Sections

Neuroinformatics - 2021


SESSION 1

Monday, October 18                    17:00 – 19:30
Lecture-hall Актовый зал

Chair: Prof. DOLENKO SERGEY

Artificial intelligence

1. * BOLSHAKOV V.E., ALFIMTSEV A.N., SAKULIN S.A., BYKOV N.V.
Bauman Moscow State Technical University
Deep Reinforcement Ant Colony Optimization for Swarm Learning

Until recently the implementation of reinforcement learning algorithms was limited by the astronomical number of the real environment possible states. Therefore these algorithms were implemented only for simple tasks in a tabular grid worlds. Deep learning technologies have expanded the use of these algorithms for multidimensional and complex virtual environments – computer video games. Moreover, modern deep learning in general and multi-agent reinforcement learning in particular are actively porting classical machine learning algorithms to neural network architecture. Combining these trends, we propose a new deep reinforcement learning algorithm based on the traditional ant colony optimization algorithm for solving the problem of cooperative homogeneous swarm learning. The algorithm shapes the collective behavior of a decentralized system, consisting of a set of independent homogeneous agents, locally interacting with each other and with the environment. The developed algorithm also represents an alternative stigmergic approach to the implementation of the leading multi-agent technology of centralized learning with decentralized execution. Local and often random interactions in the process of centralized learning lead to the emergence of agent swarm collective behavior, uncontrolled by individual agents in the process of real operation. An experimental study of the advantages of the developed algorithm was carried out in a virtual environment of the StarCraft Multi-Agent Challenge.

Neuromorphic computing and deep learning

2. * BALASHOV I.S., CHEZHEGOV A.A., CHIZHOV A.S., GRUNIN A.A., FEDYANIN A.A.
Lomonosov Moscow State University
Neuromorphic photoelectric synapses based on metal oxide nanocrystallites.

Artificial synapse is the key neuromorphic computational elements. We reports on the implementation of photoelectric synapses based on semiconductor nanocrystallites. The temporal characteristics of synaptic plasticity, short-term memory STM and long-term memory LTM, as well as additional channels of gas-assisted and temperature-assisted conduction changes, acting as neuromodulators, were investigated.

3. MALSAGOV M. YU., MIKHALCHENKO E.V., KARANDASHEV I.M., NIKITIN V.F.
Scientific Research Institute for System Analysis, Moscow
APPLICATION OF UNET ARCHITECTURE IN THE PROBLEM OF COMBUSTION OF HYDROGEN WITH OXYGEN

Traditionally, neural network models of the UNET type have been used for automatic image-to-image conversion, etc. In this work, it is shown that such an architecture can be successfully applied to the problem of modeling chemical combustion processes described by a rigid system of ordinary differential equations. Using it, we were able to train a compact model that can approximate changes in the concentration of substances in a mixture during chemical reactions with a high degree of accuracy.

Applied neural systems

4. * KASATKIN N.E., YUDIN D.A.
The Moscow Institute of Physics and Technology (State University)
Real-time approach to neural network-based disparity map generation from stereo images

Generating depth maps using mono- or stereo- images is a topic of active research. This paper is dedicated to study of different methods of depth and disparity maps generating. It includes analysis of existing methods of depth maps generating and investigation and improvement of the real-time neural-network based method AnyNet. Our approach AnyNet-M related to the models which are make prediction stage-by-stage and increase quality of disparity estimation over time. Firstly, proposed method was tested on KITTI Stereo Dataset and custom OpenTaganrog dataset with images of 1242×375 resolution. The method was proved to be real-time with approximately 38 FPS using one TeslaV100 GPU. Secondly, quality of proposed model was tested and reached 5% of Average 3-Pixel Error on KITTI dataset. Finally, it was integrated to Robotic operation system for further use as part of the navigation systems of unmanned vehicles.

5. SAKHAROVA E.K., NURLYEVA D.D., FEDOROVA A.A., YAKUBOV A.R., KANEV A.I.
Bauman Moscow State Technical University
Issues of Tree Species Classification from LiDAR Data using Deep Learning Model

Trees inventory is an important task for the ecology, calculation of carbon dioxide absorption and city arrangement reasons. But different types of trees have their own characteristics, and the problem of their automatic classification is urgent. Various sensors are currently used to meet this challenge but the most often used is LiDAR. Different classification methods are applied, including deep learning. One of the modern deep learning models for point cloud data is PointNet. Therefore, the authors of this work applied it to classify tree species. Russia has its own specific set of tree species and the authors collected a dataset for the research containing tree species typical for this area and carried out its labeling. The results showed good capabilities of PointNet for tree species classification but revealed the problem of insufficient data for high training accuracy and difficulties with manual marking of instances based on point cloud data. The paper gives recommendation for overcoming these issues.

6. * ISHKOV DENIS OLEGOVICH, TEREKHOV VALERY IGOREVICH
Bauman Moscow State Technical University
Knowledge Distillation of Convolutional Neural Network Models on Self-Labeled Data for Automatic Text Captcha Traversal

While most of the existing papers have investigated fixed-length CAPTCHA recognition, the authors propose to apply knowledge distillation to imitate the predictions of recurrent-convolutional neural networks. Such models have proven themselves well in the problem of predicting the dynamic length of characters in an image. The influence of the size and complexity of the dataset on the quality of the model is studied. The analysis of model errors allowed us to make recommendations on ways to counteract automatic recognition.

7. * TEREKHOV V.I., ZABELINA V.A., SAVCHENKO G.A., CHUMACHENKO S.I,
Bauman Moscow State Technical University
Classification Of Tree Species By Trunk Image Using Conventional Neural Network And Augmentation Of The Training Sample Using A Telegram-Bot

This paper considers the problem of creating a model of a convolutional neural network for recognizing tree species from the image of a trunk for ground-based lidar taxation of forest stands. To increase the probability of recognition, it is proposed to use a telegram bot for augmentation of the training set. Training, selection and comparison of convolutional neural network models was performed. A telegram bot has been created that allows you to automate the collection of images of the training sample. The study opens a cycle of works on modeling the carbon balance of forest plantations.

8. SOROKA A.A., TROFIMOV A.G.
National Research Nuclear University (MEPhI), Moscow
Cross-Modal Transfer Learning for Image and Sound

Recently the research on transfer learning between similar domains has become increasingly common. However, the fields of cross-domain and cross-modal knowledge transfers are more complicated and have been studied less. We propose the new transfer learning strategy between tasks on essentially different domains called as cross-modal transfer learning and consider its ideas and the algorithm. The key element of cross-modal transfer pipeline is cross-modal adapter, i.e. a neural network that transforms the target domain features to the source domain features that can be efficiently processed by a pre-trained neural network. In the experiments the dataset ImageNet and audio dataset ESC-50 are chosen as source domain and target domain respectively. It is shown that a fairly simple neural cross-modal adapter makes it possible to achieve high classification accuracy on target domain using the knowledge obtained by pre-trained neural network on the source domain. Our experiments also show that cross-modal transfer learning noticeably reduces the training time in comparison with the building target model “from scratch”.

9. E S UDEGOVA, I G SHELOMENTSEVA AND S V CHENTSOV
1
2Siberian Federal University, Krasnoyarsk
Optimizing Convolutional Neural Network Architecture for Microscopy Image Recognition for Tuberculosis Diagnosis

Globally, tuberculosis (TB) is the leading infectious killer in the world before pandemia. This paper presents the result of optimizing convolutional neural network architecture for the detection of acid-fast stained TB bacillus. The experimental set contains the segmentation results of microscopy images of the patient's sputum stained by the Ziehl–Neelsen method. The authors constructed an experimental algorithm for optimizing the original convolutional neural network model, including optimizing the model dimension, data augmentation, adjusting the model parameters, and improving regularization. The authors built few models of convolutional neural networks (CNN) models to recognize TB bacillus, which showed the maximum value of metrics in the experiment.

