RU2530270C2 - Virtual stream computer system based on information model of artificial neural network and neuron - Google Patents

Abstract

FIELD: physics, computer engineering.

SUBSTANCE: invention relates to computer engineering and intelligent computer systems and can be used to create intelligent computer systems. A virtual stream computer system comprises a unit of autonomously operating processor elements (PE); the number of operating PE in a specific physical time unit is defined by conditions of the computation process algorithm or results thereof from the previous time unit; a two-level virtual active memory unit for each PE and a group of PE, a common virtual active memory unit available for all PE, coordinate systems for information referencing and verifying results of the computation process in the domain of the two-level virtual active memory unit for each PE or group of PE.

EFFECT: high reliability of information performance in circularity, restriction, exchange and uncertainty conditions, enabling implementation of dynamically expanded architecture, high information accuracy of computation and data processing results.

1 dwg

Description

The invention relates to the field of computer technology and intelligent computing systems and can be used to create intelligent computing systems, information processes in which are caused by model and algorithmic isolation, restrictions on the computing environment, energy and information exchange and information uncertainty, development and implementation of new forms of computing based on information model of artificial neural network and neuron and stream computing virtualization technologies Processes, development of artificial and computational intelligence, etc.

The technical result is the expansion of the functional and intellectual capabilities of a computing system, increasing the reliability of their information performance in conditions of isolation, restrictions, exchange and uncertainty, providing neural networks, the possibility of implementing a dynamically expandable architecture for creating super sophisticated intelligent information virtual virtual computing systems, increasing the information reliability of calculation results and data processing. A virtual streaming computing system contains a block of autonomously functioning processor elements (PEs), the number of functioning PEs in a particular quantum of physical time is determined by the conditions of the computational process algorithm or its result from the previous time slot, a two-level virtual active memory block for each PE and PE group, a shared virtual block active memory available for all PE, coordinate system information binding and verification of the results (intermediate and final) computational about the process in the field of two-level virtual active memory for each PE or group of PE.

Known varieties of streaming computing systems are based on the “pure data stream” model: statistical aircraft and dynamic aircraft. The general principles of the organization of computing processes in such aircraft are aimed at achieving a high degree of parallelism and ensuring a high load factor of hardware, which are the main factors determining the efficiency of the aircraft both in a number of technical and algorithmic characteristics, and in general. The development and implementation of computational processes for simulation technologies for various tasks and the processing of information in such systems based on models of sequential and parallel algorithms is carried out on calculations that are performed using a strict logic diagram of a sequence of discrete steps, including accurate data (accuracy is determined by the number of significant digits or characters) . Even with such strict assumptions, an error in just a single bit in the sequence of operands or in the data often makes the result of the calculation useless or blurry. Powerful efforts to eliminate or minimize such errors by hardware and software are mainly associated with solving the dilemma: the accuracy of a computer system and the success of preventing unstable execution due to such accuracy, which are key values in the traditional computing industry. Accuracy and reliability in traditional serial and parallel computing systems are very far from reliable organization of computing processes, synthesis and information processing in intelligent living systems. The question of choosing criteria for evaluating the effectiveness and reliability of algorithms, as well as many others, remains open. Traditional streaming computing systems and computing systems for solving streaming problems are functionally systems with high computing efficiency both in technical characteristics and in algorithmic support. Efficiency is achieved, as noted above, due to the high degree of parallelism of the hardware interaction of its elements and the methodology for the synthesis of parallel circuits and algorithms of computational processes. The intellectual capabilities and intellectual effectiveness of such computing systems as an intelligent tool in the technologies of organization of computing processes, simulation and information processing in order to achieve accurate and stable results when the diffusion of information is localized on a spatio-temporal scale, have not been considered or investigated. Such questions are also open that are related to the development of information models of streaming computing in the conditions of model and algorithmic isolation, information exchange, limitations of the computing environment and information uncertainty.

The causes of errors caused by key quantities (accuracy and instability) of computational processes are based on the fact that the traditionally practiced sequential and stream computing technologies are highly unstable. This instability is largely due to the fact that the technologies of computational processes of streaming and scalar computing are implemented without taking into account the information dynamics of objects of the computing environment, which determines the occurrence of the spatio-temporal and information conditions of the controllability of computational processes. Currently, it is not known for certain how data are presented, stored and processed in living intelligent information processing systems, how computing mechanisms and many other computing aspects are organized in them. The work on the study of the identified secrets previously and currently mainly focused on the study of biochemical, neurophysiological and structural aspects of information processes in living neurosystems. The results of the studies allowed only some limited computational mechanisms to be identified, but no mathematical description of the logical data structures for representing information in such neural systems was obtained, and a general theory of data manipulation and calculations in living neural systems was not built. It has also not been created, which is the most important, acceptable general theory that would describe the computational capabilities of intelligent living systems based on information models of information transmission and processing in neurosystems, which can be adapted for the development and implementation of intelligent computing systems capable of exhibiting similar characteristics of information processing. Nevertheless, it can be argued with a high degree of certainty that living intelligent information processing systems realize their functions, including computational ones, in conditions of isolation, limitation, exchange (energy and information) and uncertainty. The indicated conditions for the functioning of living systems are also characteristic of streaming computing systems, which are currently not implemented and are not described in any model of existing computing systems (real and abstract).

