Sections

Neuroinformatics - 2019


POSTER SESSION 1

Tuesday, October 8                    16:30 – 17:30
Lecture-hall Фойе 2 этажа

Chair: Prof. BURTSEV MIKHAIL

Artificial intelligence

1. MARIA TARAN, GEORGIY REVUNKOV, YURIY GAPANYUK
Bauman Moscow State Technical University
The Hybrid Intelligent Information System for Poems Generation

Any generated text must have a “form” and “content” components. It is the “content” that is the main component of the generated text, but the “form” component is no less important. It may be necessary to generate texts in different linguistic styles among them the poetic linguistic style. The article proposes an approach for poems generation problem using hybrid intelligent information systems (HIIS). The HIIS consists of two main components: the subconsciousness module and the consciousness module. In the case of poems generation, the subconsciousness module consists of two submodules: the stress placement module and the rhyme and rhythm module. These modules use machine learning techniques. The consciousness module includes the poem synthesis module, which is rule-based. The stress placement module is based on the convolutional neural network. On the test dataset, the accuracy of the classifier is 97.66%. The rhyme and rhythm module based on neural networks with a depth of 5-7 layers. On the test dataset, the accuracy of the classifier is 91.63%.

2. ENGEL EKATERINA ALEKSANDROVNA , ENGEL NIKITA EVGENEVICH
Katanov Khakass State University, Abakan
The Photovoltaic System’s Control Model on the Basis of the Modified Fuzzy Neural Net

This paper represents the photovoltaic system control model on the basis of a modified fuzzy neural net. Based on the photovoltaic system condition, the modified fuzzy neural net provides a maximum power point tracking under random perturbations. The architecture of the modified fuzzy neural net was evolved using a neuro-evolutionary algorithm. The validity and advantages of the proposed photovoltaic system control model on the basis of a modified fuzzy neural net are demonstrated using numerical simulations. The simulation results show that the proposed photovoltaic system control model on the basis of a modified fuzzy neural net achieves real-time control speed and competitive performance, as compared to a classical control scheme with a PID controller based on perturbation & observation, or incremental conductance algorithm.

3. A.A. CHERNYSHOV, V.V. KLIMOV, A.I. BALANDINA
National Research Nuclear University (MEPhI), Moscow
BiLSTM-BASED APPROACH TO THE NATURAL LANGUAGE TEXT DEPENDENCIES ANALYSIS

This article discusses the idea of conducting a multistage process of building a search image of a query in natural language for use in the semantic search system. Modern methods and tools for processing natural language are widely used in the field of machine translation. Research on search engines and semantic search mainly focuses on data storage and further analysis. Most search engines use a huge amount of previously accumulated user queries to predict search results, without taking into account the user's intention through quality query processing. The proposed approach is based on the selection of the maximum amount of information from the original request by means of syntactic and semantic analysis, as well as the use of synonymous extension techniques. This article describes the first step in the process of building a search query model, based on extracting syntactic dependencies from the original sentence.

4. BURAKOV M.V.
Saint Petersburg State University of Aerospace Instrumentation
FUZZY ESTIMATES OF FITNESS FUNCTION FOR GENETIC SYNTHESIS OF NEUROCONTROLLERS

The task of choosing the fitness function in the synthesis of a neurocontroller by genetic algorithm is considered. It is proposed to use fuzzy evaluations of fitness, which allow for a more accurate comparison of various options for system movement and speeding up the process of finding a solution. The results of the computational experiment in Simulink MatLab to determine the parameters of a neurocontroller for a nonlinear dynamic plant are given, which confirm the practical usefulness of the proposed approach in the synthesis of neurocontrollers for a wide range of control systems.

5. V.D. KOSHUR, P.I. ROZHKOV
Siberian Federal University, Krasnoyarsk
IMPLEMENTATION OF THE MULTI-AGENT SYSTEM OF ARTIFICIAL INTELLIGENCE FOR THE SOLUTION OF THE CLASSIFICATION

Designing a multi-agent system that allows you to flexibly solve the problem of classification. This approach is based on the use of agents that are in their own isolated space and formally, they do not know anything about each other, but there is a “communication” between them - the transfer of information about accumulated knowledge. Each agent solves its own problem and is part of a common agent system.

Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling

6. VALERIIA A. DEMAREVA AND YU. A. EDELEVA
1N.I. Lobachevsky State University of Nizhni Novgorod
2Technische Universität Carolo-Wilhelmina zu Braunschweig
Is information density a reliable universal predictor of eye movement patterns in silent reading?

The role of information density as a reliable universal predictor of eye movement patterns in silent reading is considered. Density differences between Russian and English are taken to explain the difference in eye movement patterns for readers with Russian as a native language compared to English-speaking readers. An empirical eye tracking study shows that only one of four expectations got confirmed. Supposedly, the eye-movement pattern observed for Russian could be in-fluenced other factors.

7. SOZINOVA I.M., ARUTYUNOVA K.R., ALEXANDROV YU.I.
1Moscow State University of Psychology and Education
2Institute of Psychology of Russian Academy of Sciences, Moscow
Spectral parameters of heart rate variability as indicators of the system mismatch during solving moral dilemmas

Variability in beat-to-beat heart activity reflects the dynamics of heart-brain interactions. From the positions of the the system evolutionary theory, any behavior is based on simultaneous actualization of functional systems formed at different stages of phylo- and ontogenesis. Each functional system is comprised by neurons and other body cells, the activity of which contributes to achieving an adaptive outcome for the whole organism. In this study we hypothesized that the dynamics of spectral parameters of heart rate variability (HRV) can be used as an indicator of the system mismatch observed when functional systems with contradictory characteristics are actualized simultaneously. We presented 4-11-year-old children (N=34) with a set of moral dilemmas describing situations when an in-group member treated an out-group member unfairly, which benefited the in-group and put out-group members’ lives at risk. The results showed that LF/HF ratio of HRV was higher in children with developed moral attitudes for fairness toward out-groups as compared to children who showed preference for in-group members despite the unfair outcome for the out-group. Thus, the system mismatch in situations with a moral conflict is shown to be reflected in the dynamics of heart activity.

8. NEKRASOVA JULIA, SUDAREVA ANASTASIA
Moscow Aviation Institute (National Research University)
Adaptive sensorimotor brain-computer interface for managing rehabilitation facilities

The paper deals with a non-invasive sensorimotor brain-computer interface based on an electroencephalogram for managing rehabilitation tools such as exoskeletons, artificial limbs and wheelchairs. A feature of the device is the implementation of the technology of localizing sources of electrical activity of the brain (beamforming), the use of which will allow to design a neurointerface device with the number and location of electrodes that is optimal for a particular patient.

9. YURIY SKOBTSOV, OLGA CHENGAR
1Saint Petersburg State University of Aerospace Instrumentation
2Sevastopol State University
Synthesis production schedules based on multi-object ant algorithm

It is proposed to use ant algorithms together with the object-oriented simulation models. To optimize the functioning of the automated technological complex machining together with a modified ant algorithm it is designed object model, which allows calculate the fitness function value and evaluate potential solutions. The transition and calculation of the synthetic pheromone concentration are determined for supposed directed ant algorithms. Multi-object optimization with adaptive weights is proposed, where in the solution process the weights of the objective function are corrected. Experimental studies of two- and three-object optimization on the example of the production schedule have been carried out on an example of the automated technological complex of machining. For the first time, the use of ant algorithms is proposed together with object-oriented simulation modeling.

10. EVGENII VITYAEV
Sobolev Institute of Mathematics, Siberian Branch of the Russian Academy of Sciences
CONSCIOUSNESS AS A COMPLEX REFLECTION OF CAUSES OF THE EXTERNAL WORLD

In previous works, the problem of statistical ambiguity was solved and the maximum specific causal relationships were determined, the conclusion on which is noncontradictory. In this paper, the following hypothesis is argued: the brain makes all possible conclusions on perceived cause-and-effect relationships, creating an consistent model of the perceived world, manifesting itself as a consciousness. A formal model of a neuron that satisfies Hebb's rule, which reveals such maximally specific causal relationships, is proposed.

