Research relevance: Neural networks are now widely used in marketing research, in radio and hydrolocation, in control systems, in decision-making systems, in expert systems, and many other areas.
The prospects for the development of neurocomputing are wide-ranging. A person, who has once successfully applied neural network technology and obtained a positive result, will certainly tend to apply neural networks in his work further, realizing their advantages over other variants. Those who have not yet encountered neural networks will inevitably do, since neurocomputing is already becoming a massively used science. The use of neural networks in the military is very promising, but the use of neural networks in household appliances is also very active.
On the other hand, we can already see the introduction of neural computers in household appliances – LG air conditioners with a built-in neural network smart control unit, Samsung washing machines with a fuzzy logic chip inside, Panasonic household video cameras with neuro-fuzzy sharpening system, and finally Microsoft research on creating a neural network speech recognition system for future operating systems. All this shows that neurocomputing is gaining a firmer foothold in our daily lives. I believe that there is not enough research dedicated to the study. Therefore, I consider my research project relevant.
CHAPTER 1. MAIN BODY
1.1 What is a Neurocomputer
Neurocomputers are systems, in which the algorithm of problem solution is represented by a logic network of private elements – neurons with complete rejection of Boolean elements like AND, OR, NOT. As a consequence, specific connections between elements are introduced, which are the subject of a separate consideration. Unlike, classical methods of problem-solving, neurocomputers implement algorithms of problem-solving, represented in the form of neural networks. This, limitation allows to develop algorithms, potentially more parallel than any of their physical implementations.
Neural network topics are interdisciplinary, which has led to considerable variation in common terminological approaches. Neural network topics are dealt with by computer system developers and programmers as well as by medical professionals, financial and economic specialists, chemists, physicists, etc. (i.e. everyone who is not lazy). What is clear to a physicist is totally unacceptable to a medical professional, and vice versa – all this has given rise to numerous disputes, and entire terminology wars in various fields of application of everything with prefix neuro-.
1.2 The history of neurocomputers
Neurocomputers are the computers of new generation, which qualitatively differ from other classes of parallel computing systems, because they use for problem-solving not the pre-designed algorithms, but specially selected examples, which they learn from. Their appearance is due to objective reasons: the development of the element base, which allows to implement a personal computer – a fully functional computer (neuron model) on a single board, and the need to solve important practical tasks set by the reality. Attempts to create brain-modeling computers were made by neural cybernetics specialists as early as in the 40’s. They sought to develop self-organizing systems capable of learning intelligent behavior in the process of interacting with the world around them, and the components of their systems were usually models of nerve cells. However, the emergence of computer science and related sciences, especially mathematical logic and automata theory, had a strong influence on the field of brain-related research at the same time.
1.3 Advantages of neurocomputers
Compared to conventional computers, neurocomputers have several advantages.
Firstly – high performance, due to the fact that neuroinformatics algorithms have a high degree of parallelism.
Secondly – neurosystems are made very resistant to disturbances and destruction.
Thirdly – stable and reliable neurosystems can be created from unreliable elements, having significant variability of parameters.
1.4 Disadvantages of neurocomputers.
In spite of the above-mentioned advantages, these devices have a number of disadvantages:
1. they are created specifically to solve specific problems related to nonlinear logic and self-organization theory. The solution of such problems on conventional computers is possible only by numerical methods.
2. Because of their uniqueness these devices are quite expensive.
1.5 Practical application of neurocomputers.
Despite their disadvantages, neurocomputers can be successfully used in various spheres of national economy.
– Real-time control of: aircraft, rockets and technological processes of continuous production (metallurgical, chemical, etc.);
– Recognition of images: human faces, letters and hieroglyphs, radar and sonar signals, fingerprints in forensics, diseases by symptoms (in medicine) and locations to look for minerals (in geology, by indirect signs);
– Predicting: the weather, stock prices (and other financial indicators), the outcome of medical treatment, political events (in particular the results of elections), the behavior of adversaries in military conflict and economic competition;
– Optimization and search for the best options: in designing technical devices, choosing an economic strategy and in treating the patient.
This list could go on, but even this is enough to understand that neurocomputers can take their rightful place in modern society.
Neurocomputers belong to the class of MOCMD – multiple streams of single commands – multiple data streams or computing systems with parallel streams of the same commands and multiple data streams.
1.4 Modern neurocomputers
Years of effort by many research groups have resulted in the accumulation of a large number of different “learning rules” and neural network architectures, their hardware implementations, and techniques for using neural networks to solve application problems.
These intelligent inventions exist in the form of a “zoo” of neural networks. Each network from the zoo has its own architecture, training rule and solves a specific set of problems. During the last decade serious efforts have been made to standardize structural elements and transform this “zoo” into a “technopark”]: each neural network from the zoo is implemented on an ideal universal neurocomputer, which has a predefined structure.
Main rules of allocation of functional components of ideal neurocomputer (by Mirkes):
1. Relative functional separateness: each component has a clear set of functions. Its interaction with other components can be described in the form of a small number of queries.
2. Interchangeability of different implementations of any component without changing other components.
Gradually, a market of neurocomputers is taking shape. At present time various highly-parallelized neuroaccelerators (coprocessors) for different tasks are widespread. There are few models of universal neurocomputers on the market, partly because most of them are implemented for special applications. Examples of neurocomputers are the Synapse neurocomputer (Siemens, Germany), the NeuroMatrix processor. A specialized scientific-technical journal Neurocomputers: Design, Application” is being published. Annual conferences on neurocomputers are held. Technically, neurocomputers today are computing systems with parallel streams of identical instructions and multiple data streams (MSIMD-architecture). This is one of the main directions of development of computational systems with massively parallelism.
The artificial neural network can be transferred from (neural) computer to (neural) computer, just like a computer program. Moreover, specialized fast analogue devices can be created on its basis. There are several levels of alienation of a neural network from a universal (neuro)computer: from a network which learns on a universal device, and uses rich possibilities in manipulating a task, learning algorithms and architecture modification, up to full alienation without possibilities of learning and modification, only a functioning trained network.
One of the ways to prepare a neural network for transfer is its verbalization: a trained neural network is minimized with retention of useful skills. The description of the minimized network is more compact and often allows an intelligible interpretation.
1.5 Applications of Neural Computers
1. Real-time control, including: airplanes and missiles; continuous production processes (in power engineering, metallurgy, etc.); hybrid (electric gasoline) car engine; pneumatic cylinder; welding machine; electric furnace; turbo generator.
2. Image recognition: images, human faces, letters and hieroglyphs, fingerprints in forensics, speech, radar and sonar signals; elementary particles and physical processes with them (gas pedal experiments or cosmic ray observation); diseases by symptoms (in medicine); locations to look for minerals (in geology, by indirect signs); danger signs in security systems; chemical compound properties by structure (in chemoinformatics)
3. Real-time forecasting: weather; stock price (and other financial indicators); treatment outcome; political events (election results, international relations, etc.); enemy behavior (real or potential) in military conflict and economic competition; stability of marital relationships.
4. Optimization – search for the best options: when designing technical devices; when choosing an economic strategy; when selecting a team (from company employees to athletes and participants of polar expeditions); when treating a patient.
5. Signal processing in the presence of large noises.
6. Prosthetics (“smart prostheses”) and strengthening of natural functions, including by direct connection of a human’s nervous system with computers (Neuro-computer interface).
8. Telecommunication fraud, its detection and prevention by means of neural network technologies – according to some specialists is one of the most promising technologies in the field of information protection in telecommunication networks.
9. Information security