The main goal of the present invention that we are introducing is to provide the possibility of developing and implementing the architecture (structure) of a virtual streaming computing system that overcomes the disadvantages of the prior art, as well as providing the possibility of developing and implementing a logical architecture (structure) for a virtual streaming computing system, the computational advantages of which are in the fields of intelligent information processing systems and computational intelligence, software is possible the implementation of the components of a virtual streaming computing system, bearing the advantages of the information model of the neural network [4] as a unique mathematical, logical and computational structure, which was discovered by the author of the present invention and which can be the fundamental and scientific basis for the implementation of computing processes of virtual streaming calculations of simulation modeling and processing system technologies information.

SUMMARY OF THE INVENTION

The information model of the neural network, as a mathematical and logical prototype of an intelligent computing system, has important computational and informational properties of the mechanisms of computational processes and allows them to be defined and described taking into account mathematical, dynamic, informational and metrological aspects that are not obvious and support the new paradigm of computation and implementation non-traditional forms of computing, taking into account the information dynamics of computing environment objects. This paradigm has new computational advantages based on non-traditional principles of development and implementation of various forms of computing, which allow us to synthesize the effects of parallelism, virtualization and intellectualization on the information dynamics of computing environment objects. The present invention discloses the architecture (structure) of a virtual streaming computing system that allows for the implementation of this new paradigm of computational intelligence and access to these new intelligent computing capabilities.

The objectives of the invention are achieved, on the one hand, on the basis of a method for constructing a virtual streaming computing system, according to which the processor modules and memory modules of such a system logically, functionally and informationally interact with each other according to the same principles as neurons in a network information model of a neural network, defines it organization and implementation in the form of a multilayer modular system for streaming parallel data processing, where there are several layers, including a layer of processor elements (PE), a two-layer layer non-virtual memory for each PE and a processor layer module and a shared active virtual memory layer for the entire processor layer, the processor layer being divided into virtual processor modules and the shared active virtual memory layer into virtual memory modules, processor modules have a network model for organizing information connections and perform the functions of the interface both between the elements of the layers and with the external environment, initialization of the computing system, routing information in the system, configuration the architecture of the virtual streaming computing system for the structure of the information graph of the problem being solved, restoring the results of the computing process when the entire system, individual layers, modules or elements fail, and the modules of the virtual RAM layer perform the functions of storing module attributes, commands, a local coordinate system for information binding and verification the results of (intermediate and final) computing processes in the computing environment [5], the modules of the total active virtual memory functionally logically designed to store global information system of coordinates testing and verification and reflection on it of the results of the computational process at every step, and performs buffer function in a tiling operation between modules layers function virtual ports and information gateways to the outside world.

According to the indicated method, the processor modules are logically and functionally combined and configured into a virtual computing structure with the geometry and connection topology of the network information model of the neural network and each processor module operates in the corresponding area of the information space of the coordinate system for linking and checking the results, which ensures high reliability in the entire computing system and the reliability of the results, and the basic and forming elements of the module a virtual memory and a shared virtual active virtual memory cell are of two types: active and passive virtual cell. Logical and arithmetic operations in processor modules are implemented according to the following principles of computing: the principle of interaction of the operands of the computing environment with the information virtual environment [2, 3], the principle of virtual perspective in the design operator [1], the principle of the connection of the calculation results [5]: fractal and information .

And on the other hand - with the help of the structure of a virtual streaming computing system that implements a new paradigm of computational intelligence based on the network information model of a neural network [4]. The network information model of a neural network, on the one hand, describes the structure (architecture) of a virtual streaming computing system in the form of a global virtual information dynamic network in which the nature of the interaction between elements at the control and information exchange levels is defined as processes with local interaction. On the other hand, an element of such a network, a formal neuron, is defined as a local virtual computing system that combines the resources of processor elements and the memory of a real physical computing system with a rigid or reconfigurable architecture, the logical and physical functions of which are as follows. Firstly, it is a local virtual computer network that defines and defines in a physical environment a real streaming BC computing server-domain, i.e. the space of virtual identifier addresses for its processor elements and corresponding active memory segments. Secondly, the virtual computing server-domain defines not only the area of information definition in the virtual address space of its active memory - its domain, but also determines the general numerical scale for the computing process and the scale division price (numerical accuracy) in the information boundaries of the memory domain address space - its range. Elements of active memory are virtual cells of various types, differing in functional attribute, logical organization of information content.

A neuron in the indicated architecture (structure) of the global virtual streaming computing system (WWS) is defined as its component, the purpose and function of which is as follows: 1) for a parallel streaming task, it implements it in a given area of possible values and a fixed quantum of physical time; 2) the implementation of statistical synchronization of the computing process in the neural network within the selected real-time quantum. Neurons are combined into a virtual computing structure that supports a computational process with a dynamically changing population of active neurons in a fixed and specific quantum of physical time. Active neurons can implement various computational processes within a specific quantum of physical time if the interaction between them is parameterized by the task algorithm and the computing environment.

In the area of the neuron’s active memory, a local segment is set (in read-only mode), in which the coordinate system of the information binding and verification is stored for the calculation results (final and intermediate) [5], burdened by various kinds of errors and information diffusion in computational technologies. The geometric interpretation of such a system is that it is the basic information coordinate grid, which reflects the dynamics of the information processes of the computing environment. The nodes of such a lattice in the broad sense are a figurative and symbolic reflection of the geometric and informational properties of the dynamics of the processes of the computing environment [1, 6]. The information essence of the geometry of nodes on the lattice defines it as a memory segment of virtual passive cells of a fixed length and with information unchanged in them, i.e. this is a segment of constants in the form of symbol chains. Such a segment defines a domain for determining the display of the results of a computational process in the information space of the active virtual memory of a computing environment — their domain and range of values — a range. The introduction of such a logical memory organization and computing process logic defines computing systems with such an architecture as virtual streaming computing systems with variable domain ranges. If the results of a computational process in a neuron’s computing environment go beyond the range, then an interaction occurs with a neighboring neuron, into the information range of which they fall.