11. VLADIMIR G. RED’KO, GALINA A. BESKHLEBNOVA
Scientific Research Institute for System Analysis, Moscow
Evolutionary Minimization of Spin Glass Energy

The current work describes the model of evolutionary minimization of energy of spin glasses. The population of agents (modeled organisms) is considered. The genotypes of agents are coded by a large number of spins of the spin glass. The energy of the spin glass is calculated in accordance with the Sherrington-Kirkpatrick model. This energy determines the fitness of agents. The process of evolutionary minimization of the spin glass energy is analyzed by means of computer simulation. Several properties of spin glasses that are related to the model of evolutionary search are analyzed. In particular, the global energy minima of spin glasses and the variation of energy at one-spin mutation are estimated. The process of the gradual decrease of the spin glass energy is also analyzed. The gradual decrease is performed by sequential changes of signs of separate spins of spin glass. The computer simulation demonstrates that evolutionary optimization results in the finding of essentially deeper energy minima as compared with the gradual decrease. The rate and efficiency of evolutionary minimization of energy of spin glasses have been estimated and checked by computer simulation.

Neurobiology and neurobionics

12. BULAVA ALEXANDRA I., VOLKOV SERGEY V., ALEXANDROV YURI I.
1Institute of Psychology of Russian Academy of Sciences, Moscow
2P.P.Shirshov Institute of Oceanology of the Russian Academy of Sciences
A Novel Avoidance Test Setup: Device and Exemplary Tasks

This paper presents a novel rodent avoidance test procedure and device, that expands the possibilities for exploration of the memory and learning processes in a psychophysiological experiment. A novel footshock-avoidance test procedure allows flexible current-stimulating settings. In our work we use slow-rise current. The tested animal considers either a current rise or time-out intervals as a signal to action. Animals make a choice between escape or avoidance.

13. A. L. PROSKURA, S. O. VECHKAPOVA, A. S. RATUSHNYAK

DOPAMINE AND HIPPOCAMPAL SYNAPTIC PLASTICITY ДОФАМИН И СИНАПТИЧЕСКАЯ ПЛАСТИЧНОСТЬ В ГИППОКАМПЕ

In addition to its role in learning and memory formation, the hippocampus may act as a novelty detector. Hippocampal synaptic plasticity has been implicated in both spatial memory formation as well as novelty acquisition. In addition, the ventral tegmental area-hippocampal loop has been proposed to control the entry of information into long-term memory, whereas the dopaminergic system is believed to play an important role in information acquisition and synaptic plasticity. Physiological and behavioral evidence supports that dopamine receptor signaling influences hippocampal function. Hippocampal synaptic plasticity, in the form of long-term potentiation and long-term depression, has been implicated in both spatial memory formation as well as novelty acquisition. However, the molecular mechanisms of synaptic plasticity modulation by dopamine in the hippocampus remain poorly understood. The aim of this study was to achieve a better understanding of the functional association between glutamate and dopamine receptors on synapses of hippocampal CA1 field. The functional association ionotropic and metabotropic receptors facilitate synaptic plasticity processes in hippocampus. The dopaminergic system gates long-term changes in synaptic strength and these changes are a critical factor in the acquisition of novel information.

14. TATIANA A. PALIKHOVA
Lomonosov Moscow State University
EUGENE SNAIL: SOMATOSENSORY MAPS ON THE BRANCHES OF SINGLE NEURON OF LAND SNAIL УЛИТКА ЕВГЕНИЯ: СОМАТОСЕНСОРНЫЕ КАРТЫ НА ВЕТВЯХ АКСОНА НЕЙРОНА ВИНОГрАДНОЙ УЛИТКИ

The cognitive maps of the brain are present one of the main themes of neuroinformatics. Penfield’s homunculus is the classic example of presentation of sensory and motor body fields in the human cortex. In analogy with Penfield homo a term Eugene snail means a presentation of body fields at the branches of single neurons. The term was proposed in memorial of E. N. Sokolov known for his studies of information processing in the brain. The data presented here were obtained using intracellular recording and imaging of the identified neurons of land snail. We propose new level for cognitive maps distribution – the branches of single neurons.

15. KRYLOV A.K.
Institute of Psychology of Russian Academy of Sciences, Moscow
A cortex columns model based on metabolic cooperation and competition

A model of cortex columns creation is proposed. According to the known BOLD fMRI data cooperative activity of neurons of a cortex column dilates a blood vessel through the glia cells activity and thus increases influx of metabolites to the column’s neurons.

16. YARETS M.YU., SHAROVA E.V., GALKIN M.V., BOLDYREVA G.N., KULEVA A.YU., TROSHINA E.M., KROTKOVA O.A.
1Institute of Higher Nervous Activity and Neurophysiology of RAS, Moscow
2Burdenko Neurosurgical Center, Moscow
3Lomonosov Moscow State University
PECULIARITIES OF EEG CHANGES IN THERAPEUTICALLY RADIATED EXPOSURE IN PATIENTS WITH MENINGIOMA OF MEDIOBASAL AREA OF THE RIGHT AND LEFT HEMISPHERE

This study shows the informativeness of the application of analysis of spatial organization coherent EEG connections to evaluate the effectiveness of radiotherapy in patients with tumor lesions of the right and left hemispheres of the brain. There are indentified features of the brain's reactivity to therapeutic radiation exposure, as well as the adaptive strategies of the brain to the effects of radiation therapy.

POSTER SESSION 2

Tuesday, October 8                    17:30 – 18:30
Lecture-hall Фойе 2 этажа

Chair: Prof. KARANDASHEV YAKOV

Neuromorphic computing and deep learning

17. I.S. FOMIN, A.V. BAKHSHIEV
1Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
2Peter the Great St. Petersburg Polytechnic University
Research on Convolutional Neural Network for Object Classification in Outdoor Video Surveillance System

Systems of indoor and outdoor video surveillance today are very widespread. It is possible to use convolutional neural networks to classify objects. In this paper, an attempt is made to combine the results of the work of an existing object detection system in outdoor video surveillance with a convolutional neural networks for objects classification based on Keras and TensorFlow packages.

18. KUZMINA MARGARITA GEORGIEVNA, BASS LEONID PETROVICH, NIKOLAEVA OLGA VASILIEVNA
Keldysh Institute of Applied Mathematics, Moscow
On capabilities of deep convolutional neural network models and multi-agent systems in problems of hyperspectral satellite image processingls

A short overview is presented on successful applications of deep convolutional neural network models and multi-agent systems to the problems of hyper-spectral satellite image processing

19. NIKOLAY FILATOV, VLADISLAV VLASENKO, IVAN FOMIN, ALEKSANDR BAKHSHIEV
1Peter the Great St. Petersburg Polytechnic University
2Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
Application of deep neural network for the vision system of mobile service robot

The solution of object detection task is valuable in many fields of robotics. However application of neural networks for mobile robots requires the use of high – performance architectures with low power consumption. In search of suitable model a comparative analysis of the YOLO and SqueezeDet architectures was conducted. The task of detecting wooden cubes by mobile robot with the camera with the aim of collecting them was solved. A specific dataset was constructed for the training purposes. Applied SqueezeDet neural network has reached precision 89% and recall 82% for IOU ≥ 0.5

Applied neural systems

20. VASILIY E. GAI, IGOR V. POLYAKOV, OLGA V. ANDREEVA
Nizhny Novgorod State Technical University named after R.E. Alekseev
Depth Mapping Method Based on Stereo Pairs

The paper proposes a new method for solving the problem of constructing a depth map based on a stereo pair of images. The result of the depth information recovery can be used to capture the reference points of objects in the film industry when creating special effects, as well as in computer vision systems used on ve-hicles to warn the driver about a possible collision. Proposed method consists in using the theory of active perception at the stage of segmentation and image matching. To implement the proposed method, a software product in the C# language was developed. The developed algorithm was tested on various sets of input data. The results obtained during the experiment indicate the correct operation of the proposed method in solving the problem of constructing a depth map. The accuracy of depth mapping using the described method turned out to be comparable with the accuracy of the methods considered in the review. This suggests that this method is competitive and usable in practice.

21. AGEEV A.V., SMOLIN V.S, SOKOLOV S.M.
1Bauman Moscow State Technical University
2Keldysh Institute of Applied Mathematics, Moscow
Study of the network applicability for a spherical drive control

The problem of using a neural network as a control algorithm for a direct control spherical drive is considered. А spherical actuator mathematical model was built in the process of research, on the basis of which numerical experiments were performed to find the optimal structure of a neural network. By optimality is meant the most accurate device positioning, as well as the possibility of implementation of the network on microcontroller as the control device with limited computing resources.