10. MARIA O. TARAN, GEORGIY I. REVUNKOV, AND YURIY E. GAPANYUK
Bauman Moscow State Technical University
Generating a Summary of a Court Act Based on an Improved Text Fragment Extraction Module

The implementation of two software modules: “text fragment extraction module” and “automatic summarization module,” aimed at the generation of a summary of a court act is considered. These modules are parts of a hybrid intelligent infor-mation system for analysis of the judicial practice of arbitration courts, the “text fragment extraction module” is a part of the system’s subconsciousness. In con-trast, the “automatic summarization module” is a part of the system’s conscious-ness. The “automatic summarization module” allows generating summaries of different lengths depending on the information that the lawyer need. The input of the module is a text with pre-classified paragraphs. The output of a module can be a dictionary for use in other modules of the system. The improved version of the “text fragment extraction module” uses a hybrid approach for text extraction. It consists of five submodules: submodule for creating templates, clustering sub-module, classification submodule 1 (based on LGBM), classification submodule 2 (based on the convolutional neural network), classification submodule 2 (based on logistic regression implemented using the PyTorch library). The improved version of the “text fragment extraction module” shows an accuracy of 85.83%, F-measure 85.38%.

POSTER SESSION 1

Tuesday, October 19                    14:00 – 15:00
Lecture-hall Актовый зал

Chair: Prof. DOROFEEV VLADISLAV PETROVICH

SESSION 2

Tuesday, October 19                    15:00 – 16:30
Lecture-hall Актовый зал

Chair: Prof. RATUSHNYAK ALEXANDER SAVELYEVICH

Adaptive behavior and evolutionary modelling

11. * TYKHONOV I.V., ZHDANOV A.A.
1The Moscow Institute of Physics and Technology (State University)
2Lebedev Institute of Precision Mechanics and Computer Engineering, Russian Academy of Sciences, Moscow
Algorithm of knowledge extraction in vision systems for real-time applications

This paper considers the construction of a control system in the task of tracking - predicting the class of an object and its location in the image. The solution involves unsupervised learning, without prior image marking - automatic real-time formation of images and accumulation of knowledge about their influence on the process of further recognition and prediction. The system is modeled based on the real data.

Neurobiology and neurobionics

12. S.A. POLEVAYA, E.V. EREMIN, N.A. BULANOV, N.L. KOROTKOVA, L.V. SAVCHUK, A.I. FEDOTCHEV, S. B. PARIN
1Privolzhsky Research Medical University
2Nizhny Novgorod State Medical Academy
3Higher School of Economics, Nizhny Novgorod Branch
4Institute of Cell Biophysics, Pushchino, Moscow region
5N.I. Lobachevsky State University of Nizhni Novgorod
HEART RATE VARIABILITY AS A DISPLAY OF ADAPTATION AND DESADAPTATION IN THE CONDITIONS OF EVERYDAY LIFE

The paper presents the results of the analysis of the database obtained using the technology of event-related telemetry of the heart rate. Rhythmographic signs of adaptation and maladjustment were revealed upon presentation of cognitive and physical loads: the Toulouse-Pieron correction test and the orthostatic test in children and adults. It was found that the characteristic signs of an increase in the adaptive potential are: an increase in the total power of the heart rate variability spectrum, the power of the high-frequency region of the spectrum, functional reserve, and an increase in R-R intervals.

13. I.A. SMIRNITSKAYA
Scientific Research Institute for System Analysis, Moscow
The thalamic nucleus classification in relation to their engagement in the correction of initial movements

An approach to creating a model of hierarchical control of movements in the brain is proposed, based on the idea of a single center for the initiation of movements. Taken the assumption of initial random movement and then its correction, the role of individual thalamic nuclei is evaluated. Different thalamic nuclei take part in distinct phases of movement correction. The role of inhibitory structures is emphasized. The source of initial random movements postulated to be the pedunculopontine nucleus (PPN), and the parabrachial nucleus (PBN) to be the source of initial activity, which drives both the PPN and the forebrain structures.

14. NAZHESTKIN I.A.
The Moscow Institute of Physics and Technology (State University)
Integrated information theory application for brain networks plasticity research in instrumental learning

It is a complicated challenge to unambiguously set out what happens while learning process in a brain. In this study an attempt was made to use an integrated information theory (IIT) for this purpose. It has been shown that integrated information coefficient was significantly correlated with a habituation index in a spatial task with an aversive stimulus. Correlation was more strong in amygdala neurons compared to hippocampus neurons.

Applied neural systems

15. * ALSU SAGIROVA, MIKHAIL BURTSEV
The Moscow Institute of Physics and Technology (State University)
Extending Transformer Decoder with Working Memory for Sequence to Sequence Tasks

The paper introduces methods of memory augmentation in the Transformer decoder for sequence-to-sequence task. Transformer lacks storage for information associated with a context but not presented explicitly in the text. Learnable memory incorporated in Transformer provides additional space to keep necessary information. We show that augmenting the Transformer decoder with learnable working memory results in a performance boost for machine translation task.

16. * VLADIMIR I. GORBACHENKO, DMITRY A. STENKIN
Penza State University
Solving of inverse coefficient problems on networks of radial basis functions

The prospects of using radial basis function networks as physics-informed neural networks for solving direct and inverse boundary value problems described by partial differential equations are shown. To solve the coefficient inverse problem of recovering the properties of a piecewise inhomogeneous medium, an algorithm based on parametric optimization is proposed. The algorithm uses two radial basis function networks, one of which approximates the solution of direct problems, and the second approximates a function that describes the properties of the medium. Network training is performed using an algorithm developed by the authors based on the Levenberg-Marquardt method. Expres-sions for the analytical calculation of the elements of the Jacobi matrix in the Levenberg-Marquardt method are obtained. The application of the developed algorithm is demonstrated by the example of a model coefficient inverse prob-lem for a piecewise homogeneous medium. To solve the direct problem on ra-dial basis function networks for a piecewise homogeneous medium, an algo-rithm developed by the authors was used, based on solving individual prob-lems for each area with different properties of the medium and using a common error functional that takes into account errors on the border of areas.

POSTER SESSION 2

Wednesday, October 20                    14:00 – 14:45
Lecture-hall Актовый зал

Chair: Prof. LEBEDEV ALEXANDER EVGENYEVICH

SESSION 3

Wednesday, October 20                    15:00 – 16:30
Lecture-hall Актовый зал

Chair: Prof. VVEDENSKY VIKTOR

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

17. VICTOR VVEDENSKY
National Research Centre "Kurchatov Institute", Moscow
Origin of the Polysemy of Words

We observed regularities in spoken word recognition time for Russian adjectives when the meanings of the words to be recognized where close. Based on experimental results we managed to split all adjectives into groups forming distinct semantic categories. Words in each category were ordered into a list where meanings of neighbor words were close. The proximity of words was de-termined by the number of common translations into English. All these groups of words can be organized into a centrally symmetric array of radial word chains and we conjecture that this reflects cortical representation of the thesaurus of adjectives. The words are stored in individual memory cells and the size of the dedicated cortical area determines the number of words the language can use. We densely packed words into this area, which should be empty before language acquisition in early age. Presumably toddlers do not place first words into the memory so sharply as adults do and hence use multiple cells for one word. Later on new words replace initial ones in the most of semantic categories, though some “early” words remain in newly organized groups. Thus the word finds it-self in groups with different meaning and becomes polysemous

18. MARKOVA G.M., BARTSEV S.I.
Siberian Federal University, Krasnoyarsk
Coding of input stimuli by simple recurrent neural networks passing the delayed matching to sample test

Within the framework of the concept of neural correlates, reflection as a key property of consciousness was modeled on artificial neural networks. The method of cross-temporal classification was used to study the dynamic patterns of neural activity that are formed when the neural network passes the delayed matching to sample test. It was shown that identification of input stimuli using dynamic patterns of neural activity by simple means is impossible.

