machine learning.
Several concepts and principles were borrowed from biology. A neuron is a type of switch that receives and then transmits pulses (signals). If a neuron receives a sufficiently powerful impulse, it is believed that it is activated and transfers impulses to other neurons associated with it. A neuron that remains unactivated remains at rest, it does not transmit an impulse. A neuron consists of several main components: synapses that connect neurons to each other and receive impulses, an axon, whose task is to transmit impulses, and a dendrite, which receives signals from different sources. When a neuron receives an impulse above a certain threshold, it immediately transmits a signal to the following neurons.
The mathematical model is slightly different. The input of a mathematical model of a neuron is a vector that consists of a large number of components. Each of their components is 1 of the pulses that the neuron receives. The output of the model is a single number. That is, inside the model, the input vector is transformed into a scalar, which is then transferred to other neurons.
Neural networks can be trained in two ways: with and without a teacher. The learning process consists of several steps. First, an external stimulus is supplied to the network input. Then, in accordance with the rules, the free parameters of the neural network are changed, after which the network responds to input stimuli in a different way. The process must be repeated until the network solves the problem. The learning algorithm with the teacher is that during training the network already has the correct answer. This method has been successfully used to solve many applied problems, but it is often criticized for being biologically implausible. Neural networks are trained without a teacher when only the input signals are known. On their basis, the network is gradually learning to give better output values.
The use of neural networks is really diverse. Often they are used to automate pattern recognition, forecasting, create various expert systems, and approximate functionalities. With the help of such a network, it is possible to recognize audio or optical signals, predict exchange performance, create self-learning systems that can, for example, synthesize speech using a given text or park a car. Neural networks in the West are used more actively, unfortunately, domestic firms have not yet adopted this technique.
Despite the advantages of ANN over conventional computing in some areas, existing neural networks are not ideal solutions. Since they are capable of learning, they may be wrong. In addition, it is impossible to guarantee that the developed neural network will be optimal. The developer must understand the nature of the problem to be solved, have a lot of information that characterizes the problem, obtain data for testing and training the network, choose the training method, transfer function, and adder functions correctly.