Here we present an analysis of the current state in the field of development of hardware accelerators of artificial intelligence (AI). Despite the fairly good progress made over the past decades, this area is experiencing a number of significant difficulties in its development. The solution to this problem lies in the application of new approaches to the organization of computing, in particular, computing in memory enabled by nanoelectronic memristive devices. We provide an overview of state-of-art systems, as well as our own version of the experimental concept of AI accelerators based on metal-oxide memristive devices and the massively parallel architecture for information processing.

The article is devoted to the reliability of hardware implementation of artificial neural networks based on memristive devices (ANNM). On the basis of the system engineering methodology, the authors have developed a universal general approach to determining the reliability of ANNM “from the accuracy of functioning through fault tolerance”. The active and passive methods of ensuring the reliability of ANNM are described. An example of determining and ensuring the reliability of a specific version of the hardware implementation of ANNM is given.

Рассмотрен процесс разработки многослойного персептрона на основе мемристивных устройств для двунаправленного адаптивного нейроинтерфейса. Предложен общий подход к обеспечению точности его функционирования с учетом влияния дестабилизирующих факторов.

This article describes the influence of algorithms for tuning the parameters of neuromorphic systems on their fault tolerance. This is relevant to the hardware implementation of neuromorphic systems using memristors (NSM). The study is conducted using the authors developed a variant of the system approach and methods of simulation of artificial neural networks (ANN). By the example of a multilayer perceptron, it is shown that different ANN learning algorithms in the nominal mode of operation make it possible to achieve similar values of the operation accuracy. But due to the influence of production and operational factors in real conditions of operation, the ANN may fail. The range of allowable values of the destabilizing factors on ANN operation depends on the learning algorithm and may differ several times.

A general approach to the development of memristor-based artificial neural networks (ANNM) operated with specified fault tolerance (FT) is formulated and applied in the paper. It is shown that ensuring the required ANNM FT is related to ensuring the required accuracy of their operation at all the structural and functional hierarchy levels. The paper proposes a quantitative FT criterion that can be used to create ANNM reliability block diagrams, calculate and optimize reliability in accord with the actual Russian and international standards. The application of the proposed algorithm is considered on the example of the ANN performing an approximation of mathematical functions, the synapses of which are implemented with memristors. It is found that a potentially high ANNM FT cannot be achieved by itself only because of the massive parallelism of artificial neural networks. Instead it depends on many factors and requires the application of special physical and information technologies at all the ANNM life cycle stages.

The authors have reviewed interpretations of the terms “dependability” and “fault-tolerance” in Russian and interstate standards. A new quantitative criterion of fault-tolerance of the memristors-based artificial neural networks is proposed and substantiated. The authors have also proposed and provided rationalization for a revised definition of fault-tolerance as a property of the memristors-based artificial neural networks, which most fully conforms to the new version of its quantitative criterion. An example of the application practice for the fault-tolerance criterion during the design stage of an artificial neural network of a test degree of complexity is given.

The article covers a solution to a problem of defining the tolerances of information and physical parameters of components of artificial neural networks ( ANNs), which are implemented as hardware through the application of nanoscale electronic components with memristive properties (memristors). The developed method foundation is a system approach to the memristors-based ANN (ANNM) design, whereby the ANNMs should be studied as united physical and informational objects. When the ANNM is produced and operated, the errors of its components' physical parameters provoke information parameter errors. To define the tolerated errors (tolerances), a simulation methodology is used. The potential of the developed method is illustrated through the process of defining tolerances for the synaptic weights and neural biases of a two-layer feed forward ANNM.

This paper is concerned with the issue of secure radio communication of data between manned aircrafts, unmanned drones and control services. It is indicated that the use of artificial neural networks (ANN) enables correct identification of messages transmitted through radio channels and enhances identification quality by every measure. The authors designed and implemented a simulation modeling technology for ANN development, which enables signal detection with required accuracy in the context of noise jamming, natural and other types of noise.