19. ISHKOV DENIS OLEGOVICH, TEREKHOV VALERY IGOREVICH
Bauman Moscow State Technical University
The Phenomenon of Resonance in Knowledge Distillation: Learning Students by Non-Strong Teachers

In real-world scenarios, we often want to obtain a small and at the same time accurate model. Smaller neural networks require less computational power and are faster than larger ones or ensembles. As a drawback, however, the representational knowledge of small models suffers. With the help of different techniques, it is possible to achieve an improvement in the predictive ability of such models while maintaining their complexity. Knowledge distillation, one of the popular techniques, introduces an additional term in the loss function. During the training of the student network, it minimizes Kullback-Leibler divergence between the smoothed output distributions of a teacher and student. We compared existing studies in the field of Knowledge Distillation. Specifically, we explore the influence of the teacher-student predictive ability gap on student improvement. Results show the existence of a clear dependency. Experiments on the image domain demonstrate the same tendency when increasing the accuracy of teacher's predictions. Revealed insights make it possible to choose teacher network's complexity to distil knowledge into the student network most efficiently in terms of student's accuracy-complexity tradeoff.

20. MIKRYUKOV ANDREY, VAZUROV MIKHAIL
Plekhanov Russian University of Economics
MODEL FOR FORECASTING UNIVERSITY ACTIVITIES BASED ON A GRAY FUZZY COGNITIVE MAP

An approach to solving the problem of predicting the performance indicators of a university as a weakly structured system based on cognitive modeling is proposed. The university activity model is a gray fuzzy cognitive map that allows taking into account the uncertainty of experts' opinions about the state of the cognitive map concepts, which provides more reliable modeling results. The novelty of the approach is determined by the use in the construction of a fuzzy cognitive map of special constructions in the form of interval estimates to assess the strength of connections between concepts that provide an adequate cognitive model. The proposed approach allows, under the given constraints, to find the most acceptable scenario for planning the increment of basic indicators to target values ​​by identifying the latent factors influencing them and impulse influences (increments) on them, ensuring guaranteed achievement of the set goal and obtaining more reliable results.

Neural network theory, neural paradigms and architectures

21. A.V. BAKHSHIEV, A.A. DEMCHEVA, L.A. STANKEVICH
1Peter the Great St. Petersburg Polytechnic University
2Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
CSNM: The compartmental spiking neuron model for developing neuromorphic information processing systems

The paper considers a compartmental spiking neuron model, which allows to describe a neuron as a tree structure consisting of separate segments, each of which can contain an arbitrary number of excitatory and inhibitory synapses. The main feature of the model is simplicity of implementing neurons with different configurations of dendritic and synaptic devices, while maintaining the ability to describe the dynamic processes of converting impulse flows in a neuron. The proposed structural description of a neuron makes it possible to independently compute both individual neurons and each equation de-scribing the synapse and the segment of the neuron membrane. This ap-proach should significantly facilitate the modeling of complex, dynamically changing neural structures. A number of experiments carried out have shown the possibility of structural adjustment of the model responses by op-timizing the model hyperparameters — neuron size, dendrite length, and the number of synapses.

22. ISAEV I.V., SARMANOVA O.E., BURIKOV S.A., DOLENKO T.A., LAPTINSKIY K.A., DOLENKO S.A.
1Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
2Lomonosov Moscow State University
An Inverse Problem Involving Integration of Optical Spectroscopic Methods: Study of Influence of Feature Selection on Resilience of Neural Network Solution to Noise in Data

A neural network solution to the inverse problem of optical spectroscopy for determining ion concentrations in a multicomponent aqueous solution of heavy metal salts and methods for increasing its resilience to noise in the data are considered. It is shown that simultaneous use of the selection of significant input features and the integration of physical methods with the correct method of application allows one to achieve the best result.

SESSION 4

Thursday, October 21                    11:00 – 13:00
Lecture-hall Онлайн

Chair: Prof. DMITRY YUDIN

Artificial intelligence

23. LEV ALEXANDROVICH STANKEVICH
Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
Cognitive technologies and artificial mind for humanoid robots

The work is devoted to application of cognitive technologies for creating artificial mind of humanoid robots. It is assumed that artificial mind is imitation of human mind. The imitation demands modeling psychical processes related to thinking and consciousness and, in particularly, modeling cognitive and creative functions. The functions are proposed to realize using network means of logical and associative processing information. As example cognitive dialogue system for robot is presented

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

24. ALEXEI V. SAMSONOVICH, ALEXANDR D. DODONOV, MATVEY D. KLYCHKOV, ANTON V. BUDANITSKY, IGOR A. GRISHIN, AND ALYONA S. ANISIMOVA
National Research Nuclear University (MEPhI), Moscow
A Virtual Clown Behavior Model Based on Emotional Biologically Inspired Cognitive Architecture

A cognitive model of the virtual clownery paradigm is developed, in which two clowns perform on a virtual stage, interacting with each other. Agents controlling virtual clown behavior were designed and implemented based on the eBICA cognitive architecture and embedded into the virtual environment, where they controlled two avatars. Clownery action was generated using the paradigm of ad hoc improvisation, without any given scenario. In this study, the output was produced as text, and the action was not visualized. In parallel, a number of scenarios were written by a group of experts using the same paradigm and behavioral repertoire. The two sets of scenarios were compared to each other in a number of characteristics. Results show which particular features of the model result in believable and socially attractive behavior. Implications concern socially-emotional intelligent agents for practically useful domains.

25. SHCHERBAN I. V., LAZURENKO D. M., SHAPOSHNIKOV D. G., KIRILENKO N. E.
1Scientific Research Center of Neurotechnologies, Southern Federal University, Rostov-on-Don
2
Development of Algorithms to Detect EEG Patterns Specific for Arbitrary Motor Activity of a Human in the BCI Applications

A new algorithm for patterns associated with the execution of voluntary motor activity reflected in electroencephalogram (EEG) signals has been developed. The Hausdorff metric was used as a search function. An adaptive low-pass filter was synthesized with the multivariate singular spectrum analysis. The use of the developed method has provided a reliable automatic search for induced premotor EEG patterns and the correct determination of their time boundaries (accuracy up to 96%).

26. DMITRY M. LAZURENKO, DMITRY G. SHAPOSHNIKOV, IGOR E. SHEPELEV, PAVEL D. SHAPOSHNIKOV, VALERY N. KIROY
1
2Scientific Research Center of Neurotechnologies, Southern Federal University, Rostov-on-Don
3Southern Federal University, Rostov-on-Don
Motor Imagery-related Quasi-Stationary EEG Patterns for Neural Interfaces

It is known that relatively stable patterns of brain activity are observed in the EEG, suitable for solving the problem of neural control and developing brain-computer interfaces (BCIs). In the present work, the variability of the Spectral Power and Cross-Correlation characteristics of the EEG was assessed to search for quasi-stationary patterns associated with voluntary motor imagery. It was shown that the most stable characteristics of the EEG recorded from the areas of the frontal, premotor and central cortex provide planning and control of the im-plementation of purposeful behavior. The instability of the characteristics of beta and gamma frequencies increased the variability and reduced the reproducibility of the target motor imagery EEG patterns recorded in these areas, as well as the efficiency of control in the closed-loop Brain-computer Interface.

