Sections
Neuroinformatics - 2022
SESSION 1
Monday, October 17 17:30 – 19:45
Lecture-hall Концертный зал
Chair: Prof. SOKHOVA ZAREMA BORISOVNA
Neural network theory, neural paradigms and architectures
1. MIKHAIL KISELEV, ANDREY LAVRENTYEV1The Chuvash state university named after I. N. Ulyanov
2Kaspersky Lab, Moscow
“Gas” instead of “Liquid”: which Liquid State Machine is Better?
The crucial feature of liquid state machines (LSM) is memory which is nec-essary to the convert spatio-temporal input patterns and their sequences scat-tered in time into multidimensional representation in the form of instant neuronal activity. In this work, we utilize a methodology for exact estima-tion of LSM memory ability and to find parameters of the LSM “liquid” (the chaotic spiking neural network) optimal from viewpoint of memory depth using genetic algorithm. We applied this technique to chaotic networks of leaky integrate-and-fire neurons with adaptive threshold. The result of the optimization was rather unexpected – best memory was demonstrated by an ensemble of unconnected neurons more corresponding to metaphor of “gas” than “liquid”. This result compels to revise the traditional view on liquid state machines and efficiency of spiking neural networks playing the role of “liquid” in it.
2. TATIANA LAZOVSKAYA, DMITRY TARKHOV, DARYA CHERNUKHA, ALEXANDER KORCHAGIN, ALEXANDER VASILYEV, AND GALINA MALYKHINA
Peter the Great St. Petersburg Polytechnic University
Analysis of predictive capabilities of adaptive multilayer models with physics-based architecture for Duffing oscillator
The development and implementation of cyber-physical systems requires the use of fast algorithms for creating adaptive dynamic models of real objects. In this paper, the adaptation of two multilayer models with different basis methods are presented and tested. Multilayer models refer to models with physics-based architecture. The quality of constructing a medium-term forecast based on the processing of dynamic measurements including random errors obtained during the interaction of the object and software in real time is compared. As an example, constructing an approximate functional solution of the Dung equation with a dynamical parameter is considered. The results of comparing models with approximately the same net construction time and assessing the presence of significant differences in the predictive capabilities of the methods used are presented. The optimal models for practical applications depending on the specifics of the specific problem being solved are discussed.
3. A.YU. DOROGOV
Saint Petersburg Electrotechnical University "LETI"
Application of fast neural networks for correlation measurements
The paper presents methods for constructing topology and parametric training of fast neural networks (FNN) for problems of correlation measurements of one-dimensional signals and images. A mathematical model of the FNN is described. The connection of fast algorithms with self-similar structures is noted. A method of multiplicative representation of arbitrary discrete functions and images is shown. Algorithms for constructing the topology of the FNN and training for the implementation of the correlation discriminator are proposed.
4. MARK NAGOVITSIN, DENIS KUZNETSOV
1The Moscow Institute of Physics and Technology (State University)
2DeepPavlov
DGAC: Dialog Graph AutoConstruction up on data with a regular structure
A script graph is often used in development when the dialog has a regular structure. In this paper, we propose a method for recovering or extracting regular structures from data by building a dialog graph. Based on the MultiWOZ dataset, the dialog graphs can be used for a more accurate pre-selection of candidates for various response selection models: statistical, pre-trained and fine-tuned. Obtained results show, that our approach can be applied to automatic graph construction of a scenario when developing a dialogue system.
5. O.P. MOSALOV
Generative adversarial networks as an approach to unsupervised link prediction problem
An unsupervised approach for the problem of link prediction in a graph representing a subject area is described. The approach is based on the usage of generative adversarial networks architecture to generate vectors that can be transformed into a set of edges. The approach is validated on a graph built with the data extracted from a file archive. Main results achieved are discussed, together with advantages and disadvantages of the approach and potential prospects of the algorithm.
6. * SERGEY CHERVONTSEV, ALEXANDER TROFIMOV
National Research Nuclear University (MEPhI), Moscow
On the similarities between denoising diffusion models and autoencoders
Denoising diffusion probabilistic models (DDPMs) have recently emerged as a powerful paradigm for generative modelling, outperforming adversarial methods in various domains and applications using a comparable amount of computation resources. Recent findings demonstrate that there is a connection between them and more venerable methods like denoising autoencoders. In this article, we present a simplified description of this connection that maintains all essential components. We also demonstrate empirically that, at least on toy datasets, one can indeed obtain a similar performance to DDPM by training a timestep-aware denoising autoencoder using an instance noise trick, which doesn't require any reparameterizations.