22. SOKOLOVA E.S, TELNYKHA A.A
1Nizhny Novgorod Research Institute of Radio Engineering
2The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
Clustering and identification of thunderstorm convective cells in mesoscale systems

The problem of clustering and identification of thunderstorm convective cells in mesoscale systems according to their geometrical features is considered. The clustering of thunderstorm convective cells is carried out by meteorological radar reflectivity. An original method is proposed for identifying thunderstorm convective cells, the result of which is the construction of a thunderstorm risk function. The weights of the risk function are found using machine learning algorithms.

23. VLADIMIR B. KOTOV, ALEXANDR N. PALAGUSHKIN, FEDOR A. YUDKIN
Scientific Research Institute for System Analysis, Moscow
Metaphorical Modeling of Resistor Elements

Abstract. The variable resistors changing their resistance during the process of functioning may become the basis for creation of neural networks elements (synapses, neurons, etc.). The processes leading to resistance change are extremely complicated and are not yet amenable to correct description. To master the possibilities of using the variable resistors it is reasonable to use the metaphorical modeling, i.e. to replace a complex physical system with a simple mathematical system with a small number of parameters, reproducing the important features of real system’s behavior. A simple (elementary) resistor element with state determined by a single scalar variable is considered as the modeling unit. The equations describing the change of the state variable are written down. The choices of functions and parameters in equations, as well as the methods of such elements combination with traditional electronic components (fixed resistors, capacitors, diodes, etc.) are discussed. The selection of these functions from a small set and the adjustment of several parameters allow to obtain the characteristics close to real ones. The scheme of measuring the "volt-ampere characteristics" is considered. An example of specific selection of functions determining the resistor element behavior is given.

24. BERGALIYEV T.K., MAZUROV M.E.
1The Moscow Institute of Physics and Technology (State University)
2Plekhanov Russian University of Economics
Effectiveness of government funding in development and implementation of educational neurotechnologies

One of the successful examples of modern digital laboratories for schools in the field of neurotechnologies is the digital laboratory by russian company BiTronics Lab. The results of the implementation of the digital laboratories were considered. A linear regression equation was constructed, which characterizes the dependence of the involved schoolchildren in neurotechnologies using indicators: allocated budgetary funds, the number of events taken to familiarize with neurotechnologies.

25. DMITRIY A. TARKHOV, ALEXANDER N. VASILYEV
Peter the Great St. Petersburg Polytechnic University
The construction of the approximate solution of the chemical reactor problem using the feedforward multilayer neural network

A significant proportion of phenomena and processes in physical and technical systems is described by boundary value problems for ordinary differential equations. Methods of solving these problems are the subject of many works on mathematical modeling. In most works, the end result is a solution in the form of an array of numbers, which is not the best for further research. In the future, we move from the table of numbers to more suitable objects, for example, functions based on interpolation, graphs, etc. We believe that such an artificial division of the problem into two stages is inconvenient. We and some other researchers used the neural network approach to construct the solution directly as a function. This approach is based on finding an approximate solution in the form of an artificial neural network trained on the basis of minimizing some functional which formalizing the conditions of the problem. The disadvantage of this traditional neural network approach is the time-consuming procedure of neural network training. In this paper, we propose a new approach that allows users to build a multi-layer neural network solution without the use of time-consuming neural network training procedures based on that mentioned above functional. The method is based on the modification of classical formulas for the numerical solution of ordinary differential equations, which consists in their application to the interval of variable length. We demonstrated the efficiency of the method by the example of solving the problem of modeling processes in a chemical reactor.

26. INGA G.SHELOMENTSEVA

APPLICATION OF NEURO-FUZZY NEFCLASS SYSTEM FOR RECOGNITION OF ZIEHL-NIELSEN STAINED SPUTUM SMEAR IMAGES

The task of constructing a neuro-fuzzy classification system for medical images is considered on the example of sputum analyzes stained by the Ziehl-Nielsen method using the NEFClass model. The optimal parameters for using the NEFClass-PC program were determined based on the color and shape feature vector.

27. VICTOR V. DERYABIN
Admiral Makarov State University of Maritime and Inland Shipping
The prediction of a vessel’s geodetic latitude by using of a neural network

A new neural network based model predicting a vessel’s latitude is proposed. The model is a one-hidden layered feedforward network satisfying the conditions of the universal approximation. The hidden layer of the network consists of 30 neurons with hyperbolic tangent activation functions. The input vector is a set of variables that defines a vessel’s kinematics on an Earth’s ellipsoid. The output variable is the latitude of a ship. Random number generators are used to form the set of training and validation samples. The Levenberg-Marquardt training algorithm was used to adjust the network parameters. After training, the neural network was tested in 1000 navigation situations. Each situation was a four-hour sailing during which a vessel’s heading and speed, the vector of current remained unchanged. The results of the test have shown that, in the model situations, the prediction quality of the net is not worse than 3.6 angular minutes per four hours. The feature of the proposed model is that it does not require to suppose a certain view of the time-behaviour of the integrated function unlike, e.g., the standard trapezoidal rule does.

28. S.V. ALEKSEEV, A.D. KONEVA, Y.A. SEREDA
N.I. Lobachevsky State University of Nizhni Novgorod
ECG Segmentation by Neural Networks: Error and Correction

In this study we examined the question of how errorcorrection occurs in an ensemble of deep convolutional networkstrained for an important applied problem: segmentation ofElectrocardiograms(ECG). We also explore the possibility ofusing the information about ensemble errors to evaluate a qualityof data representation built by the network

29. DMITRY M. IGONIN AND YURY V. TIUMENTSEV
Moscow Aviation Institute (National Research University)
Semantic segmentation of images obtained by remote sensing of the Earth

In the last decade, computer vision algorithms, including those related to the problem of understanding images, have developed a lot. One of the tasks within the framework of this problem is semantic segmentation of images, which provides the classification of objects available in the image at the pixel level. This kind of segmentation is essential as a source of information for robotic UAV behavior control systems. One of the types of pictures that are used in this case is the images obtained by remote sensing of the earth's surface. A significant number of various neuroarchitecture based on convolutional neural networks were proposed for solving problems of semantic segmentation of images. However, for some reasons, not all of them are suitable for working with pictures of the earth's surface obtained using remote sensing. Neuroarchitectures that are potentially suitable for solving the problem of semantic segmentation of images of the earth's surface are identified, a comparative analysis of their effectiveness as applied to this task is carried out.

Neural network theory, neural paradigms and architectures

30. KOROTKOV A. G.
N.I. Lobachevsky State University of Nizhni Novgorod
Ensemble of neuron-like elements with excitatory couplings

The new model of ensemble of neuron-like elements is suggested. Ensemble has two elements with excitatory couplings. Coupling is modelled by using function of phase of presynaptic element.

31. TELNYKH A.A., NUIDEL I.V., SHEMAGINA O.V.
The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
Biomorphic model of the cortical column for semantic image analysis

Currently, using functional and morphological studies, it has been shown that neurons are combined into vertical columns consisting of six layers. The columns are the main structural and functional units for the processing of sensory information. The paper presents a biomorphic model of the cortical column, which has a topical structure. In the process of learning to objects of recognition, the weights of the ascending connections are activated and amplified, and thus a column is formed. An important feature of the presented model is that the structure of the cortical column has a universal nature, consists of a small number of types of neurons, and allows to perform semantic image analysis in order to detect objects of various types and indicate their characteristic properties. The learning process is visualized. It is possible to visualize the formation of a column, which is a detector of specific stimuli such as a face, a figure, a number, etc.

32. BOLOTOV M.I, OSIPOV G.V.
N.I. Lobachevsky State University of Nizhni Novgorod
Collective dynamics of neuron-like elements coupled by a common pulsed field

The collective dynamics in an ensemble of neuron-like elements coupled by a common pulsed field is studied. The dynamics of the ensemble is analyzed numerically for different values of system parameters. The calculated data, as functions of the synchronicity parameter, lead to a conclusion being drawn on the collective behavior of ensemble elements, the character of evolution of the ensemble field, and the dependence of the behavior of the ensemble field on the synchronicity of ensemble elements.