27. SAEVSKIY A.I., SHEPELEV I.E., SHAPOSHNIKOV D.G., LAZURENKO D.M.
1Scientific Research Center of Neurotechnologies, Southern Federal University, Rostov-on-Don
2
Search for Informative Frequency Range and EEG Time Boundaries for Solving the Problem of Motor Imagery Patterns Classification

Аn approach based on machine learning methods to the classification of EEG recorded during 3 type of motor imagery is presented. Moreover, methods for determining the most informative frequency ranges and time characteristics of target movements have been developed based on frequency spectrum calculation and machine learning methods. It is shown that the selection of the optimal frequency ranges separately for each subject and each type of movement, as well as the selection of an informative time segment, significantly increase the classification accuracy. The comparison between 3 different features is given, namely, powers of frequency spectrum bands, Hjorth parameters and inter-channel correlations. The last two were shown to be advantageous when both search for informative frequency range and search for informative time segment are performed.

28. P.RUDYCH, K.LADONOVSKAYA
The Institute of Computational Technologies of SB RAS (ICT SB RAS), Novosibirsk
EEG correlates of perception of emotional adjectives analyzed by neural network’s interpretability values EEG correlates of perception of emotional adjectives analyzed by neural network’s interpretability values

The brain processes and its signals are highly affected by the variation of interpersonal, intrapersonal and situational data variability and it’s make the experiments and analysis complicated. We propose to use the tools of interpretability machine learning techniques to pick the experimental results from the noise environment, well described in metafactors. We tested it on the EEG ERP measurements of adjectives perception processes and studied it’s connection with the frequency of usage each word. We used the shapley values analysis to show the result correlation with the peak 350..400 ms latency and analyzed it’s nature.

29. VLADIMIR B. KOTOV AND ZAREMA B. SOKHOVA
Scientific Research Institute for System Analysis, Moscow
Possibility of Harmonious Coexistence of Human and Artificial Beings

We examine a way to achieve harmony in the community of people and intelligent artificial beings (robots). A simple model allows us to show that equilibrium is possible when robots are busy producing resources and people are busy manufacturing the robots. We determine conditions under which the ratio of people population to robot population tends to a constant with time, and simultaneously the populations tend to stationary values or increase exponentially. We examine the cases of a non-stop robot manufacturing and the case of stoppage in their production when consumption is low. We show that the first possibility is safer from the point of view of the civilization evolution and discuss the issues of organizing a correct production of robots.

SESSION 5

Thursday, October 21                    14:00 – 16:30
Lecture-hall Онлайн

Chair: Prof. MALSAGOV MAGOMED

Adaptive behavior and evolutionary modelling

30. NGUYEN NGOC DIEP, ZHDANOV A.A., SMIRNOV S.P.
1
2Lebedev Institute of Precision Mechanics and Computer Engineering, Russian Academy of Sciences, Moscow
Adaptive copter control

The article presents the results of model experiments on the use of a biologically inspired system of autonomous adaptive control (AAC) in copters. The general principles of building such control systems are considered and examples are shown concerning control at the levels of "mechanic", "pilot", "navigator" and "commander" both by individual copters and by their group. It is shown that the autonomous adaptive control system is able to automatically improve the quality of control as it learns itself.

31. SAMER EL-KHATIB, YURI SKOBTSOV , SERGEY RODZIN
1Saint Petersburg State University of Aerospace Instrumentation
2Southern Federal University, Rostov-on-Don
Comparison of Modified Hybrid object detected graph cut and Hybrid Ant Colony optimization - k-means for MRI images segmentation

. The Hybrid Ant Colony Optimization (ACO) – k-means and Modified Hybrid object detected Graph Cut image segmentation algorithms for MRI images seg-mentation are considered in this paper. The proposed algorithms and sub-system for the medical image segmentation have been implemented. The experimental re-sults show that the proposed algorithm has good accuracy in comparison to modfied Graph cut algorithm.

Neuromorphic computing and deep learning

32. MARIIA PUSHKAREVA, IAKOV KARANDASHEV
Scientific Research Institute for System Analysis, Moscow
Weight elimination trough structural pruning and targeted dropout

In this article we research the application of structured pruning to neural networks with batch normalization layers. The possibilities of pruning without further training are also being studied, this approach reduces computational costs. In this paper, we propose an algorithm for rebuilding the network, which reduces the memory required for network. The ResNet-32 network for the CIFAR-10 classification problem was compressed by 4.5 times.

Applied neural systems

33. VLADIMIR KRYZHANOVSKIY, GLEB BALITSKIY, NIKOLAY KOZYRSKIY, ALEKSANDR ZURUEV
Scientific Research Institute for System Analysis, Moscow
QPP: REAL-TIME QUANTIZATION PARAMETER PREDICTION FOR COMPRESSION AND ACCELERATION OF DEEP NEURAL NETWORKS

We suggested the approach of quantization parameter prediction of feature maps of hidden layers by NN input. The prediction is made before the inference of the input. This allows, from one hand, using efficient inference frameworks or CNN accelerators, which supports only static quantization parameters and getting the performance of dynamic quantization methods, on the other hand.

34. TARKOV MIKHAIL
Rzhanov Institute of Semiconductor Physics, Siberian Branch of Russian Academy of Sciences, Novosibirsk
SPICE-MODEL OF 2G FEFET TRANSISTOR AND ITS APPLICATION TO THE CONSTRUCTION OF THE ADAPTIVE ADDITER

A SPICE model of a two-gate ferroelectric transistor (2G FeFET) based on the use of the gate-drain I-V characteristic is proposed. An analytical approximation of the curves of the current-voltage characteristic is constructed. An adaptive adder model is built on the basis of the 2G FeFET model. It is shown that the adaptive adder is able to recognize distorted vectors corresponding to the reference vector of the adder weights (a property of associative memory).

35. MOSALOV O.P., IVANOV I.A.

The usage of an autoencoder as a filter in the task of key words retrieval

In this article the task of implementing a filter to assess the quality of key words retrieved from a scientific article is presented. An artificial neural network with the autoencoder architecture is used as such a filter. Different ways of internal structure of the artificial neural network are considered and compared. The description of computational simulations are provided, and their results are discussed.

36. SHUTYAK DENIS VYACHESLAVOVICH, LITVINOV OLEG STANISLAVOVICH
Bauman Moscow State Technical University
Study of the applicability of neural networks for frequency identification of objects.

We consider the problem of using a convolutional neural network for constructing the direction finding relief of an equidistant antenna array when superresolution of two signals is required. Resolution refers to the determination of the direction from which the signals come, provided that the angular distance between the sources is small. In the course of the research, a model of an equidistant antenna array was implemented, on the basis of which the "thermal noise" algorithm was implemented.

37. FEDYAEV O.I., RESHETNYAK Y.A.
Donetsk National Technical University
Neural network detection of objects of a given type on the path of a car

The problems of searching and classifying objects in images using modern computer vision algorithms are considered. To solve the problem of object detection, the original image is divided into fragments. Fragmentation made it possible to simplify the solution of the object classification problem based on the use of a convolutional neural network. The performance of the neural network object detection system is confirmed by the simulation results.