7. KHARCHENKO B.V., PROTASOV B.I.
Moscow Aviation Institute (National Research University)
Creation of neural network committees for image recognition with small values of erroneous decisions
The paper considers the task of creating and developing a committee based on convolutional neural networks, working on the principle of evolutionary solution matching, which allows increasing the probability of correct image recognition of a high degree of difficulty compared to a conventional neural network and significantly reducing the probability of making erroneous decisions in difficult cases.
8. MARGARITA PREOBRAZHENSKAIA
P.G. Demidov Yaroslavl State University
Relay System of Differential Equations with Delay as a Perceptron Model
We study the perceptron model, which is a system of differential-difference equations. The model has three types of neurons: sensory, associative, reactive. It is analytically shown that with appropriate parameters the membrane potentials of associative neurons are either periodic functions with one burst per period, or exponentially tends to zero. For a reacting neuron, the existence of a mode with any finite number of bursts, after which the neuron goes into a state of refractoriness.
9. KRYZHANOVSKY B.V., LITINSKII L.B., KAGANOWA I.M.
Scientific Research Institute for System Analysis, Moscow
THERMODYNAMICS OF THE INTERACTION OF TWO HOMOGENEOUS GROUPS OF BINARY AGENTS
We previously examined a set of fixed points of a neural network with a block-constant connection matrix, which has two blocks along its main diagonal. Such a matrix describes interactions between two homogeneous groups of binary agents (spins, neurons, etc.). In the present paper we analyze thermodynamic characteristics of this system and their dependence on the external parameters of the problem.
SESSION 2
Tuesday, October 18 12:00 – 13:00
Lecture-hall Зал ученого совета
Chair: Prof. LEBEDEV ALEXANDER EVGENYEVICH
Applied neural systems
10. BELKIN I.V., REZANOV A. D., YUDIN D. A.The Moscow Institute of Physics and Technology (State University)
Center3dAugNet: effect of Rotation Representation on One-Stage Joint Car Detection and 6D-Pose Estimation
Joint car detection and 6D-pose estimation on monocular images is a topic of active research. This paper discusses the fast one-stage method based on CenterNet deep architecture, named Center3dAugNet. Particular attention is paid to the study of rotation representation effect on network training results. It is of high importance due to discontinuity of 3D-object yaw, pitch and roll angles. Relative quaternion representation has shown the greatest positive effect on the mean average precision of network. Also we introduce new loss function for rotation estimation in terms of quaternions. The performance of the proposed approach is evaluated both on powerful hardware platforms based on NVidia Tesla V-100, RTX2080Ti, RTX2070, as well as on embedded computing boards Jetson Xavier and TX2. Our experiments indicate that the presented approach is suitable for real-time computer vision systems of unmanned vehicles and mobile ground robots. Source code is made publicly available: https://github.com/cds-mipt/center_3d_aug_net
11. * BASHAROV ILYA, YUDIN DMITRY
The Moscow Institute of Physics and Technology (State University)
Multitask learning for extensive object description to improve scene understanding on monocular video
Simultaneous localization, classification and association - a pool of the most important tasks in computer visionюIn this paper, we extend the nuScenes dataset with segmentation masks by using the heavy HTC neural network, and also present an algorithm for merging masks with current ground-truth observations. Also, we are introducing a model that allows us to solve 4 tasks simultaneously using monocular video only: instance segmentation, 2D and 3D rotated box detection, object tracking.
12. I G SHELOMENTSEVA
Classification of Light Microscopy Image Using Probabilistic Bayesian Neural Network
Probabilistic learning is supervised learning and involves the formation of a probability distribution function for each class. Probabilistic Bayesian learning for each new input vector uses Bayes' rule to assign it to the class with the highest posterior probability. The authors illustrate the implementation of a classification procedure based on probabilistic Bayesian neural networks for classifying light microscopy images. The article proposes two hybrid models of probabilistic Bayesian neu
13. KABACHENKO FEDOR, SAMARINA ALENA, MIKHAYLIK YAROSLAV
1Peter the Great St. Petersburg Polytechnic University
2University of Manitoba
Development of the convolutional neural network for defining the renal pathology using computed tomography images
It is known that impaired kidney function leads to a deterioration of human health and quality of life. Modern medicine offers the method of computed tomography (CT) for the diagnosis of various types of renal pa-thology. At the same time, the risk of incorrect diagnosis by the doctor himself is still high, even with wide practical experience. The quality of analysis by radiologists of CT scans can be significantly improved using Convolutional Neural Networks (CNN). This paper represents the devel-opment of the CNN for the task of multiple classification of renal patholo-gy, as well as the modification of the algorithm using alternative activation functions.