Computational processes in the indicated systems are quantized by physical time. The duration of the quanta can be arbitrary and is determined by the conditions of statistical synchronization and the computational process. In each quantum of physical time, the number of active neurons is determined by the result of calculations in the previous quantum and the conditions of the problem.

Virtual cells of the active memory of a neuron that do not belong to the coordinate system of information binding and verification are active cells of variable length with dynamically changing information in them, i.e. these are working virtual cells. The process of interaction of character chains stored in passive and active virtual cells is defined as an information process with local interaction. Any operation between cells of different types is specified in the form of two operands - interaction and design. The first step in the operation is the implementation of the information alignment procedure for character chains of different lengths, which is defined as interaction with the environment. Next, the operation of interaction between the character chains is implemented and the result is formed, which is projected into the node of the base coordinate lattice. The meaning of the design operation is as follows. If the information distance between the symbol chain of the result of the operation and the lattice node is less than the information neighborhood of the node, then the result is in this node [1, 2, 3].

In accordance with the present invention, the structure of a virtual streaming computing system based on information models of the formal neural network and neuron, as a physical platform, can have a streaming computing system (for example, a transputer, etc.) or a multi-conveyor computing system (with a rigid or reconfigurable architecture), from processor elements and a block of multilevel virtual active memory. Such virtual computing systems support three phases of virtualization - virtual space, image, and computing environment.

The practical consequences of using the network information model of the neural network and neuron in computing intelligence technologies and new forms of computing mean that many unattainable and unknown properties and cognitive capabilities of intelligent computing systems for processing and analyzing information in the conditions of isolation, limitation, exchange and uncertainty become available already in finished form. These properties and cognitive capabilities of the indicated systems include the ability, similar to living information systems, of creating new forms of intelligent streaming virtual parallel computing, new forms of reliability of their performance in conditions of isolation, limitations, exchange and uncertainty, and a new class of computer computing systems. By implementing a new computational paradigm, the present invention not only provides computational advantages in the fields of computational intelligence and the implementation of new forms of computing, but promises to generalize traditional computational science by moving to a different style of reasoning in terms of information dynamics and figurative representation of objects of computer processes in a virtual environment of intelligent computing.

The conclusions are both having theoretical development and practical continuation with wide and far-reaching direct and immediate results. This technology, on the one hand, promises a direct breakthrough not only in the field of AI and computer sciences, but also in understanding the mechanisms of intelligent information processing in living systems. And on the other hand, it promises significant new results in many other areas, including mathematics and computer science, metrology, psychology, and even sociology.

Brief Description of the Drawings

Figure 1 is a schematic (network) representation of a network information model of a neural network in the form of a two-dimensional "streaming" oriented graph.

To understand the essence and content of the present invention, it is necessary to understand the methodology of the method for describing and reflecting the dynamics of information processes of the computing environment under conditions of model and algorithmic isolation, exchange (energy and information), and information uncertainty. To understand the dynamics of information processes in computing systems and technologies, it is necessary to understand the mechanism and nature of the interaction of information processes in a virtual computing environment. To understand the mechanisms of information dynamics of objects of computational processes in the indicated computing systems, it is necessary to understand the dynamic, information, mathematical and metrological aspects of their information evolution, which form the methodological basis and ideology of the architecture of virtual streaming computing systems based on the network information model of the neural network and neuron. Therefore, the author of the present invention operates with the concepts and entities used in an order that reflects the dominant factors of the new ideology of developing virtual streaming computing systems.

The network information model of the neural network, which was developed by the author of the present invention and introduced into the ontology of the theory of neural networks [4], is defined as a multidimensional "flow" oriented graph with a set of input and output information links, and in Fig. 1 for a better presentation of the content and visual perception, it is presented in the form of a two-dimensional oriented graph. The vertices of the graph correspond to the neurons of the network, and the edges (bi-oriented) denote the logical and physical information buses for the storage, "transportation" and control processes in the network. Three types of information flows are defined: control flows, information flows, address flows. The organization of relations between vertices satisfies the condition that each vertex of any layer of a graph (subgraph) has a connection with all vertices of only neighboring layers. The type of information interaction between network elements is defined as processes with local information interaction. The information-spatial interaction of the network element K _ij with its environment in the network (see Fig. 1) is defined within the boundaries of neighboring columns (layers, which in the multidimensional case are defined as subgraphs). Any element of the j-th column interacts only with elements of the (j + 1) -th or (j-1) -th column. The mechanism of informational interaction of a neuron with its environment is characterized by vertical diffusion of information and localization of horizontal interaction processes in the restrictions on the right and left. Connections determine the type of information interaction: vertically - this is information diffusion, and horizontally - local dynamics with an offset. The functional, dynamic and topological properties of the information interaction of its elements are described at two levels: 1) detailing the topology and metrics of the neural network; 2) management and control of information dynamics on the elements of the neural network. The neuron’s memory region (the prototype domain of the neuron) is logically and functionally structured and is defined as a virtual local dynamic information system or network in which the interaction or functioning of its elements is described not at the level of signal transmission-reception, but at the information level: exchange, storage and generation information in storage and transmission channels. A neuron in such a dynamic information environment is defined as a local virtual computing system or virtual processor element. From a mathematical point of view, the meaning of the entity introduced in this way is as follows. First, the information prototype of a neuron in the address space of active memory will be the segment that defines the area of its information definition in the space of virtual addresses - its domain, and the range of values - its range. Secondly, the memory elements of such segments are virtual cells of various types, which differ in functional characteristics, logical organization, and information content. Cells that belong to the passive class are functional elements, i.e. in addition to information storage and transmission operations, virtual information interaction operators are defined on them, namely: 1) between cells of various types; 2) between cells and the information environment. Active-type cells are virtual memory elements in the usual sense and definition, i.e. perform the functions of storing and transmitting information, but the logical structure of such cells and the organization of information in them changes dynamically.