38. KNYAZEVA IS, PLOTNOKOV AA, MEDVEDEVA TV, MAKARENKO NG
1Sanct-Petersburg State University
2The Central Astronomical Observatory of the Russian Academy of Sciences at Pulkovo, Saint-Petersburg
Multi-output deep learning framework for solar atmospheric parameters inferring from Stokes profiles

Spectropolarimetric observations are broadly used for the extraction of physical information in the field of solar physics. Inferring magnetic and thermodynamic information from these observations includes inversion problem solving. Assuming that spectropolarimetric profiles are produced by a given atmospheric model, it is required to find the best sets of parameters within such a model corresponding to particular observations. Standard optimization approach often requires large computational resources and even in this case still performs very slowly. Previously it was suggested to use different strategies with artificial neural networks to overcome problems with computational power. It was previously shown that neural networks could be a viable alternative to the standard least square approach, but they could not replace it. Most papers only cover Magnetic Fields Vector parameter inferring, whereas the commonly used solar atmosphere model includes 11 parameters. In this paper we provide an end-to-end deep learning framework for full parameter inferring as well as comparison of several approaches for multi-output predictions. For this purpose, we trained one common network to predict all parameters, a set of parameter-oriented independent networks to deal with each parameter, and finally a combination of the above: a set of parameter-oriented independent networks built upon several layers of the pretrained common network. Our results show that using a partly independent network built upon a pretrained network provides the best results and demonstrates better generalization performance.

Neural network theory, neural paradigms and architectures

39. MIKHAIL KISELEV, ALEXANDER IVANITSKY, ANDREY LAVRENTYEV
1The Chuvash state university named after I. N. Ulyanov
2
Comparison of Memory Mechanisms Based on Adaptive Threshold Potential and Short-Term Synaptic Plasticity

Working memory is a crucial property of spiking neural networks (SNN) recognizing long complex spatio-temporal patterns since the features characterizing onset of the pattern should be stored until its end. Many different ways to implement memory mechanism in SNN have been proposed. Here, we compare two of them – one is based on the variable threshold potential while the other – on the short-term synaptic plasticity (STP). The methodology for objective quantitative memory quality measurement using a special kind of artificially generated signals is described. We demonstrate that the both mechanisms have comparable time characteristics while the STP-based memory shows significantly less capacity – at least, for the moderate size SNN (~10000 neurons).

SESSION 6

Thursday, October 21                    16:30 – 19:00
Lecture-hall Онлайн

Chair: Prof. TEREKHOV SERGE

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

40. BURLAKOV E.O., VERKHLYUTOV V.M., USHAKOV V.L.
1V. A. Trapeznikov Institute of Control Sciences of Russian Academy of Sciences, Moscow
2Institute of Higher Nervous Activity and Neurophysiology of RAS, Moscow
3National Research Centre "Kurchatov Institute", Moscow
SIMPLE HUMAN BRAIN MODEL REPRODUCING EVOKED MEG BASED ON NEURAL FIELD THEORY

Macro-scale models of the human brain are based on the structure of the connectome in most cases. However, they do not take into account the features of neuronal activity propagation observed experimentally. Our model develops the idea of the dynamics of local neural network activity, providing an opportunity to simulate traveling waves of excited neuronal populations. In this work, we describe a spatial-structural model involving an imitation of mesoscale electrical activity of the brain calculated based on neural field equations with no account for the structure of the connectome.

41. MAIOROV VLADIMIR I.
Lomonosov Moscow State University
The Tao of Artificial Intelligence

To train artificial intelligence from scratch, you need a consistent list of examples, marked up according to what is good and what is bad. As an initial filter when compiling the list, one can take maxims from the "Metaphysics of Morals" by I. Kant: "one's own perfection as a goal, which is at the same time a duty"; "the happiness of another as a goal, which is at the same time a duty."

42. MEILIKHOV E.Z., FARZETDINOVA R.M.
National Research Centre "Kurchatov Institute", Moscow
Value-based decision making, ˗simple analytic theory

In the framework of modern ideas, decision making and, hence, the choice of the optimal behavior, take place on the level of specific neuron networks. Those networks accumulate information on the possible behavioral alternatives, estimate all pro et contra and make the optimal decision. In formal network models, the standard one, describing the dynamics of that process, is the attractor dynamics, where the switching from one decision to another one is the transfer of the system from one attractor to another. That attractor model is formal, is not practically associated with the neurophysiology and is based on the numerical solving some ad hoc equations. More prominent place in the hierarchy of models is hold by analytic models, where system properties are investigated by analytic methods of the mathematical analysis. In the present work, we just consider one of simple analytic models. In that model, the dynamical process of decision making is likened (as in the attractor model) to moving the system over the energy landscape in the presence of strong enough noises. This process could be readily presented by the ball movement over the rough hummocky surface (in such a simulator, more or less intensive surface shaking could play the role of the noise). Different actual decisions correspond to landscape wells of different depths being separated by energy barriers from each other. In the process of decision making, the ball overcomes those barriers, due to noises, and moves from one well to another one. If the probability of the ball transfer to one of the wells is superior to probabilities of other transfers, that process results in the unique decision; conversely, different decisions could be made with comparable probabilities. As against to attractor models with the explicit consideration of noise effects, we use the known Arrhenius-Kramers formula which associates the mean system life-time in the certain quasi-stationary state with the height of the energy

Neurobiology and neurobionics

43. VASILY S. TISELKO, ANTON V. CHIZHOV
Ioffe Physical Technical Institute, Russian Academy of Sciences, St Petersburg
Contribution of multilayer interactions to neural activity retaining in response to flash stimulus in simple and complex models of an orientational hypercolumn of visual cortex

Cortical neural networks in vivo are able to retain their activity evoked by a short stimulus. Our previous study has shown the major role of strong excitatory recurrent connections and slow NMDA receptor kinetics into the retention effect in case of single layer of recurrently-connected neuronal populations. In the present work, we investigate two multilayer models to study the role of interactions between the layers into the effect. Simulations with the simple firing rate-based multilayered ring-structured model of an orientational hypercolumn of the visual cortex, taking into account recurrent connections, synaptic depression and slow NMDA receptor currents, show that strong feed-forward connections between the layers could compensate for weak lateral recurrent connections and support the retention in the subsequent layers. In case of developed recurrent both lateral and feedforward connections, a similar response of populations in all layers is observed. Complex biologically detailed 2-d conductance-based refractory density model confirms the results. The retention effect can play an important role in such cognition processes as continuous motion perception.

44. MARIA R. KOTIKOVA, ANTON V. CHIZHOV
Ioffe Physical Technical Institute, Russian Academy of Sciences, St Petersburg
Propagation and relaxation of neuronal membrane mechanical deformations in mathematical model

Neuronal membrane undergoes mechanical deformations in response to electrical pulse or mechanical influence. We propose a one-dimensional mathematical minimal model that describes propagation and dissipation of the deformations. The model is based on fluid dynamics equations and the Lippmann equation connecting electrical effects and membrane elasticity. The effects of internal heterogeneity of neuron, e.g. cytoskeleton, is investigated and specified as relaxation.

45. MARGARITA M. PREOBRAZHENSKAIA
P.G. Demidov Yaroslavl State University
Three unidirectionally synaptically coupled bursting neurons

We investigate the bursting-behaviour of three unidirectionally coupled neurons with electrical coupling. The mathematical model of this association is the system of relay differential-difference equations. The unidirectional synaptic coupling was modeled based on the idea of fast threshold modulation. We show that the system has periodic mode with bursting effect with different bursts for every component.