SESSION 3
Tuesday, October 18 14:00 – 16:00
Lecture-hall Зал ученого совета
Chair: Prof. MALSAGOV MAGOMED
Applied neural systems
14. * GADZHIEV I.M., MYAGKOVA I.N., DOLENKO S.A.1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Use of Classification Algorithms to Predict the Grade of Geomagnetic Disturbance
This paper presents different approaches for predicting the grade of geomagnetic Kp index using machine learning algorithms. The Kp index is considered to be an indicator of the energy input from the solar wind into the Earth’s magnetosphere. Kp index forecasting can help predicting magnetic storms causing electrical machinery malfunction.
15. * GUSKOV A.A., LAPTINSKIY K.A., BURIKOV S.A., ISAEV I.V.
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Integration of data and algorithms in solving inverse problems of spectroscopy of solutions by machine learning methods
This study presents the results of solving the inverse problem of determining the concentrations of heavy metal ions of multicomponent solutions by their Raman and absorption spectra using integration of machine learning algorithms and integration of optical spectroscopy methods. It is shown that if the integrated methods differ much by their accuracy, then their integration is not effective. This is observed both for algorithmic integration and for integration of physical methods.
16. * BARINOV ROMAN OLEGOVICH, GAI VASILIY EVGENEVICH, POLIAKOV IGOR VLADIMIROVICH, KUZNETSOV GEORGIY DMITRIEVICH, TISHCHENKO ANDREI ALEKSANDROVICH
Nizhny Novgorod State Technical University named after R.E. Alekseev
Model and algorithms for automatic evaluation of learning outcomes of a neural network
The problem of automatic evaluation of learning outcomes of a neural network based on the analysis of learning curves is considered. The analysis of learning curves is reduced to extracting the proposed feature description and subsequent classification using a classical machine learning model. The proposed automatic estimation model serves to increase the degree of automation and interpretation of the learning process of the neural network.
17. * KARIMOV E.Z., SHIROKII V.R., MYAGKOVA I.N.
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Application of machine learning methods for domain adaptation spacecraft data
In this paper, the problem of domain adaptation of data used to predict the geomagnetic Dst index using machine learning methods is considered. Domain adaptation is necessary when switching from the data of one spacecraft to the data of another. Several methods of data translation from the DSCOVR (Deep Space Observatory) spacecraft domain to the ACE (Advanced Composition Explorer) spacecraft domain are considered and compared. It is shown that domain adaptation of data makes it possible to improve the quality of forecasting.
18. * RODIONOV D.M., KARCHKOV D.A., MOSKALENKO V.A., NIKOLSKY A.V., OSIPOV G.V., ZOLOTYKH N.YU.
N.I. Lobachevsky State University of Nizhni Novgorod
Possibility of using various architectures of convolutional neural networks in the problem of determining the type of rhythm
It is known that the most dangerous predictors of sudden cardiac death is the appearance in a person of signs of atrial and ventricular fibrillation. At the same time, this type of rhythm can rarely be detected during a quick screening examination, which greatly complicates the diagnosis and, as a result, the choice of effective treatment. Within the framework of this article, the experience of using the most popular architectures of convolutional neural networks adapted for the analysis of the electrocardiogram signal is presented to localize areas of sinus rhythm and fibrillation. As the networks under study, such architectures as ResNet, DenseNet and XceptionNet were considered. Each of their architectures showed good results, but the DenseNet network showed the best results. Due to a number of advantages described in the article, the DenseNet network was chosen. It should be noted that the chosen architecture of the neural network, after a number of necessary modifications and improvements, will be integrated into the existing diagnostic complex for cardiovascular diseases, developed on the basis of Lobachevsky University.
19. * RUSOV D., GONCHAROV P., SHCHAVELEV E., LUBCHENKOV L., NIKOLSKAIA A., REZVAYA E., OSOSKOV G., ZHEMCHUGOV A.