Models of algorithms and procedures that describe information processes on interacting chains of characters in the information environment of a neural network are based on the mechanism of processes with local interaction and the virtual perspective method, which are described in detail and in detail in the works of the author of the invention [1, 2, 7].

The level of informational state of neurons is determined by an assessment of a probability measure given on a fuzzy subset of informational signs of passive virtual cells of the corresponding active memory segments. A neuron is defined as a subset of neighboring passive virtual memory cells that are logically organized and functionally connected. The computing and information environment of a neuron is defined by a subset of passive and active virtual cells in the local area of active memory. If the measure on the passive cells of the segment is nonzero, then the corresponding element of the neural network is in the active state, and otherwise, in the passive state.

Virtual active memory of a neural network is divided into the following types of virtual cells. The first type is passive cells, the dimension of which is fixed and the information in them does not change. They are also elements of a data binding system. The second type is passive-active cells, the dimension of which is fixed and the information in them dynamically changes. The third type is active cells, the dimension of which and the information in them dynamically changes. The fourth type is active-passive cells, the dimension of which is dynamically changing, but the semantic information does not change. The nature of the information interaction between symbol chains in active cells is defined as processes with local interaction and interaction with a virtual information environment, and the type of information interaction between symbol chains is defined by an evolutionary operator defined on the information space of chains of interacting symbols. The law of information interaction between chains of characters in passive and active cells is set in the form of two operators: an interaction operator and a design operator.

The interaction operator f is defined between the character chains x _i and x _{i + 1} stored in active cells, for a simple implementation it has the following form:

where x _i is the address of the active virtual cell in which the first operand x _{i of the} operator f is stored, and x _{i + 1} is the address of the active virtual cell where the second operand x _{i + 1 of the} operator f is stored and where the result will be written. It should be noted that equation (1) will be true for the vector form.

The design operator □ defines the mechanism of information interaction between symbolic chains stored in active and passive virtual cells, and Fig. 5 is a schematic drawing illustrating the operation of the operator □. Here z _i is a passive cell, and {m _k } is a subset of active cells; d _ik is the informational distance between symbol chains in cells z _i and m _k , d _∂on is the informational sphere of influence of a passive cell z _i . For information objects of the subset {m _k } of active cells interacting with the information object in the passive cell z _i , the operator □ is defined as the following generalized expression:

□: {m _k } → z _i , k = 1, 2, ...

And the weight function □ (z _i ) of the cell z _i is determined on the basis of the following expressions:

□ (z _i ) = n _i / □,

. where n _i is the cardinality of the subset {m _k } of the set of active cells that fall into d _∂on is the neighborhood of the cell z _i , and

.

The mechanism of action of the operator □ is as follows.

Because Since symbol chains in passive virtual cells have a fixed length that is always less than the variable length of chains in active virtual cells, the operation of informational “alignment” of symbol chains of different lengths is first implemented. The logical scheme and mechanism of such an operation is defined as interaction with the information environment. A numerical estimate of the information distance between symbol chains stored in active and passive virtual cells is calculated by the following formula:

where M is the length of the character chains x _i and x _{i + 1} , w _i is the weight of the i-th position of the character chain, and _i and b _i are the i-th characters of the chains corresponding to x _i and x _{i + 1} . If it does not exceed the admissible criterion d _∂on , which is a given information neighborhood of a symbol chain in a passive virtual cell, i.e. fair inequality

then it is believed that the result of the interaction operator is in a passive virtual cell for which condition (2) is satisfied.

The information neighborhood of such a passive virtual cell may have arbitrary morphometry or topological mosaic. The action of the mechanisms of information interaction of symbolic chains stored in active and passive virtual cells of active memory is that information from active cells is “projected” into passive. This means the following: if the value of the symbol chain of the active cell falls into the information neighborhood of the symbol chain stored in the passive cell, then its information weight increases.

The network information model of a neural network and a neuron created by the author of the present invention is the methodological basis for the development and implementation of virtual streaming computing systems and technologies for solving a wide class of streaming problems in conditions of model and algorithmic isolation, exchange (energy and information), limitations of hardware and software capabilities and information uncertainties based on the implementation of new forms of computing. It should be noted that the concept of computing is defined and expressed in terms of fundamental principles, according to which is the science of information processes in computer computing systems.

The practical implementation of new forms of computing under the indicated conditions, limitations, and opportunities involves, on the one hand, the development of real and virtual computing components based on new logical and topological architectures. And on the other hand, equipping new forms of computing with appropriate logic and software, since there is a need to develop both new logical schemes for organizing stream computing under conditions of isolation, limitation, exchange (information and energy) and uncertainty, and new algorithms.