46. MYSIN I.E., CHIZHOV A.V.
1Institute of Theoretical and Experimental Biophysics of RAS, Pushchino
2Ioffe Physical Technical Institute, Russian Academy of Sciences, St Petersburg
Image encoding by orientation hyperсolumn

Activity maps V1 raise a question of encoding images using a small number of orientation hyperсolumn (HC). What features and how many of them should be coded by the HC to ensure visual acuity as a human? We have shown that encoding in each HC only the average brightness and its gradient gives visual acuity 10 times worse than that of a human. We took into account that the HC has detectors in the form of first and second derivatives of Gaussians with different scales, and we have obtained high-quality encoding of 1-d signals.

47. MAIOROV VLADIMIR I.
Lomonosov Moscow State University
LEARNING IN BIOLOGICAL AND ARTIFICIAL NEURAL NETWORKS

Learning algorithms in biological and artificial neural networks are very different. In biological recurrent networks, neural attractors are formed. Artificial recurrent networks are tuned to process vector sequences, but do not provide the formation of attractors. Artificial networks consist of “point” neurons with high-dimensional synapses. Biological neurons are connected by synapses with a small number of states, but with the possibility of local integration in numerous dendritic branches.

48. I.V. MAKUSHEVICH, N.G. BIBIKOV, I.N. PIGAREV
1N.N. Andreyev Acoustics Institute, Moscow
2Institute for Information Transmission Problems (Kharkevich Institute), Moscow
Comparison of the dynamics of the background activity in the located nearby neurons of the cat's auditory cortex

The characteristics of the activity of six simultaneously recorded neurons of the auditory cortex of an unanesthetized cat in the absence of controlled sound stimuli are analyzed in detail. All cells characterized by a short refractory period and a site of increased excitability after the generation of the spike. Slow changes in activity were monitored by successive estimates of the Hurst index. It was shown that the trends in the firing rate of the background activity of neighboring cells are positively interrelated, forming jointly functioning clusters.

POSTER SESSION 3

Friday, October 22                    14:00 – 16:00
Lecture-hall Онлайн

Chair: Prof. KAGANOV YURI

Adaptive behavior and evolutionary modelling

49. RED'KO VLADIMIR GEORGIEVICH
Scientific Research Institute for System Analysis, Moscow
Modeling of Interaction between Learning and Evolution for Kauffman’s NK Networks

The model of interaction between learning and evolution for Kauffman’s NK networks has been designed and analyzed by means of computer simulation. The evolving population of autonomous agents is considered. Any agent has the genotype and the phenotype. Genotypes and phenotypes are coded by Kauffman’s NK networks. The interaction between learning and evolution is studied. The effect of hiding is demonstrated: intensive learning can suppress the evolutionary optimization of genotypes.

50. TALALAEV DMITRY V.
Lomonosov Moscow State University
Mutations in Hopfield neural network

In this note, we consider the mutations of the graph defining the Hopfield model that preserves the asymptotic behavior of the model. This question can be of significant importance in optimizing the network topology. From a philosophical point of view, this question is close to the study of the post-priori equivalence of the functioning of networks with different knowledge.

Artificial intelligence

51. VLADISLAV DOROFEEV
Scientific Research Institute for System Analysis, Moscow
Object-Process Methodology for Intelligent System Development

Development of the new artificial systems with unique characteristics is very challenging task. In this paper the application of the hybrid super intelligence concept with object-process methodology to develop unique high-performance computational systems is considered. The methodological approach how to design new intelligent components for existing high-performance computing development systems is proposed on the example of system requirements for "MicroAI" and "Artificial Electronic" systems.

52. KURYAN V.E.

AUTOMATIC TRANSLATION BASED ON THE WORLD MODEL GRAPH

Building a world model graph is made automatically based on the processing of an array of pairs of texts in the input and output languages. The approach is based on the representation of the situation as a subtree of the general model of the world in the input language. The corresponding subtree in the output language is converted into plain text, which is the translation of the source text

53. CHEREMISINOVA OLGA NIKOLAEVNA, ROSTOVTSEV VLADIMIR SERGEEVICH
Vyatka State University
Improving the quality of image recognition by automatic selection of hyperparameters of a convolutional neural network

The analysis of common methods of automatic selection of hyperparameters of a convolutional neural network (CNN) is carried out. A hybrid hyperparameter optimization method is proposed that combines the advantages of Bayesian and evolutionary methods. The experimental results showed that the quality of the SNS tuning for the considered methods is the same.

54. OLEG O. VARLAMOV, LARISA E. ADAMOVA, YURIY E. GAPANYUK, VALERIY I. TEREKHOV, DMITRIY V. ALADIN, DMITRIY A. CHUVIKOV
Bauman Moscow State Technical University
The Multidimensional Open Gnoseological Active Network (MOGAN) Approach for Creating an Artificial General Intelligence

To create artificial general intelligence, we need new tools that can work in different areas and directions. Long-term research of mivar technologies of logical artificial intelligence has allowed creating a new powerful, universal, and fast tool called “the multidimensional open gnoseological active network” (MOGAN). The generalization of evolutionary databases and knowledge with linear logic and computational processing has been completed. The restrictions on the exhaustive search for the construction of cause-and-effect reasoning algorithms have been removed. This tool allows one to quickly and easily work with logical reasoning in the “If-then” format and can be used to create various new-level AI application systems. It is compatible with existing developments of reflex and logical AI levels. The development of the MOGAN approach involves models based on complex networks. Thus, the MOGAN approach can be considered as one of the possible ways to create an AGI.

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

55. DARIA V. TIKHOMIROVA
National Research Nuclear University (MEPhI), Moscow
The Contact Cycle for modeling the behavior of the Virtual Bot-Presenter

We are faced with the question of when, why, and what emotions a presenter should express during a presentation in order to get interest of the audience. To answer the question, you need to understand the motive of a presenter. We can assume that his motive is to interest, engage, impress the listener so that he shares his feelings about the report, so that thoughts about the content of the report keep the listener engaged, so that he finds the reflection of the report ideas in his own experience. In this paper we propose to use the contact cycle and structures of emotions for modeling the behavior of the virtual bot-presenter.

56. SELEZNEVA EKATERINA IGOREVNA, TIHOMIROVA EKATERINA ALEXANDROVNA
N.I. Lobachevsky State University of Nizhni Novgorod
Pupil dynamics and heart rate during the reporting of deliberately false information under induced cognitive load

e consider theoretical approaches and methods related to the relevant paradigm of lie detection - selective influence of cognitive load on liars. The advantages of using eye-tracking and heart rhythm variability dynamics analysis (event-related heart rhythm telemetry) technologies are proved. We conduct an experiment to differentiate the influence of emotional arousal and cognitive load on information markers of lying, as well as to test the technology of event-related telemetry of heart rhythm in the area of lie detection.

57. VORONKOV G. S.
Lomonosov Moscow State University
THE MIRROR IMAGE FACTOR IN VISION AND THINKING: MORPHOLOGICAL AND FUNCTIONAL MANIFESTATIONS

The paper considers the manifestations of the mirror image factor (MIF) in vision and thinking. The MIF itself is discussed in a debatable way - it is suggested that the presence of the mechanism of mirror transformation (correspondence) in vision is due to the need for the coordinate system of the subjective visual image to be identical to the coordinate system of the real visual space. Taking into account MIF, an explanation is given for some "strange" phenomena (both morphological and functional) in vision and thinking.