33. ALENA GUSEVA, GALINA MALYKHINA
Peter the Great St. Petersburg Polytechnic University
Team of neural networks to detect the type of ignition

The article is about the development of a modern multisensory fire system, which has sensors for temperature, CO concentration and smoke concentration.The presence of several different types of sensors allows determine the type of source of ignition, which make possible automatically determine the means of fire extinguishing at the very beginning of the ignition process. The study was carried out on the basis of simulation results obtained in the supercomputer center.It simulated the processes of ignition in the ship’s rooms for various sources of fire: paper, household waste containing plastic, gasoline, alcohol-containing substances and electrical cables.As the study showed, a good result can be obtained with the help of a team of specially organized neural networks. A team of neural networks divided into two levels has been proposed to solve this problem. At the first level, neural networks with partial training are used. At the second level, a probabilistic neural network.The fire system is highly flexible at the hardware level because it has a wireless interface that allows quick reconfiguration. The software of the fire system, in this case also has a high flexibility, allows for simple expansion, contraction or modification of software modules in the conditions of changing sources of ignition in the room.

34. KRASNOV, M.M., SMOLIN V.S.
Keldysh Institute of Applied Mathematics, Moscow
LOCAL CURVATURE AND NEURAL NETWORK APPROXIMATION ACCURACY

The accuracy increasing problem of the one-dimensional functions approximation on a neural network with one hidden layer is considered. It is possible to reduce the root mean square approximation error by more than a hundred times. This is achieved by replacing in a number of places of the network training by the method of reverse propagation of error to other laws of parameters changing, as well as choosing nonlinear functions of neural-like elements and adjusting their properties.

35. ANDREY MIKRYUKOV, MIKHAIL MAZUROV
Plekhanov Russian University of Economics
CLASSIFICATION OF THE TASK DECISION ON THE BASIS COMMITTEE OF ELECTORAL NEURAL NETWORKS

The article considers an approach to the improvement of methods for solving classification problems based on neural network committees (ensembles). To build a neural network committee, a new class of neural networks has been proposed, called electoral neural networks. The scientific novelty of the work is to improve the quality of the solution of the classification problem by the committee of electoral neural networks. The proposed approach can be applied in the decision support subsystems of the situational information security center.

36. CHERNYSHOV ANDREW VADIMOVICH, LOBOV SERGEY ANATOLYEVICH
N.I. Lobachevsky State University of Nizhni Novgorod
Hebb’s and competitive learning of spiking neurons

We study a possibility of Hebb’s and competitive types of learning in spiking neurons in temporal and frequency coding problem. We show that use of Hebb’s learning as pair and triplet STDP rules is sufficient for temporal coding but is not for frequency coding. For the obtaining of selectivity of synapses in case of frequency coding it is necessary to use synaptic forgetting that depends on neuron activity.

SESSION 1

Wednesday, October 9                    10:30 – 11:45
Lecture-hall Ауд. 4.18 (5.17)

Chair: Prof. ALEKSANDR PANOV

Artificial intelligence

37. STAROVEROV ALEKSEY AND ALEKSANDR I. PANOV
1Federal Research Center "Informatics and Control" RAS, Moscow
2The Moscow Institute of Physics and Technology (State University)
Hierarchical actor-critic with hindsight for mobile robot with continuous state space

Multi-level hierarchies have the potential to accelerate learning in sparse reward tasks because they can divide a problem into a set of short horizon subproblems. In order to realize these potential, Hierarchical Reinforcement Learning (HRL) algorithms need to be able to learn the multiple levels within a hierarchy in parallel, so these simpler subproblems can be solved simultaneously. Most famous existing HRL methods that can learn hierarchies are not able to efficiently learn multiple levels of policies at the same time, particularly in continuous domains. To address this problem, we had analyzed the newest existing framework, Hierarchical Actor-Critic with Hindsight (HAC) and test it in simu-lated mobile robot environment.

38. KOPELIOVICH M. V., KOZUBENKO E. A., KASHCHEEV
1Southern Federal University, Rostov-on-Don
2Scientific Research Center of Neurotechnologies, Southern Federal University, Rostov-on-Don
Impact of assistive control on operator behavior under high operational load

This work describes impact of artificial assistant on operator’s performance which is applied to correct operator’s actions in case of unsafe or ineffective behavior. In order for assistive control to be effective a method to evaluate and predict operator performance should be applied. This paper presents a model of operator activity based on histograms of distribution of reaction times to particular stimuli. The model is then applied to the task of monitoring operator activity in a controlled environment, designed to emulate certain actions performed by an aircraft pilot. For each subject an individual behavioral portrait is made. Then, performance changes under high operational load conditions and impact of assistive control are evaluated.

39. KARAVAEV YU. L., EFREMOV K.S., ZVONAREV I.S.
Kalashnikov Izhevsk State Technical University
Optimal trajectories for traning artificial neural network to control the mobile wheeled robot

The work is devoted to the development of an intelligent control system for a wheeled robot, an algorithm is proposed for training an artificial neural network and forming training samples. Samples are formed on the basis of modeling the equations of motion of the multilink wheeled robot, ensuring its movement along trajectories in the form of Euler's elastica, ensuring minimization of control.

40. ALEKSANDR STIKHARNYI, ALEXEY OREKHOV, ARC ANDREEV, YURIY GAPANYUK
Bauman Moscow State Technical University
The Hybrid Intelligent Information System for Music Classification

The article proposes an approach for music classification problem using hybrid intelligent information systems (HIIS). The HIIS consists of two main components: the subconsciousness module and the consciousness module. The subconsciousness module is implemented as a set of binary classifiers based on the LSTM network. The output of the subconsciousness module is the metadata, stored in the metadata buffer. The consciousness module is implemented using decision trees approach. The implementation is based on the CART algorithm from the scikit-learn library. The output of the consciousness module is the predicted class of the music classification problem. The experiments were conducted using custom dataset. The algorithms of three levels of complexity were used for experiments: the logistic regression approach (the simplest model), the multilayer perceptron approach (the model of medium complexity), the HIIS approach (the model of high complexity). The results of the experiments make sure the validity of the proposed HIIS approach.

41. V.B. KOTOV, Z.B. SOKHOVA
Scientific Research Institute for System Analysis, Moscow
On possible consequences of neuroinformatics development

A model of coexistence of human beings and dominating artificial being is considered. There are different variants of system evolution with some of them leading to vanishing of the human population. The conditions are presented, which are necessary for the acceptable future of mankind.

SESSION 2

Wednesday, October 9                    11:45 – 13:15
Lecture-hall Ауд. 4.18 (5.17)

Chair: Prof. LITINSKII LEONID

Neuromorphic computing and deep learning

42. DMITRY VADIMOVICH NEKHAEV AND VYACHESLAV ALEKSANDROVICH DEMIN
National Research Centre "Kurchatov Institute", Moscow
Competitive maximization of neuronal activity in Convolutional Recurrent Spiking Neural Networks

Spiking neural networks (SNNs) are the promising algorithm for specific neurochip hardware real-time solutions. SNNs are believed to be highly en-ergy and computationally efficient. We focus on developing local learning rules that are capable to provide both supervised and unsupervised learn-ing. We suppose that each neuron in a biological neural network tends to maximize its activity in competition with other neurons. This principle was put at the basis of SNN learning algorithm called FEELING.

43. VIKTOR MOSKALENKO, NIKOLAI ZOLOTYKH, GRIGORY OSIPOV
N.I. Lobachevsky State University of Nizhni Novgorod
Deep Learning for ECG Segmentation

We propose an algorithm for electrocardiogram (ECG) segmentation using a UNet-like full-convolutional neural network. The algorithm receives an arbitrary sampling rate ECG signal as an input, and gives a list of onsets and offsets of P and T waves and QRS complexes as output. Our method of segmentation differs from others in speed, a small number of parameters and a good generalization: it is adaptive to different sampling rates and it is generalized to various types of ECG monitors. The proposed approach is superior to other state-of-the-art segmentation methods in terms of quality. In particular, F1-measures for detection of onsets and offsets of P and T waves and for QRS-complexes are at least 97.8%, 99.5%, and 99.9%, respectively.

44. VLADIMIR V. KNIAZ, PETER V. MOSHKANTSEV, VLADIMIR A. MIZGINOV
State Research Institute of Aviation Systems, Moscow
Deep Learning a Single Photo Voxel Model Prediction from Real and Synthetic Images

Reconstruction of a 3D model from a single image is challenging. In this paper, we evaluate the impact of the synthetic data in the dataset on the performance of the trained model. We use a recently proposed Z-GAN model for single-view voxel model prediction as a starting point for our research. We generated a new dataset with 2k synthetic color images and voxel models. We train the Z-GAN model on synthetic, real, and mixed images.