The methodology of the present invention is based on an inductive method for developing virtual sweat (parallel) computing systems and information processing technologies and logic circuits for organizing computing processes. The nature of the dynamics of objects of information processes (these are processes with local interaction) in the indicated systems and technologies and the nature of their interaction in the computing environment are the dominant ideological factors in the way they are developed and implemented. Information objects in a virtual environment of streaming computing systems and technologies are interacting chains of characters (these are objects of non-numerical nature), and any formal object of a simulated procedure, operation or task in a computing environment is defined and described as the logical structure of such chains. The environment for the interaction of objects of computer processes of the indicated computing systems and technologies includes the following main intermediary attributes: a virtual logarithmic variable, interaction operators, a logical structure in the address space of the physically limited memory of the computing system, and others that are components of the computing environment. From an intuitive and logical point of view, the computing environment contains all or almost everything that relates to obtaining a result. All information objects of computer processes and the virtual computing environment in which the simulation and the process of obtaining the result are reflected and described on the basis of the formalism of the information dynamics of interacting chains of characters and the principles of constructing non-traditional computing computing technologies: the principle of interaction of information objects and the principle of virtual perspective [4, 6 ].

Everything that is done in computing comes down to some basic principle: what is considered to be the identifier of the data stream or an individual element is taken and what is considered to be its value is constructed relative to the environment. It is this construction process that is considered to be calculation, and computing itself develops construction implementation technologies. The relationship between the identifier and its value is parameterized by the computing environment.

The forms of computing in the framework of dynamic models of computational technologies (dynamics models of interacting chains of characters) are primarily associated with the development of new organization schemes and new algorithms, i.e. there is a situation of transition from sequential algorithms of scalar (distributed) computing systems to algorithms for computational technologies of streaming parallel virtual systems under conditions of model isolation, exchange restrictions, and information uncertainty.

The main principles and principles of the approach under consideration for the methodology for creating virtual streaming (parallel) computing systems and technologies can be defined as follows: the basis of the approach in terms of the development and implementation of new forms of computing is the principles of interaction of information objects of the computing environment on quantum discrete information spaces and the principles of virtual prospects and interactions of objects of the computing environment with the information environment of the computing system (ac virtual memory). Within the framework of the indicated approach, the organization and implementation of computing assumes that the dynamic evolution of interacting symbol chains in a computing environment is determined on quantum discrete information spaces of the active memory of computer systems, and the set of interacting and evolving chains of symbols in the cells of the address space of such memory, which is the physical medium of such interaction form a quantum discrete dynamic information system with the environment. In formalizing, interpreting, and implementing such a system, it is necessary to separate its dynamic, informational, mathematical, and metrological aspects. In the mathematical aspect, such a system means the following. Firstly, it is defined on a discrete quantum information space, the element of which is an information quantum, which is also an information attribute in this space. The logical prototype of such a quantum is a chain of characters, and the physical is an algorithmic variable, a virtual memory cell, or a subset of them. By means of such an attribute, the spatiotemporal region of the object of the computing environment in virtual active memory is localized, and the measure of informational measurement is the depth of the quantum - this is the reciprocal of the length of the symbol chain that can fit in the designated memory region and is an information identifier of the quantum type. Secondly, the operator of interaction of symbolic chains in operations and algorithms of computing technologies of such virtual computing systems is defined not in the form of traditional and familiar arithmetic and logical operations, but in the form of operators of interaction of symbolic chains with the information environment and among themselves, as well as the operator of designing the resulting symbolic chains to cell nodes of a virtual information base coordinate system [1, 2, 3]. In this case, we have a completely different type of information dynamics of interacting chains of characters than the one generated by the computing technologies of currently existing computing systems. A striking difference is that existing computing systems generate information dynamics on the cells of real physical memory of the same physical length, although in the virtual version this cell length can be increased, but it is fixed anyway. In the computing environment of such computing systems, information processes of computing technologies (operations and procedures) are generated and proceed under the following conditions: 1) isolation (relative to memory elements and the information environment); 2) conservatism or potentiality (as a relatively limited hardware and software capabilities, energy-information exchange with an external real and virtual environment); 3) information uncertainty - any interaction of two character chains fixed as a result generates information uncertainty.

In reality, no information dynamic system, in particular a computing system, is closed and conservative, on the one hand, in the sense of activity and passivity of its processor elements and memory elements, and on the other, information dynamics in such information systems of its objects and information environment it is static and there is no way to implement information control, diagnostics and control of the computing process. The existing ideology of developing the implementation of computing systems (streaming and scalar) does not allow the implementation of new forms of computing in the framework of the concept of isolation and potentiality. A new way is needed here (foundations, principles, formalism) - this is a rejection of the conditions of isolation and potentiality, i.e. transition to open information dynamic systems, in which the idea of the completeness of the information dynamics of computing processes is determined and composed of their dynamic elements and subsystems, as well as the nature of their interaction with the information environment.

The considered and proposed method of creating virtual computing systems and technologies just offers a rejection of the principles of their information isolation and conservatism. The processor elements and memory elements of such systems are considered as multifunctional virtual subsystems within its hardware-topological architecture, information-logical organization of the relationships between elements and hardware-functional capabilities. Virtual information subsystems in the hardware-physical and information environment of such systems are both static and dynamic objects. The formation or generation of such subsystems is implemented at a logical level in the information environment of support systems (software systems and technologies) of the life of a real computing system.

The dynamic aspects of quantum discrete information systems, one way or another, are associated with the presence of quantum attributes in dynamic systems, such as time quantization (physical and informational) and space (real and virtual), etc.