58. PROSHINA E.A., SAPRIGYN A.E., LEBEDKIN D.A.

EEG patterns of cognitive processing of adjectives of different frequency of occurrence and emotional coloring

The aim of this work is to analyze the EEGs of healthy subjects using the methods of source localization in order to study the temporal dynamics and localization of functional activity in the cognitive processing of verbal stimuli.

Neurobiology and neurobionics

59. DICK O.E.
Pavlov Institute of Physiology of the Russian Academy of Sciences, St.Petersburg
Wavelet analysis of correlations between EEG patterns and heart rate variability in subjects with discirculatory encephalopathy

The task is to find connections between the brain and heart bioelectrical activities in subjects with discirculatory encephalopathy. To solve the problem, the method of synchrosqueezed wavelet transform of simultaneously recorded electroencephalogram (EEG) and electrocardiogram (ECG) is used with subsequent analysis of the ratio of instantaneous frequencies. It was shown that the time of occurrence of the correlation of instantaneous frequencies of the EEG delta range and heart rate variability correlates with the severity of discirculatory encephalopathy.

60. S.V. BOZHOKIN
Peter the Great St. Petersburg Polytechnic University
Wavelet analysis of the non-stationary Rose-Hindmarsh model describing neural activity

A new approach is proposed to developing a model of non-stationary sig-nals, whose spectral properties change in time. The approach is based on improving the Rose-Hindmarsh (RH) system of equations describing the electrical activity of individual neurons. The RH model analyzed the de-pendence of the number of spikes in each burst of neuronal activity on the control parameter. The new approach to the RH model is that instead of a constant value of the control parameter in the RH model, a time-dependent function is considered. The transitional period of neuronal activity, in which the rhythm of bursts and the number of spikes in each burst change, is stud-ied using the continuous wavelet transform. For two neurons with different electrical activity, continuous wavelet transformations are constructed, de-pending on the corresponding time and frequency. The concept of spectral integrals is introduced, which are the average value of the local density of the signal energy spectrum, integrated over a certain frequency interval. Spectral integrals are calculated for two different neurons in the range of low, medium and high frequencies. Spectral integrals give a visual represen-tation of the change in the amplitude-frequency properties of the neuron in time and show the changes in the neuronal activity over time in a certain frequency range.

61. OLEG S. VASILYEV, OLGA G. SAFONICHEVA, EVGENY E. ACHKASOV,
1
2Sechenov University
"Smoothness" as a measure of diagnostics and effectiveness of musculoskeletal system rehabilitation in the of athletes.

The movement of a healthy person is characterized by smooth trajectories, while the damaged link of the musculoskeletal system produces intermittent movements and it is not able to produce completed movement on the range of acts, available to it. Studies of the “smoothness” parameter as a diagnostic measure and a criterion for the effectiveness of the musculoskeletal system rehabilitation in sports medicine have not been carried out yet. The research studied the average smoothness of the trajectories of markers, fixed on the lower limb during performing the Grand-plié test choreographic movement. Correlation between the level of pain syndrome and the normalized measure of smoothness for the hip joint is 0.85; for the ankle it  is 0.74. Correlation between the dynamics of recovery and the normalized measure of smoothness for the hip joint = 0.96; for the ankle = 0.62. Thus, smoothness is considered to be a diagnostic indicator reflecting the adequacy of the rehabilitation load and the entire rehabilitation process.

62. PROSKURA A.L., VECHKAPOVA S.O., ZAPARA T.A., RATUSHNYAK A.S.
1The Institute of Computational Technologies of SB RAS (ICT SB RAS), Novosibirsk
2
EFFECTS OF LEPTIN AND INSULIN ON HIPPOCAMPAL SYNAPTIC PLASTICITY

The paper presents the results of the analysis of regulatory protein-protein interactions in the interactome of the dendritic spine of the pyramidal neuron of the hippocampal CA1 field under the control of hormones produced by the peripheral tissues of the body. A molecular mechanism has been proposed for the realization of their effects on excitatory neurotransmission in the hippocampus.

Neuromorphic computing and deep learning

63. KOTOV VLADIMIR BORISOVICH, PUSHKAREVA MARIIA MIKHAILOVNA
Scientific Research Institute for System Analysis, Moscow
The Bidirectional Variable Resistor Model

The successful realization of capabilities of variable resistors as elements of neuromorphic systems calls for a simple model that can reproduce the basic properties of such elements. The paper considers a model of a variable resistor using two simple parallel opposed resistance elements. The equations reducing the variation of a resis-tor’s state to the movement of a representative point in the two-dimensional space are written and investigated. The model is shown to reproduce many properties of real variable resistors and allow us to evaluate their usability in particular data processing systems. Some potentially useful modes of operation of variable resistors are analyzed. The conditions under which a variable resistor can be used as a memory unit, sign-variable signal detector are found. The important role of characteristic functions for estimating variable resistors is pointed out.

64. SHAMIN A. Y., KARANDASHEV Y. M.
Scientific Research Institute for System Analysis, Moscow
ON THE NEURAL NETWORK APPROACH TO THE SOLUTION OF DIFFERENTIAL EQUATIONS

A generalization of the neural network method for the numerical solution of differential equations with given initial and boundary conditions is proposed. At the same time, it is proposed to train the network along with the initial conditions. Thus, the network learns to solve not only one problem with fixed initial conditions, but a whole class with conditions in a certain range. Examples of the method for a number of problems are given: harmonic oscillator, linear with exponentially growing solutions, wave equation).

65. ANTON KORSAKOV, ALEKSANDR BAKHSHIEV, LYUBOV ASTAPOVA AND LEV STANKEVICH
1Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
2Peter the Great St. Petersburg Polytechnic University
Application of the compartmental spiking neuron model for the conditioned reflex implementation

The question of modeling and implementation of the conditioned reflex is considered. The reflex model comprises a neural network based on a spiking compartmental model of a neuron with the possibility of structural ad-aptation of the dendritic tree to the input spikes pattern. The neuron model allows to describe and change its structure online (the size of the cell body, the number and length of dendrites, the number of synapses), depending on the incoming spikes pattern at its inputs. A brief description of the compartmental spike model of a neuron is given. The method of structural adaptation of the model to the input spikes pattern is described. To study the work of the proposed neuron model in a network, the choice of a conditioned reflex as a special case of the mechanism of forming associative bonds is justified as an example. The structural scheme of the conditioned reflex formation is described. A description of the experiment on the reflex formation is presented. The conclusion is made about the prospects of using spiking compartmental models of neurons to increase the bio-likelihood of behavioral functions implementation in neuromorphic control systems.