45. YURIY S. FEDORENKO
Bauman Moscow State Technical University
The Simple Approach to Multi-label Image Classification using Transfer Learning

The article deals with the problem of image classification on a relatively small dataset. The training deep convolutional neural net from scratch requires a large amount of data. In many cases, the solution to this problem is to use the pretrained network on another big dataset (e.g. ImageNet) and fine-tune it on available data. In the article, we apply this approach to classify advertising banners images. Initially, we reset the weights of the last layer and change its size to match a number of classes in our dataset. Then we train all network, but the learning rate for the last layer is several times more than for other layers. We use Adam optimization algorithm with some modifications. Firstly, applying weight decay instead of L2 regularization (for Adam they are not same) improves the result. Secondly, the division learning rate on the maximum of gradients squares sum instead of just gradients squares sum makes the training process more stable. Experiments have shown that this approach is appropriate for classifying relatively small datasets. Used metrics and test time augmentation are discussed. Particularly we find that confusion matrix is very useful because it gives an understanding of how to modify the train set to increase model quality.

46. DMITRY A. YUDIN, ALEXANDR V. DOLZHENKO AND EKATERINA O. KAPUSTINA
1The Moscow Institute of Physics and Technology (State University)
2Belgorod State Technological University named after V.G. Shukhov
The Usage of Grayscale or Color Images for Facial Expression Recognition with Deep Neural Networks

The paper describes usage of modern deep neural network architectures such as ResNet, DenseNet and Xception for the classification of facial expressions on color and grayscale images. Each image may contain one of eight facial expression categories: “Neutral”, “Happiness”, “Sadness”, “Surprise”, “Fear”, “Disgust”, “Anger”, “Contempt”. As the dataset was used AffectNet. The most accurate architecture is Xception. It gave classification accuracy on training sample 97.65%, on cleaned testing sample 57.48% and top-2 accuracy on cleaned testing sample 76.70%. The category “Contempt” is worst recognized by all the types of neural networks considered, which indicates its ambiguity and similarity with other types of facial expressions. Experimental results show that for the considered task it does not matter, the color or grayscale image is fed to the input of the algorithm. This fact can save a significant amount of memory when storing data sets and training neural networks. The computing experiments was performed using graphics processor using NVidia CUDA technology with Keras and Tensorflow deep learning frameworks. It showed that the average processing time of one image varies from 4 ms to 30 ms for different architectures. Obtained results can be used in software for neural network training for face recognition systems.

47. KNYAZEVA IS, OHINKO TL, MAKARENKO NG, RYBINTSEV A.S
1Sanct-Petersburg State University
2The Central Astronomical Observatory of the Russian Academy of Sciences at Pulkovo, Saint-Petersburg
Deep-learning approach for McIntosh-based classi cation of solar active regions using HMI and MDI images

Solar active regions (ARs) the primary source of solar flares. There are plenty of studies where the statistical relationship between ARs magnetic field complexity and solar flares are shown. Usually, the complexity of ARs described with different numerical magnetic field parameters and characteristics calculated on top of them. Also, there is well known and widely adapted McIntosh classification scheme of sunspot groups, consists of three letters abbreviation. Solar Monitor's flare prediction system's based on this classification. Up to date, the classification is done manual once a day by the specialist. In this paper, we describe an automatic system based on convolutional neural networks. The best approach for describing the complexity of solar ARs is using so-called magnetogram images, where each pixel contains information about the magnetic field. For neural network training, we used images from two big magnetogram databases (HMI and MDI images) covered together period from 1996 to the 2018 years. Our results show that the automated classification of Solar ARs is possible with a moderate success rate, which allows to use it in practical tasks.

SESSION 3

Wednesday, October 9                    14:15 – 16:00
Lecture-hall Ауд. 4.18 (5.17)

Chair: Prof. VVEDENSKY VIKTOR

Applied neural systems

48. ANDREY LITVIN
Immanuel Kant Baltic Federal University, Kaliningrad
Clinical Decision Support System for Prediction of Infected Pancreatic Necrosis

Infected pancreatic necrosis is associated with high morbidity and mor-tality and is mandatory for surgical or minimally invasive intervention. The aim of this study was to construct and validate the clinical decision support system (CDSS) to predict infected pancreatic necrosis (IPN). All patients who presented with severe acute pancreatitis from January 2011 to December 2017 were reviewed. The study included 398 patients: age (median, range) - 49 (23-77); sex males/females (n) - 302/96; BMI, kg/m2 (median, range) - 28 (24-33); SAPS II at admission, grades - 12.9±3.8; pancreatic necrosis etiology, %: alcohol induced 62.6%, biliary 20.6%, idi-opathic 12.2%, trauma induced 4.6%. All patients in our study (n=398) our divided randomly into three groups. The ANN we constructed was able to determine the presence or absence of infected pancreatic necrosis based on the patients' clinical, laboratory, and radiographic parameters with an AUC= 0,92. High sensitivity is im-portant in the evaluation of IPN. Clinical decision support system was able to predict the development of infected necrotizing pancreatitis with considerable accuracy and outper-formed other clinical risk scoring systems.

49. IGOR ISAEV, SERGEY BURIKOV, TATIANA DOLENKO, KIRILL LAPTINSKIY, AND SERGEY DOLENKO
1Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
2Lomonosov Moscow State University
Diagnostics of Water-Ethanol Solutions by Raman Spectra with Artificial Neural Networks: Methods to Improve Resilience of the Solution to Distortions of Spectra

In this study, we consider adding noise during training of a neural network as a method of improving the stability of its solution to noise in the data. We tested this method in solving the inverse problem of Raman spectroscopy of aqueous ethanol solutions, for a special type of distortion caused by changes in the power of laser pump leading to compression or stretching of the spectrum. In addition, we tested the method on the spectra of real alcoholic beverages.

50. ALEXANDER EFITOROV, SERGEY DOLENKO, TATIANA DOLENKO, KIRILL LAPTINSKIY, AND SERGEY BURIKOV
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Use of wavelet neural networks to solve inverse problems in spectroscopy of multi-component solutions

Wavelet neural networks (WNN) are a family of approximation algorithms that use wavelet functions to decompose the approximated function. They are more flexible than conventional multi-layer perceptrons (MLP), but they are more computationally expensive, and require more effort to find optimal parameters. In this study, we solve the inverse problems of determination of concentrations of components in multi-component solutions by their Raman spectra. The results demonstrated by WNN are compared to those obtained by MLP and by the linear partial least squares (PLS) method. It is shown that properly used WNN are a powerful method to solve multi-parameter inverse problems.

51. DMITRY S. KOZLOV AND YURY V. TIUMENTSEV
Moscow Aviation Institute (National Research University)
Semi-Empirical Neural Network Models of Hypersonic Vehicle 3D-motion Represented by Index 2 DAE

We consider a problem of mathematical modeling and computer simulation of nonlinear controlled dynamical systems represented by differential-algebraic equations of index 2. The solution of the problem is proposed within the framework of a neural network based semi-empirical approach that combines theoretical knowledge of the modeling object with training tools applied to artificial neural networks. We propose particular form semi-empirical models implementing implicit Runge-Kutta integration formulas inside the activation function. The training of the semi-empirical model makes it possible to elaborate on the models of aerodynamic coefficients implemented as a part of it. We present a semi-empirical model that uses as theoretical knowledge the equations of a full model of hypersonic vehicle motion in the specific phase of descent in the atmosphere. The simulation results for the problem of identifying the aerodynamic coefficient, implemented as an ANN-module of a semi-empirical model of the movement of a hypersonic vehicle, are presented.

52. MOHIE MORTADHA ALQEZWEENI
Penza State University
LEARNING RADIAL BASIS FUNCTIONS NETWORKS FOR SOLVING THE PROBLEM OF APPROXIMATION AND PARTIAL DIFFERENTIAL EQUATIONS

The use of radial basis functions networks for meshfree approximation of functions and the solution of boundary value problems for partial differential equations is considered. Algorithms for learning radial basis functions networks based on the Nesterov accelerated gradient method and Levenberg-Marquardt method are proposed, implemented and investigated.

53. SOROKIN DMITRY, NUZHNY ANTON
1
2The Nuclear Safety Institute of the Russian Academy of Sciences
Program of hierarchical search in a textual documentation corpus on the disposal of radioactive waste problem

The program of hierarchical search is presented. The program represents documents as a map of semantically-similar clusters. For each cluster it automatically calculates principal words and builds topics based on keywords for each cluster. This allows the user to search for fragments of text that meets the selected topic. The thematic analysis of the text corpora allows user to detect the presence or absence of topics, assesses the completeness of the information provided.