In terms of information, such systems mean the following. Firstly, the information dynamics of interacting chains of characters in operations and procedures of computational technologies is not based on the general principles and methods of traditional computational mathematics, which operate with arithmetic objects (numbers are logical and algebraic structures), not reflecting their information meaning and content, but on principles of joint quantum action on a quantum discrete information system of the interaction energy in the information field and information in a closed environment, changes, uncertainty and a strong deviation from the dynamic and informational equilibrium. And this means that there is no closed interaction (when memory cells for operands are closed within a fixed length) of logical and arithmetic objects in the classical (familiar) sense, but interaction occurs at the level of quanta and virtual information environment. Secondly, in the virtual environment of the flow of information processes, on the one hand, there are other laws that determine the information dynamics of those objects (information quanta), through which the dynamics of symbol chains are generated and developed in operations and procedures of computing technologies that interact with the virtual information environment, and on the other hand, in the information environment of virtual computing systems that operate in conditions of isolation, limitation, exchange and uncertainty, other units of measurement and scale systems (physical, informational, virtual).

The metrological aspects of such systems are due to the following facts. Firstly, the computing objects of such systems have coordinate systems of information binding, which numerical objects of computing technologies of existing computing systems based on algorithms of traditional computational mathematics do not have. Secondly, the possibility of implementing models of algorithms and procedures of technologies of information metrological verification and verification of computational information objects in conditions of isolation, limitation, exchange and uncertainty.

The main advantages of the paradigm of quantum discrete information dynamical systems, such as virtual streaming computing systems and new forms of computing implementation, are that they, from the standpoint of a single conceptual scheme and, in particular, a single mathematical, logical, and algorithmic formalism, allow one to naturally take into account the fundamental points who brought and bring to the theory of building streaming computing systems and technologies the ideas of openness, locality, virtualization of quantum discrete spaces and information quanta.

The fundamental points in constructing a methodology for creating virtual streaming computing systems and technologies based on the above principles and approaches are the following main factors. Firstly, the information dynamics of computing process objects cannot be closed and limited in the information space of the computing environment, nor can they be defined in the virtual information and address space of a real computing system (at the logical and virtual level), but it can only be localized using the information quantum. Secondly, in the allocated (final) quantum of physical (or informational) time, the dynamics of objects of computational processes under conditions of physical isolation, restrictions, exchange, and uncertainty have infinite informational evolution of possible quantum states, which can be interpreted as wave functions on an ordered set of time quanta, those. vectors with unit norm.

Multilayer modular implementation scheme of a virtual streaming computing system.

The physical and virtual implementation of the virtual streaming BC architecture based on the network information model of the neural network and neuron is carried out according to the principle of a multilayer modular computing system. The organization of such a system includes several layers: a layer of processor elements (PE) (1st layer: n1), layers of active virtual memory for each PE and a group of PE (2nd layer: n2), layers of shared active virtual memory, accessible to all PE (3rd layer: n3). Such a scheme for organizing a multilayer modular computing system can be implemented both on the physical and logical level, and on the virtual. The number of layers of such a system is determined by the geometry and topology of the neural network and the graph of the problem to be solved, the requirements for the reliability of the result, the conditions for organizing the computing process, etc. Each layer contains elements (physical and virtual) for the functional and logical purpose of one type for each layer (1, 2, 3 ) The interface between the elements of the layers is organized at the logical and physical levels using inputs and outputs: 1st layer - i | o1; 2nd layer - i | o2; 3rd layer - i | o3. The processor elements perform the functions of an interface both between layer elements and with the external environment, initializing the computing system, routing information in the system, configuring the topology of the neural network as a logical image of the architecture of a virtual streaming computing system under the structure of the information graph of the problem to be solved, restoring the results of the computing process in case of failure system, individual layers or elements. The processor layer n1 consists of m processor elements (PEs) with an arbitrary topology of links and an optionally unified logical pooling of interaction between them. In this case, PE is the functional module of the layer. The communication channels between them and bandwidth also need not be the same. Each PE has a two-level random access memory (OP). The set of OP blocks for each PE forms a layer n2, i.e. OP blocks are memory layers.

The first level of the OP module is available only to a specific PE or PE module. The memory module of this level is divided into the following fields:

1st field is a field of attributes of PE, which are the following characteristics: 1) the logical number of PE is an information attribute by which the division number in the quantum scale for measuring the values of the calculated value is identified; 2) a sign and level of information activity of PEs, through which, during each allocated quantum of physical time of streaming processing, new PEs can turn on and their activity level is estimated; 3) attributes of communication with memory of lower levels;

2nd field is a command memory field;

3rd field - the field of active basic virtual memory, which is divided into passive virtual cells of arbitrary fixed length, unchanged during the computational process, and the information in them also does not change. This field is that area of memory (inaccessible for writing) in which the local coordinate system of information binding and verification of the results (intermediate and final) of computing processes in the environment of PE calculations is stored. The coordinate information system, in turn, is a global information attribute, through which it is specified to define the mapping of the states of the dynamic evolution of computing process objects in the information space of active memory — their domain — and their range of values — their range.

The second level of memory in its organization (logical, virtual and informational) with priority access to the fields of this level by various PEs has the following features. Firstly, the fields of this level form a dynamically active virtual memory, the logical organization elements of which are two types of virtual cells: 1) active cells, the dimension of which and the information in them change; 2) active-passive cells, their dimension changes, but not semantic information. Secondly, this memory level is logically divided: into fields accessible to a specific PE and into fields available to other PEs of a virtual streaming computing system.