1
2Sanct-Petersburg State University
3Joint Institute for Nuclear Research
Recurrent and Graph Neural Networks for Particle Tracking at The BM@N Experiment
This work presents a new two-step approach for elementary particle tracking that combines the advantages of both local and global tracking algorithms. On the first stage, where the graph of possible track-candidates is too big to fit into memory, a recurrent neural network model TrackNETv3 is used for building track candidates. On the second stage there is a graph neural network GraphNet needed for clearing the graph from the fake segments. The results of testing the proposed approach on the 3,2 GeV Ar+Pb simulation for the BM@N RUN7 are presented.
20. * V.V. OLONICHEV, B.A. STAROVEROV, S.D. TARASOV
Kostroma State Technological University
Dynamic Processes Serial Parametric Identification with the Neural Networks
The problems of solving the problem of the function approximating changing temporary sequencies parameters identification and nonlinear dynamic objects identification with the help of neural networks are considered in the current work. A quazyreservoir sequental method of deep learning using the TensorFlow library and a scheme of its realization based on using two neural networks, namely identificational and imitational, is suggested. A serial-parallel scheme of nonlinear dynamic object identification is also considered. The research results are presented. The possibility of practical utilisation of suggested methods is proven, and the practical results of research are presented. The suggested methods are more efficient compared with the methods of direct parametrical identification with the help of neural networks.
21. * V.I. TEREKHOV, D.O. ISHKOV
Bauman Moscow State Technical University
NEURAL ARCHITECTURE SEARCH IN SOLVING THE PROBLEM OF APPROXIMATION OF DYNAMIC ANAMORPHIC METHOD
The article searches for optimal configurations of three neural network architectures using two methods. The neural network solves the problem of approximation of the dynamic anamorphic method. A dataset consisting of visual images and irregular transformation rasters has been collected. The differences and similarities between the genetic algorithm and Bayesian search are investigated. Both methods find optimal configurations with comparable speed, while the found configurations of the genetic algorithm are more diverse.
POSTER SESSION 1
Tuesday, October 18 16:30 – 18:00
Lecture-hall Фойе 2 этажа
Chair: Prof. BAKHSHIEV ALEKSANDR VALERYEVICH
Adaptive behavior and evolutionary modelling
22. ZAREMA B. SOKHOVA AND VLADIMIR G. RED’KOScientific Research Institute for System Analysis, Moscow
A model of predicting and using regularities by an autonomous agent
In this article, a computer model of an autonomous agent that functions in an environment with different levels of environmental illumination is developed and investigated. The level of illumination changes periodically. The task of the agent is to learn to eat in those moments of time when the illumination is radically increasing (at sunrise), and do not anything in the rest of the time. When the agent is feeding at sunrise, the agent receives essential positive reinforcement and its resource grows, otherwise the agent loses a small amount of resource. Two computer methods were used to train the agent: 1) the SARSA method, which is well known in the theory of reinforcement learning, and 2) the neural network method. Computer simulation demonstrated how the agent successfully learns and functions, accumulating a resource. It is shown that the results obtained by the SARSA method and the neural network method coincide. The model can be considered as a stage of investigation of more general properties of autonomous cognitive agents.
23. SAMER EL-KHATIB, YURI SKOBTSOV , SERGEY RODZIN
1Southern Federal University, Rostov-on-Don
2Saint Petersburg State University of Aerospace Instrumentation
Optimal number of ants determination in image segmentation method for complexly structured images
images segmentation, complexly structured images, ant colony optimization, k-means algorithm
24. S.A. POLEVAYA, L.V. SAVCHUK, GROMOV K.N., A.I. FEDOTCHEV, S.B. PARIN
1N.I. Lobachevsky State University of Nizhni Novgorod
2Institute of Cell Biophysics, Pushchino, Moscow region
Display of school disadaptation in the mode of vegetative regulation
An experimental model has been developed that makes it possible to reproduce the basic components of cognitive activity. The experimental model includes 3 functional contexts: extreme cognitive load - "stress"; optimal cognitive load - "school activity"; relaxation - relaxation. It was found that successful children show a direct relationship between the intensity of cognitive load and the tension of regulatory systems. In children with maladaptation, such a relationship is absent.