54. SCHEKALEV A.A., KITOV V.V.
1Lomonosov Moscow State University
2Plekhanov Russian University of Economics
Style transfer with adaptation to the central objects of the scene

Style transfer is a problem of rendering image with some content in the style of another image, for example a family photo in the style of a painting of some famous artist. The drawback of classical style transfer algorithm is that it imposes style uniformly on all parts of the content image, which perturbs central objects on the content image, such as faces or text, and makes them unrecognizable. This work proposes a novel style transfer algorithm which automatically detects central objects on the content image, generates spatial importance mask and imposes style non-uniformly: central objects are stylized less to preserve their recognizability and other parts of the image are stylized as usual to preserve the style. Three methods of automatic central object detection are proposed and evaluated qualitatively and via a user evaluation study. Both comparisons demonstrate higher quality of stylization compared to the classical style transfer method.

SESSION 4

Wednesday, October 9                    16:30 – 18:00
Lecture-hall Ауд. 4.18 (5.17)

Chair: Prof. YAKHNO VLADIMIR

Neurobiology and neurobionics

55. STASENKO SERGEY VICTOROVICH, LAZAREVICH IVAN ALEXANDROVICH, KAZANTSEV VICTOR BORISOVICH
N.I. Lobachevsky State University of Nizhni Novgorod
Generation of quasi-synchronous bursts in the neuron-glial network model

The article discusses the effect of the activity of glial cells (astrocytes) on the synaptic dynamics of interneuronal contacts and network dynamics. The model is a network of synaptic-coupled pulsed neurons, where the dynamics of synaptic connections is modulated by the action of a gliatransmitter. In the network of synaptically connected neurons, the effect of spontaneous transitions to the mode of generation of quasi-synchronous burst discharges, typical of the epileptiform activity of the neural networks of the brain, was detected.

56. ANTON V. CHIZHOV, ELENA G. YAKIMOVA AND ELENA Y. SMIRNOVA
1Ioffe Physical Technical Institute, Russian Academy of Sciences, St Petersburg
2Pavlov Institute of Physiology of the Russian Academy of Sciences, St.Petersburg
Direction Selectivity Model Based On Lagged And Nonlagged Neuron

Direction selectivity (DS) of visual cortex neurons is modelled with a filter-based description of retino-thalamic pathway and a conductance-based population model of the cortex as a 2-d continuum. The DS mechanism is based on a pinwheel-dependent asymmetry of projections from lagged and non-lagged thalamic neurons to the cortex. The model realistically reproduces responses to drifting gratings. The model reveals the role of the cortex in sharpening DS, keeping interneurons non-selective.

57. O.E. DICK
Pavlov Institute of Physiology of the Russian Academy of Sciences, St.Petersburg
WAVELET AND RECURRENCE ANALYSIS OF EEG PAT-TERNS OF SUBJECTS WITH PANIC ATTACKS

The task of analyzing the reactive patterns of electroencephalogram (EEG) in individuals with panic attacks before and after non-drug therapy associated with the activation of artiphicial stable functional connections of the human brain is considered.The quantitative measures of the photic driving reaction for the suggested frequency are estimated by increasing the energy of the wavelet spectrum during the photostimulation and the parameters of the joint recurrence plot of the light stimulus and EEG pattern.

58. SMIRNITSKAYA I.A.
Scientific Research Institute for System Analysis, Moscow
The impact of dorsal and ventral visual streams on the control of grasping

Since 1982 Ungerleider and Mishkin’s paper about the different roles of dorsal and ventral visual streams, the first as “where” and the last as “what”, there is no consensus, what these pathways really do and are they really exist. In this review the contribution of parietal, premotor and prefrontal cortical regions in the control of grasping in the context of the existence of two visual streams is discussed. There is evidence that each of the two streams consists of two subdivisions. The roles of the subdivisions in control of grasping such as: the remembering of the features of object for grasping, the calculation of value of the object for grasping, the control of the movement’s precision, the retention of the movement’s goal in working memory, and so on, are analyzed. The separate problem is the coherency of the execution of all this tasks. The presence of multiple interconnections between the subdivisions of two streams gives the full opportunity for their synchronization.

59. S.D. GLYZIN AND M.M. PREOBRAZHENSKAIA
P.G. Demidov Yaroslavl State University
Two delay-coupled neurons with a relay nonlinearity

The models of association of two coupled neural oscillatorswith synaptic coupling are considered. An important feature of this as-sociation is an additional delay in the chain of connections. The systemof relay differential-difference equations was chosen as the mathematicalmodel of this association. The synaptic connection was modeled based onthe modified idea of fast threshold modulation (FTM). The delay allowsobtaining new effects which is an essential complication of the systemdynamics, and an appearance of coexisting special form attractors. Weshow that there coexist asymptotically orbitally stable solutions withsummaryN∈INspikes on a period. Moreover, the first oscillator hasmspikes, and the second one has N−m (m= 1,2,...,N−1) spikes on aperiod. We conclude that the additional delay leads to an accumulationof coexisting attractors with a given number of spikes on a period.

60. I.V. NUIDEL, A.V. KOLOSOV, S.A. POLEVAIA, V.G. YAKHNO
1The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
2N.I. Lobachevsky State University of Nizhni Novgorod
Mathematical model of the dynamics of alpha-rhythm eeg in rhythmic photostimulation for neurofeedback

We introduce neuro-information solutions to the problem of personalizing neurofeedback [1]. Computational experiments have shown the effectiveness of the phenomenological mathematical model of an elemental thalamocortical cell for describing the dynamics of the EEG alpha rhythm during rhythmic photostimulation during neurobiological control. The substantiation of the possibility of using the model as an adaptive simulator of individual EEG rhythmic patterns providing the generation of personalized protocols for neurofeedback optimizing brain functions is given.

61. I.V. MAKUSHEVICH, N.G. BIBIKOV
N.N. Andreyev Acoustics Institute, Moscow
USING THE HURST INDEX FOR THE STUDIING OF THE SPONTANEOUS ACTIVITY DYNAMIC CHANGES IN THE AUDITORY NEURONS

The every single neuron in the brain is a dynamic system that is in continuous change with elements of self-similarity in the time domain. It is natural to study such activity by the methods of fractal analysis. In this work, an approach based on the dynamic analysis of the variability of the time point process using the Hurst index was applied to analyze the spontaneous activity of the neurons in the auditory system of the grass frog. This approach allowed showing that spontaneous activity changes its variability, moving from random fluctuations to trend dependence and vice versa.

SESSION 5

Thursday, October 10                    12:00 – 13:00
Lecture-hall Ауд. 4.18 (5.17)

Chair: Prof. TEREKHOV SERGE

Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling

62. * ANTON A. PASHKOV, IVAN S. DAKHTIN
South Ural State University (National Research University), Chelyabinsk
Consistency across functional connectivity methods and graph topological properties in EEG sensor space

One of the most widely used topological properties of brain graphs is small-worldness. However, different functional connectivity methods can generate quantitatively different results, particularly when they are applied to EEG sen-sor space. In this manuscript, we sought to evaluate the consistency of values derived from pairwise correlation between selected functional connectivity methods. We showed that the alpha band yielded maximal values of correlation coefficients between small-worldness indices obtained with different methods. In contrast, delta and gamma bands demonstrated the least consistent results.

Neuromorphic computing and deep learning

63. * TROFIMOV A.G., BOGATYREVA A.A.
National Research Nuclear University (MEPhI), Moscow
A method of choosing a pre-trained convolutional neural network for transfer learning in image classification problems

A method of choosing a pre-trained convolutional neural network (CNN) that is more suitable for transfer learning on the new image classification problem is proposed. The method can be used for quick estimation of which of the CNNs trained on the ImageNet dataset images (AlexNet, VGG16, VGG19, GoogLeNet, etc.) will be the most accurate after its fine tuning on the new sample of images. It is shown that there is high correlation between the characteristics of the features obtained at the output of the pre-trained CNN’s convolutional part and its accuracy on the test sample after fine tuning. The proposed method can be used to make recommendations for researchers who want to apply the pre-trained CNN and transfer learning approach to solve their own classification problems and don’t have sufficient computational resources and time for multiple fine tunings of available free CNNs with consequent choosing the best one.