The fields available to a particular PE are subsets of virtual addresses for various types of virtual cells in real physical memory. The logical structure of the organization of these fields in computational processes is dynamically changeable. The mechanism for implementing a dynamic variable logical structure of such fields in the address space of various types of virtual cells in a limited area of the real physical space of the aircraft memory is simple enough to implement various forms of computing and is not tied to either data types or addressing systems. In this case, the contents of the virtual cell are interpreted as a binary field (set), and the addresses of virtual cells (the logical structure of the address space) can change after each operation or their combination. These fields also form a virtual address space for working cells and storing intermediate results of a particular PE. The third layer n3 forms the space of the total active virtual memory in the address and information space of the real physical memory of the aircraft. This layer is logically divided into memory modules, which are used to implement the following goals. First, a logical structure of virtual memory is formed on a selected subset of modules, which reflects the coordinate lattice of the information system of coordinates for checking and verifying the results of a computational process, the nodes of which correspond to passive virtual cells [1, 2, 3]. The results of calculations (intermediate and final) in successive time quanta into which the computational process is divided are stored in passive memory cells of the coordinate lattice and form a logical raster structure for the reflection of the results in the form of symbol chains for values of a discrete blurred random function. The logical structure of the results at the nodes of the lattice is an information display of the final topology of the image of the solution of the simulated problem or computational process by layers. Secondly, the modules of layer n3 perform buffer functions in tiling operations between layer modules, the functions of virtual ports and information gateways with the outside world.

Organization of the computing process in the computing environment of a virtual streaming computing system

Computing systems based on the network information model of a neural network and a neuron are universal computing systems of a new generation with new capabilities. In this section, the author of the invention confirms this statement and describes the logical configuration diagram of the architecture of the virtual streaming computing system according to the geometry and topology of the neural network under the structure of the information graph of the problem being solved.

The fundamental unit of a virtual streaming computing system is a neuron (see figure 1). The physical prototype of a neuron in the processor layer n1 of a computing system is the union of a certain amount of PE (1), and in the layer n2 it is the union of the corresponding memory modules (2). The amount of PE in a neuron is determined by the functional purpose and logical organization of the computing process according to the structure of the information graph of the problem being solved. The organization of information input and output connections for various types of buses between the modules of layers n1, n2, n3 within the boundaries of the information space of both the neural and the neuron is carried out at the inputs and outputs i | o1, i | o2, i | o3 of the multilayer modular scheme of the computing system. PEs perform arithmetic and logical operations on symbolic chains of variable and fixed length, which are stored in the active virtual memory of module 2. Since Since computing processes in a virtual environment of a computing system are implemented and proceed under conditions of model isolation, restrictions, exchange, and uncertainty, models of algorithms and procedures in a computing environment are constructed using stochastic Monte Carlo-type schemes. Computation algorithms constructed by such schemes belong to the class of algorithms on fuzzy subsets, and the results of calculations based on such algorithms are fuzzy stochastic quantities whose values (symbol chains of a fixed length) are defined on the lattice nodes of the information coordinate system, the prototype of which is a set of passive virtual active virtual memory cells.

The logical scheme for constructing calculations of the type of iterative processes, the accuracy of the calculations and the effectiveness of which cannot be verified by direct substitution, based on vague algorithms of operations and procedures can be described as follows. The accuracy of the information representation of the source data, operands of operations, and the results obtained in the computing environment of any computer system in the form of a symbol chain is specified by the number of significant characters. In order to ensure the required accuracy specified by the number of significant characters, over the entire interval of such an iterative computational process, it is necessary that the length of the symbol chains of the operands of logical or arithmetic operations be much greater than the required accuracy. In this case, chains of different lengths participate in the operations. Therefore, at the beginning of the main operation, as described above (the dynamics of such computational processes are described in more detail in the works of the inventor [1, 2, 3, 4, 5, 7]), the operation of interaction with the information environment is implemented, and then a specific logical or arithmetic operation. It should be especially noted that in the computing environment of streaming virtual computing systems, the difference between the concepts of the calculated value and the calculated value is strictly defined and indicated, which consists in the fact that the calculated value is a deterministic value, and the calculated value is a blurred value, the information image of which (symbolic fixed-length chain) is defined on the lattice nodes of the information coordinate system. After the interaction operation between operands is completed, the operation of designing the result of the previous operation to the lattice node of the coordinate information system is performed. The initial conditions of the indicated iterative process are determined and set in the information neighborhood of one or more nodes of the lattice of the information coordinate system. The lattice of the coordinate information system is logically divided into layers of possible values (domain of definition) of the result of the calculations, interconnected in time or in a parametric variable, i.e. Each neuron or module has its own domain-range in the value space of the general scale of the information coordinate system. The time interval (physical or informational) of the implementation of the computing process is a multiple of the physical quanta of time of a fixed duration. In the selected quantum of physical time, only information objects of neighboring layers interact. It should also be noted that the lattice of the information coordinate system has a geometry similar to a neural network. In each quantum of physical time in the computing environment, the number of active neurons is fixed and determined, but after the calculations are completed at a specific fixed time quantum, their number in the next quantum step can change quantitatively and qualitatively, when new neurons can be connected to the calculations. This is due to the fact that during informational interaction of calculation results with symbolic chains in the nodes of the lattice of the information coordinate system, they can fall into the information zone of inactive neurons and then such a neuron is activated in the next quantum step.

When organizing a computing process in the computing environment of a computing system having the architecture of an information model of a neural network, the coordinate lattice of the information system for linking and checking the results of calculations is logically and informationally divided into parts and each neuron functions in the corresponding area of the information space of the coordinate lattice.