Applied neural systems

66. VLADIMIR B. KOTOV AND GALINA A. BESKHLEBNOVA
Scientific Research Institute for System Analysis, Moscow
Generation of the Conductivity Matrix

Vector-matrix multipliers are key units of neuromorphic systems. Resistor arrays can be used to realize this sort of multipliers. Voltages connected to the resistor array input define the vector in the product, while resistor conductivities determine the matrix multiplier. If current-controlled variable resistors are the elements of the array, it is possible to use the same resistor matrix in many vector-matrix multiplications. The key problem is to set correct conductivities to variable resistors. The paper considers a resistor array whose variable resistors are generated by coupling conductors from a particular set. All the resistors are considered directional; their directions are determined by a signature matrix. It is shown that the generation of the correct conductivity matrix is not always possible. In fact, only very simple conductivity matrices can be readily generated and the operation may require a special signature matrix. However, a signature matrix is defined by a particular technologic process, which imposes constraints on the form of the signature matrix. The optimization approach, which ignores the specific form of a conductivity matrix, looks more promising. On the other hand, it is possible to vary this matrix in different directions, which allows us to choose the best conductivity matrix in terms of the whole neuromorphic system. We can get a high-dimensional manifold of conductivity matrices by applying different voltages to the resistor array input and obtaining the stationary state of the conductivity matrix

67. ENGEL EKATERINA ALEXANDROVNA, ENGEL NIKITA EVGENYEVICH
Katanov Khakass State University, Abakan
The Fault Forecasting System in a Solar Plant on the Basis of a Modified Fuzzy Neural Net

This paper presents a fault forecasting system in a solar plant on the basis of a modified fuzzy neural net which maximizes the reliability of a solar plant and fi-nancial outcome of investment. In order to automatically generate an optimum ar-chitecture of the modified fuzzy neural net we modified a quantum-behaved PSO which fulfills a multi-dimensional quantum-behaved search of swarm particles for both positional and dimensional optimum. We developed a fault forecasting system in a solar plant on the basis of a modified fuzzy neural net which optimum architecture automatically generated by the modified quantum-behaved PSO based on Supervisory Control and Data Acquisition data of basic sensors. The results from 120 runs show that the proposed modified quantum-behaved PSO prevents an early convergence to local minimum, exhibits the better performance, fast global convergence and calculation speed than does the case of the modified PSO. The developed fault forecasting system in the solar plant on the basis of the created modified fuzzy neural net effectively forecasts a failure free operating pe-riod of the solar plant.

68. IVAN FOMIN, ANDREY ARKHIPOV
Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
Research on neural networks for landmark detection in robotic vehicle navigation task

The article discuss the problem of visual landmarks detection on video images received by a fish-eye camera using neural networks. The problem is important in the framework of navigation of the autonomous vehicle according to the image sequence from the camera using a minimum amount of sensory information. A fish-eye camera lens allow system to view and detect objects in a much wider range of angles, increases the time of tracking an object and using it to clarify the position. To choose the most relevant solution, neural network methods for object detection of different architectures and creation time were considered. To perform an experimental research, a dataset was formed for landmarks detection, containing images of various sizes from cameras with different sensors and types of wide-angle lenses. The collected data set was used to train and test selected neural network architectures. Based on the analysis of the results, a conclusion was made about which of the presented architectures can be used in the navigation task when object detection process performed on embedded and desktop PC.

69. LITVINOV O., MURODYANS D. AND KARANDASHEV I.
1Bauman Moscow State Technical University
2Scientific Research Institute for System Analysis, Moscow
Research of useful signal losses in adaptive antenna arrays with neural network control of phase shifters in various modes of their operation

In this work, we analyzed the characteristics of noise suppression in an adaptive antenna array under various control modes - traditional, phase, discrete phase, and with amplitude and phase coupling at discrete phase. The studied characteristics were the signal/noise and the power loss of the useful signal at the output of the adaptive antenna. For the last three modes, a specially trained neural network was used. It has been shown that the use of phase control modes shows better results in maintaining the power of the useful signal than traditional amplitude-phase control.

70. NIKOLAY FILATOV, TIM ISAKOV, ALEKSANDR BAKHSHIEV
1Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
2Peter the Great St. Petersburg Polytechnic University
Research on the applicability of monocular 3d object detection using CARLA simulator

Monocular 3d object detection methods are promising in the field of making au-tonomous robots without lidar, which can reduce costs of production significant-ly. However monocular 3d object detection methods tend to have low precision due to inaccurate inference of distances to objects. Nevertheless, there are several ways to measure the impact of detection precision on the downstream autono-mous driving task. In this work, autonomous agents which use lidar, monocular camera, and ground truth for 3d object detection are compared in the CARLA simulator. Each agent has passed a set of routes with challenging traffic situa-tions, totaling 122.5 km driven. Quality of movement was assessed using the collisions statistics, as a result, the agent using a monocular camera performed 4.5 % better than the agent using lidar. This indicates the applicability of monocular 3d object detection algorithms in certain cases.

71. KERBENEVA A.Y., MILOV V.R.
Nizhny Novgorod State Technical University named after R.E. Alekseev
Modeling of procedures for semantic analysis of students' short answers to open questions

The paper proposes a methodology for the intellectual assessment of knowledge based on users' answers to open questions. The procedures for preprocessing, text vectorization and classification of vector representation are considered. The results of training and testing of the proposed algorithms are presented.

72. BELYANOVA M.A., ANDREEV A.M., GAPANYUK U.E.
Bauman Moscow State Technical University
Neural Text Question Generation for Russian Language using Hybrid Intelligent Information Systems Approach

The paper considers the task of natural language text generation, particularly the task of question generation based on the given text. Recent research in the area of text generation shows that better quality is achieved using the autoregressive generative models, trained on language modeling tasks. However, such models do not use the meta-information of the text structure. In the paper, this approach is applied to Russian language texts. The architecture of the intelligent question generation system is described in detail.

73. MARYASIN O.YU., LUKASHOV A.I.
Yaroslavl State Technical University (YSTU)
FORECASTING PEAK LOAD HOURS OF THE REGION OF RUSSIA

The paper considers the problem of forecasting the peak load hours for an upcoming month using artificial neural networks. Two different methods are used to forecast peak load hours. The first method is an indirect method based on forecasting the total energy consumption of the region for the next month. The second method is based on the direct use of information on peak load hours in previous months.

74. PAVEL A. KOLGANOV, YURY V. TIUMENTSEV
Moscow Aviation Institute (National Research University)
Semantic Image Segmentation as Tool for Situational Awarkness in Unmanned Vehicle Control Tasks

This paper considers situational awareness formation, which is essential for controlling the behavior of unmanned aerial vehicles (UAVs). The convolutional neural networks, which perform semantic segmentation of the image obtained by UAV video surveillance tools. Finally, as a demonstration example, the task of selecting a site in unfamiliar terrain for landing the UAV, satisfying the conditions of safety of performance of this operation is considered.

Neural network theory, neural paradigms and architectures

75. LEBEDEV ALEXANDER
Scientific Research Institute for System Analysis, Moscow
Learning Independent Sparse Representation In Application to Symmetry Detection Problem

We continue to research biologically plausible learning methods and apply them to symmetry detection problem. We used the Hebbian-like learning rule, proposed in our previous work and upgraded it with a new algorithm of interaction of neurons of hidden layers. This algorithm relays only on pair-wise interaction between neurons. Considering activation statistics, neurons suppress each other in case of statistically dependence. This allow them to learn statistically independent patterns. Together such neurons form what we call “sparse independent representation”. We test the proposed learning method on symmetry detection task. This task is characterized by practically infinite number of training samples, so meth-ods should demonstrate a generalization ability to solve it.

76. MIKHAIL KRASNOV, YURI RYKOV, VLADIMIR SMOLIN
1Keldysh Institute of Applied Mathematics, Moscow
2
LINEAR PROPERTIES OF BPE AND DIFFERENTIAL EQUATIONS SOLUTIONS APPROXIMATION

The consideration of BPE (Back Propagation error) is proposed to divide into identifying the dependencies of the hidden elements activities and matching the output activity with the output transformation function Y ⃗(X ⃗). Often this does nothing but improves your understanding of BPE. But for non-standard tasks, it allows you to reconcile the description of neural networks with the formulas for calculating the error function. As an example, the approximation of the Hopf differential equation solutions is given.



Российская академия наук Министерство образования и науки Российской Федерации МФТИ НИЯУ МИФИ НИИСИ РАН МАИ ТРИНИТИ