64. * KHAYROV E.M., MALSAGOV M.YU., KARANDASHEV I.M.
Scientific Research Institute for System Analysis, Moscow
Post learning quantization of deep neural networks weights

In this paper, we consider a weights quantization task designed to decrease file size of pre-trained neural network without re-training. We analyse methods of quantization that allows to quantize weights by uniform or exponential distribution and provide their comparative characteristics. Besides, we show the results of our quantization on such networks, as VGG16, VGG19 and ResNet50.

65. * TEREKHOV VALERY IGOREVICH, SABIROV ARTUR AIRATOVICH, CHERNENKY IVAN MIKHAILOVICH, CHERNENKY VALERY MIKHAILOVICH
Bauman Moscow State Technical University
SAR IMAGES PREPROCESSING FOR ICE SITUATION ANALYSIS USING DEEP LEARNING METHODS

The paper studies the effectiveness of specialized algorithms for preprocessing images obtained by synthetic aperture radar (SAR) in order to improve their quality. Images pre-processed by deep learning methods are used to recognize the boundaries of ice cover, icebergs and thin ice floes. The results obtained in the training of deep neural network show that the preprocessing of SAR images plays a decisive role in the analysis of the ice situation.

SESSION 6

Thursday, October 10                    14:00 – 16:00
Lecture-hall Ауд. 4.18 (5.17)

Chair: Prof. TEREKHOV SERGE

Neurobiology and neurobionics

66. * M.A. ROZHNOVA, V.B. KAZANTSEV, E.V. PANKRATOVA
N.I. Lobachevsky State University of Nizhni Novgorod
Brain extracellular matrix impact on neuronal fi ring reliability and spike-timing jitter

In this work, the role of the brain extracellular matrix (ECM) in signal processing by a neuronal system is examined. For excitatory postsynaptic currents in the form of Poisson signal, we study the changes of the interspike intervals duration, spike-timing jitter and coefficient of variation in the presence of a background noise with varied intensity. Without ECM, noise-delayed spiking phenomenon reflecting worsening of both reliability and precision of signal processing is revealed. It is shown that, the ECM-neuron feedback mechanism allows enhancing the robustness of the neuronal firing in the presence of noise.

Applied neural systems

67. * DARIA A. EROSHеNKOVA, VALERI I. TеRеKHOV, DMITRY R. KHUSNеTDINOV AND SERGEY I. CHUMACHеNKO
Bauman Moscow State Technical University
Automated Determination of Forest-Vegetation Characteristics with the use of a Neural Network of Deep Learning

The article proposes a method of automated solution for determining the species composition, stock coefficient and other characteristics of forest plantations with the use of deep learning. The analysis of existing approaches and ways of forest inventory, which include the use of LiDAR systems and machine learning methods, is carried out. An algorithm is proposed for solving this problem and features of its implementation are given. The problem of combining the data of a “dense cloud” and a lidar survey is considered, a possible solution is proposed. The problem of segmentation of tree crowns among many other objects in this data is also considered. For the segmentation of crowns, it is proposed to use the PointNet neural network of deep learning, which allows segmentation of objects by submitting a point cloud to the input. The description of the architecture and the main features of the neural network use are briefly given. The path of further research is determined.

68. * TELYATNIKOV L.S., KARANDASHEV Y.M.
Scientific Research Institute for System Analysis, Moscow
Linear prediction algorithms for lossless audio compression

The problem of lossless audio data compression is considered using linear interpolation algorithms: LPC, FLPC and their combination Wise-LPC. In addition to interpolation, problems of optimal coding and finding the optimal sampling window are investigated. It is shown that the algorithm Wise-LPC allows to improve the compression of the audio signal by 1-5% compared with the classical LPC and FLPC. The prediction error has the Laplace distribution, with an increase in the width of the window, its dispersion smoothly decreases and reaches “saturation”.

69. * I.M. GADZHIEV
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Methods of Joining Classes in the Algorithm of Adaptive Construction of Hierarchical Neural Network Classifiers

The object of this study is the hierarchical neural network classifier (HNNC), which uses a special algorithm to solve the problem of multi-class classification. During the operation of each HNNC node, similar classes are combined into groups; the subsequent division of groups is carried out by recursive application of the algorithm at the following levels of the hierarchy. In this paper we consider the problem of accidental merging of classes, and compare different algorithms of joining classes into groups.

70. * FOMIN VLADIMIR, TEREHOV VALERY, GUNKIN MIKHAIL
Bauman Moscow State Technical University
Predictive model of demand for theater tickets using machine learning methods

The article analyzes the demand and sales of theater tickets using various methods of machine learning. Studies have shown that using the methods of machine learning can significantly improve the sales forecast of theater tickets for performances. The result of the work is the study of six predictive models, from which a model based on a recurrent LSTM network was selected, giving the smallest prediction error. In conclusion, areas of further research are given.

71. * STENKIN DMITRY ALEKSANDROVICH, VLADIMIR IVANOVICH GORBACHENKO
Penza State University




Neural network theory, neural paradigms and architectures

72. * DEMIDOVSKIJ A.V.
Higher School of Economics, Nizhny Novgorod Branch
TOWARDS AUTOMATIC MANIPULATION OF ARBITRARY STRUCTURES IN CONNECTIVIST PARADIGM WITH TENSOR PRODUCT VARIABLE BINDING

Building a bridge between symbolic and connectionist level of computations requires constructing a full pipeline that accepts symbolic structures as an input, translates them to distributed representation, performs manipulations with this representation equivalent to symbolic manipulations and translates it back to the symbolic structure. This work proposes neural architecture that is capable of joining two structures which is an essential part of structure manipulation step in the connectionist pipeline. Verification of the architecture demonstrates scalability of the solution, a set of advice for engineering practitioners was elaborated.

SESSION 7

Friday, October 11                    12:00 – 13:00
Lecture-hall Физтех.Био, ауд. 107

Chair: Prof. PARIN SERGEY

Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling

73. Y.R. MURATOV, M.B. NIKIFOROV, A.S. TARASOV, A.M. SKACHKOV
Ryazan State Radio Engineering University
VIDEO-COMPUTER TECHNOLOGY DEFINITIONS OF REDUCTIONS CONCENTRATION ATTENTION DRIVER

Fatigue of the person making (executing) control, management or decision-making, and him attention concentration decreasing on the object can lead to critical consequences. The most effective person physiological state control is video control based on the eyes state analysis. The algorithm on based on convolutional neural network and its hardware implementation, providing search of the face in the image, eye detection and analysis of their state by "open-closed» principle is proposed.

74. POLEVAIA SOFIA ALE[ANDROVNA, BULANOV NIKITA ALEXEEVICH, PARIN SERGEY BORISOVICH
1N.I. Lobachevsky State University of Nizhni Novgorod
2Higher School of Economics, Nizhny Novgorod Branch
COMPUTER TECHNOLOGIES FOR SCREENING, DIAGNOSTICS AND DIGITAL DISPLAY OF COGNITIVE DISTURBANCES

Within the framework of the concept of digital psychophysiology, modern Internet-oriented computer technologies are considered, providing objective monitoring of the current state of cognitive functions, as well as screening, diagnosis and correction of cognitive impairments. A brief description of the functionality of the ApWay.ru web platform, developed with the participation of the authors for this purpose, is given, and examples of its testing and use are given.

75. TEREKHIN S.A., SIDOROV K.V.
Tver State Technical University
Indexes attractors of the biomedical signals characterizing of the emotional responses of human

The possibility of the analysis dynamics of the emotional reaction of the person on the basis of nonverbal information is investigated. The technique and bioengineering system for monitoring emotional reactions are described. The attractors were reconstructed from the obtained EEG patterns. The estimation of indicators attractors at different stages of emotional stimulation is carried out. The influence of the noise factor of the initial data is analyzed. According to the results of the EEG analysis, the general trend in the development of the dynamics of emotions is observed.

76. VADIM L. USHAKOV, VYACHESLAV A. ORLOV, YURI I. KHOLODNY, SERGEY I. KARTASHOV, DENIS G. MALAKHOV, MIKHAIL V. KOVALCHUK
1
2National Research Centre "Kurchatov Institute", Moscow
The role of stem structures in the vegetative reactions based on fMRI analysis

This study was aimed at studying the role of brain stem structures in vegetative responses upon presentation of personally significant stimuli (own name) using the functional MRI method. The subjects, based on the data of the MR-compatible polygraph, were divided into approximately three groups with different severity of vegetative reactions to personally significant stimuli: 100% skin galvanic reaction - 7 subjects, mixed skin-galvanic reaction and changes in heart rate - 6 subjects and low autonomic reactivity - 5 subjects. The obtained statistical maps of brain neural network activities showed activation of the brain stem structures upon presentation of personally significant stimuli in the mixed skin-galvanic response group and heart rate changes, the complete absence in the group with low autonomic reactivity and low in the group with 100% skin-galvanic response. The effectiveness of using an MR-compatible polygraph for the selection of fMRI data in statistical analysis was shown.