After completion of the computational process at the lattice nodes of the information coordinate system, we obtain a fuzzy image of the image of the calculation result in the form of a topological complex on the set of active nodes, the values of which are reflected in the set of passive virtual cells of virtual active memory.

Next, the procedure for extracting from a blurred image the result of calculating a deterministic image of the result is implemented. To implement this procedure, the operation of topological incarnation of a blurred image of the result is carried out, which takes into account the criteria for the connectedness of the calculation results: fractal connectedness and informational connectedness [5]. The adequacy of the result is determined by the degree of proximity of the estimates of the result obtained according to the specified criteria.

What virtual streaming computing systems based on the information model of the neural network and neuron are really effective and new in is not emulation or synthesis of computing technologies in the computing environment of traditional computers, but emulation and synthesis of intelligent information processing technologies in the systems of perception of living systems. For example, the eye of a living system detects and processes information based on the following operations: interaction and projection into the nodes of the fundus lattice. The ear processes information in the same way. The unique and sustainable nature of the information model of the neural network and its elements determines and reflects its analogue in the architecture of the corresponding virtual streaming computing system, which creates wide opportunities for the development of artificial and computational intelligence systems, a wide range of logical and software designers related to the implementation of solutions in the field of artificial intelligence and pattern recognition, through the fact that they allow you to implement new principles for the implementation of computing.

An additional advantage of a virtual streaming computing system is a unique form of parallel processing of information and organization of parallel computing processes. Since all neurons of a neural network form local dynamic computing networks, which are autonomous within a quantum-step of calculations, even for processes united by a single algorithm, they allow the use of a neural network with its geometry and connection topology as a template environment in which neurons can simultaneously participate in solving various problems and implementing a wide range of logic circuits.

Conclusions, consequences and scope of invention

Thus, the described network information model of the neural network and the theory of the dynamics of the information interaction of computing environment objects in it show the practical nature of the present invention. New entity-objects were described: a network information model of a neural network and a neuron, an active virtual memory model, passive and active virtual memory cells, a multilayer modular scheme for the implementation of a designated computing system, processes with local information interaction, interaction and design operators, an information coordinate system of computing environments . Each of these entities is an architecture or its element for a system that can be broadly defined as a virtual streaming computing system based on a network information model of a neural network and neuron and which is an object of the present invention.

The formalism of the information model of the neural network and neuron described above at the level of network models and algorithms allows us to implement such an architecture (structure) of a virtual streaming computing system in which the computing environment and information processes are logically and functionally organized based on the following formulated principles of computing: 1) the principle of interaction of objects (character chains) computing environment with a virtual information environment; 2) the principle of virtual perspective in the design operator; 3) the principle of connectedness of the calculation results.

The foregoing disclosure and description of the invention are an interpretation and explanation of the invention. Those skilled in the art to which this invention relates will understand that various modifications of the details and embodiments of the processes and structures described herein in order to explain the nature of the invention can be made without departing from the principles and / or essence of the invention.

Any equivalent formulation of the above-described entity-objects is within the scope of the present invention. Finally, any equivalent formulation of a virtual streaming computing system based on a network information model of a neural network and a neuron is within the scope of the present invention. Accordingly, the scope of the invention should be determined by the attached claims and their legal equivalents.

Literature

1. Myshev A.V. The virtual perspective method in modeling "fuzzy" tasks. - Information Technology and Computing Systems, 2011, No. 3, p.66-79

2. Myshev A.V. Virtual perspective method and modeling under constraints and uncertainty. - Software products and systems, 2012, No. 2, 50-58.

3. Myshev A.V. Computing and modeling the fuzzy Cauchy problem using the virtual perspective method - Software products and systems. 2012, No. 3, p. 217-227.

4. Myshev A.V. The information model of the neural network in the technologies of computational intelligence and forms of implementation of computing. - Information Technology. 2012, No. 1, p. 63-79.

5. Myshev A.V. Metrological theory of the dynamics of interacting objects in the information field of a neural network. - Information Technology, 2012, No. 11, p. 52-62.

6. Myshev A.V. Dynamics of information processes in computational technologies of computer modeling. / Proceedings of ISA RAS, vol. 58. - M .: KRASAND, 2010, p.137-148.

7. Myshev A.V. Virtual perspective method and neural network algorithms in computer modeling technologies. Neurocomputers: development and application. - M .: Radio engineering, 2007, p.395-410.

Claims (1)

A method of constructing a virtual streaming computing system based on an information model of an artificial neural network and neuron, characterized in that the logical organization of the architecture of the virtual streaming computing system is implemented according to the principle of a multilayer modular system that includes a processor layer, active virtual memory layers for each processor element (PE) and PE groups, layers of shared active virtual memory accessible to all PEs, layers - processor and active virtual memory - have modular decomposition, processor modules are logically and functionally implemented in the form of a virtual computational structure in accordance with the geometry and topology of the information model of the neural network, the elements of which logically and informatively perform the functions of neurons, and the interaction between them is implemented according to the same principles as the neurons in the designated neural network model , the basic and forming elements of modules of active virtual memory of different layers are virtual cells of various types (active and passive), the environment module of the processor layer with the corresponding modules of different layers of active virtual memory includes coordinate systems for information binding and verification of the results of computational processes with a given number of significant characters (digits), ensuring their control, reliability and accuracy at certain spatio-temporal scales, based on fractal criteria and informational connectivity of limited and fuzzy values of the results of calculations, and a model of a virtual information environment.