SESSION 8

Friday, October 11                    14:00 – 15:15
Lecture-hall Ауд. 4.18 (5.17)

Chair: Prof. PARIN SERGEY

Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling

77. MEILIKHOV E.Z., FARZETDINOVA R.M.
National Research Centre "Kurchatov Institute", Moscow
BISTABLE PERCEPTION OF AMBIGUOUS IMAGES – ANALYTICAL MODEL

Watching an ambiguous image leads to the bistability of its perception, that randomly oscillates between two possible interpretations. We sugest the analytical model of the phenomenon considering the "walk" of the system along its energy landscape under the action of the "noise". The model predicts the hysteretic behavior of the perception with the logarithmic dependence of the hysteresis width on the period of cyclic sweeping the parameter, controlling the perception.

78. SOKHOVA Z.B., RED’KO V.G.
Scientific Research Institute for System Analysis, Moscow
Comparison of Two Models of a Transparent Competitive Economy

The article compares two models of a transparent competitive economy. In both models, the interaction between investors and producers is considered. In the first model, the producers do not take into account their own contributions to their capitals, in the second model, the producers take into account their contributions to their own capitals, i.e. the producers themselves play the role of investors. The analysis of these two models by computer simulation was performed. It is shown that in the first model when the producers give half of their profits to investors, the capital in the producer community is redistributed by investors more efficient-ly.

79. VICTOR L. VVEDENSKY, KONSTANTIN G. GURTOVOY, MIKHAIL V. SOKOLOV, MIKHAIL O. MATVEEV
1National Research Centre "Kurchatov Institute", Moscow
2Child Technopark of Kurchatov University, Moscow
3Institute of Linguistics RAS, Moscow
Ordering of words by the spoken word recognition time

We measured the time needed to recognize spoken words in a group of 12 subjects. We see that recognition time varies for different words with the same sound duration and they can be ordered from the word perceived most quickly to the “slowest” one. Every subject “generates” his own ordered list of 24 words used. The individual lists are similar to some extent, so that the robust average list can be compiled. Presumably, it reflects distribution of the word representations in the cortex and the time required to retrieve any word depends on its position.

80. V.A. ANTONETS
N.I. Lobachevsky State University of Nizhni Novgorod
Simulation of intuitive evaluation of unlike semantic objects

In this article two quantitative models of the mechanisms of human intuitive decision making when buying a product (goods) are constructed. The first model demonstrates a possibility of quantitative description of a subjective evaluation of a product as a complex semantic object. The second model, constructed by the example of forming a product basket from the offered range of products, quantitatively describes a mechanism of comparison and subjective evaluation of a group of semantically unlike objects. The models have been constructed based on the results of Ch. Osgood, S. Stevens, and D. Kahneman.

81. ALEXANDROVA N. SH., ANTONETS V. A., NUIDEL I. V. SHEMAGINA O. V., YAKHNO V. G.
1Sprachbrücke-Hamburg e. V.
2N.I. Lobachevsky State University of Nizhni Novgorod
3The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
MODELING OF SOME PECULIARITIES OF THE FORMATION OF THE NATURAL BILINGUALISM

The variants of dynamic modes of teaching children two languages on the basis of the proposed mathematical model. The modes in the model correspond to the experimental data.

SESSION 9

Friday, October 11                    15:15 – 16:45
Lecture-hall Ауд. 4.18 (5.17)

Chair: Prof. KAZANOVICH YAKOV

Neural network theory, neural paradigms and architectures

82. SUSAN YU. GORDLEEVA, YULIA A. LOTAREVA, MIKHAIL I. KRIVONOSOV, ALEXEY A. ZAIKIN, MIKHAIL V. IVANCHENKO AND ALEXANDER N. GORBAN
1N.I. Lobachevsky State University of Nizhni Novgorod
2University College London
3University of Leicester, Great Britain
Astrocytes organize neural associative memory

We investigate one aspect of the functional role played by astrocytes in neuron-astrocyte networks present in the mammal brain. To highlight the effect of neuron-astrocyte interaction, we consider simplified networks with bidirectional neuron-astrocyte communication and without any connections between neurons. We show that the fact, that astrocyte covers several neurons and a different time scale of calcium events in astrocyte, alone can lead to the appearance of neural associative memory.

83. TARKOV MIKHAIL SERGEEVICH
Rzhanov Institute of Semiconductor Physics, Siberian Branch of Russian Academy of Sciences, Novosibirsk
Building Neural Network Synapses Based on Binary Memristors

The design of an analog multilevel memory cell based on the use of resistors and binary memristors is proposed. This design provides a greater number of resistance levels with a smaller number of elements than the well-known multilevel memory devices. The cell is designed to set the synapse weights in hardware-implemented neural networks. The neuron vector of weights can be represented by a crossbar of binary memristors and a resistor set. An algorithm is proposed for mapping the neuron weight to the proposed multilevel memory cell. The proposed approach is illustrated by the construction example of a neuron for partitioning a set of vectors into two classes.

84. LEONID LITINSKII AND INNA KAGANOVA
Scientific Research Institute for System Analysis, Moscow
Bimodal coalitions and the Hopfild model

We give an account of the Axelrod – Bennet model that describes formation of a bimodal coalition. We present its initial formalism and applications. Then we reformulate the problem in terms of the Hopfield model. This allows us to analyze a system of two homogeneous groups of agents, which interact with each other. We obtain a phase diagram describing the dependence of the bimodal coalition on external parameters.

85. ALEXANDER E. LEBEDEV*, KSENIYA P. SOLOVYEVA AND WITALI L. DUNIN-BARKOWSKI
1Scientific Research Institute for System Analysis, Moscow
2The Moscow Institute of Physics and Technology (State University)
The large-scale symmetry learning applying Pavlov principle

Symmetry detection task in the domain of 100-dimension binary vectors is considered. This task is characterized by practically infinite number of training samples. We train an artificial neural network with binary neurons to solve the symmetry detection task. Weight changing of hidden neurons is performed according to Pavlov Principle. In the presence of error, synaptic weights are adjusted considering a matrix of random weights. After training on a relatively small number of data samples our network obtained generalization ability and detects symmetry on data not present at the training set. The obtained averaged percentage of correct recognition of our network is better than those of classic perceptron with fixed weights of synapses of neurons of hidden layer. We also compare performance of different modifications of the architecture including different number of hidden layers, different number of neurons in hidden layer, different number of neurons` synapses.

86. A.A.BRYNZA, M.O.KORLYAKOVA
Bauman Moscow State Technical University Kaluga Branch
An approach to predicting the behavior of a dynamic system beyond the boundaries of learning

The problem of predicting the behavior of a complex dynamic system is considered. The analysis of approaches that allow in conditions of limited information about the nature of behavior, as well as parametric uncertainty, to predict the behavior of systems with high accuracy, for situations in which the value of the control parameters goes beyond the boundaries of the training set is carried out. The forecast results were evaluated, the corresponding graphs were presented, and conclusions were made.

87. KISELEV MIKHAIL
The Chuvash state university named after I. N. Ulyanov
Chaotic Spiking Neural Network Connectivity Configuration Leading to Memory Mechanism Formation

Chaotic spiking neural network serves as a main component (a “liquid”) in liquid state machines (LSM) – a very promising approach to application of neural networks to online analysis of dynamic data streams. The LSM ability to recognize complex dynamic patterns is based on “memory” of its liquid component – prolonged reaction of its neural network to input stimuli. A generalization of LSM called self-organizing LSM (LSM including spiking neural network with synaptic plasticity switched on) is studied. It is demon-strated that memory appears in such networks under certain locality conditions on their connectivity. Genetic algorithm is utilized to determine parameters of neuron model, synaptic plasticity rule and connectivity optimal from point of view of memory characteristics.



Российская нейросетевая ассоциация Российская академия наук Министерство образования и науки Российской Федерации МФТИ НИЯУ МИФИ НИИСИ РАН МАИ Институт перспективных исследований мозга МГУ
AIRI iLabs Приоритет 2030