Artificial intelligence
25. SAMARINA ALENA IGOREVNA, KORCHIGIN ALEXANDR ALEXANDROVICHPeter the Great St. Petersburg Polytechnic University
AUTOMATION OF HISTOLOGICAL STUDIES TO DETECT EARLY ONCOLOGICAL PATHOLOGY OF THE BREAST
The article describes the automation of histological studies for the detection of early invasive ductal carcinoma of breast carcinoma using a convolutional neural network. The subject of the study is histological images of breast tissue, classified according to the characteristics of "healthy tissue" and "cancer tissue". In this paper, we compared transfer learning models based on neural networks EfficientNetB7, InceptionResNetV2, VGG19. The best learning result was shown by the model based on EfficientNetB7, it was chosen for integration into the software and hardware complex.
26. VICTOR E. KURYAN
MODELING OF THE HUMAN LEARNING PROCESS. AXIOMATIC APPROACH
The paper describes an approach to the construction of an automatic translation system from one natural language to another, which simulates the process of teaching a foreign language to a person. The training of the system consists in the automatic construction of a model of the world based on the analysis of the input training expressions in the input and output languages. The world model is built in the form of a graph. The rules for constructing the graph of the world model are formulated in the form of rules - learning axioms. Learning axioms do not depend on the choice of specific languages. The use of the world model in translation leads to the appearance of explainability when obtaining the results of the automatic translation system.
27. SEREDA IANA ALEXANDROVNA, NIKONOROV ILYA DMITRIEVICH, PREOBRAZHENSKAYA YULIA DMITRIEVNA
N.I. Lobachevsky State University of Nizhni Novgorod
AGENT LEARNING MODEL BASED ON COMPLETING A COGNITIVE MAP
An approach to the training of intelligent agents is proposed, in which the agent's long-term memory stores the learned action programs. It is shown that with such a memory structure, it is possible to use the classical apparatus for testing statistical hypotheses as a mathematical apparatus for learning from several samples. The proposed approach to learning has biological prerequisites.
28. E.I. ZAYTSEV, E.V. NURMATOVA
MIREA - Russian Technological University
ON THE ORGANIZATION OF THE MULTI-AGENT KNOWLEDGE REPRESENTATION AND PROCESSING SYSTEM
This paper discusses the structure and functional organization of the Multi-Agent Knowledge Representation and Processing System (MAKRPS). An agent-based approach to the development of distributed intelligent systems, which allows to integrate knowledge-based reasoning mechanisms with neural network models, is proposed. The structure of multi-agent solver, methods that implement different types of queries to the Knowledge Base, and methods for reinforcement learning virtual agents are described. The problem of optimizing the logic of data structure distribution among the nodes of the system in order to increase the speed of processing requests to the knowledge base is considered. Data collection from remote nodes of the Multi-Agent Knowledge Repre-sentation and Processing System is carried out in the operational databases of large volumes, which are synchronized to solve various kinds of analytical problems. The optimal structure ensures the efficiency of the MAKRPS on multi-computers. As a result of solving the optimization problem, the system is divided into a number of clusters that have minimal information connectivity with each other. Solving the problem of synthesizing the optimal data structure is of great practical value for designing the logical structures of knowledge bases, for forming the specifications of the queries and customizing the knowledge base.
29. GRISHIN I.A., SAKHAROVA E.K., USTINOV S.M., KANEV A.I., TEREKHOV V.I.
Bauman Moscow State Technical University
Tree Inventory with LiDAR Data
At present, the rational use of forest resources requires a constant assessment of the state and the implementation of a forecast of the dynamics of the forest fund. For these purposes, it is necessary to update existing data on forest areas in a timely manner. With the spread of technologies for remote data collection, there is a need to improve methods for measuring the taxation parameters of stands. Thus, the paper proposes a methodology based on various methods for calculat-ing taxation parameters and obtained results, in some cases better than the results of existing solutions. The methods of fitting a circle for further estimation of the diameter of a tree trunk are compared: Least Square fit and HyperLS fit. A com-parative analysis of the results of measurements of parameters with the results of measurements made using similar software in relation to the data collected by field measurement methods was carried out.
30. SMOLIN VLADIMIR SERGEEVICH, ZHURAVLEV DMITRY VLADIMIROVICH
1Keldysh Institute of Applied Mathematics, Moscow
2
THEORIES OF KNOWLEDGE, CONSCIOUSNESS, EMOTIONS AND OTHER "HUMAN" PROPERTIES FOR ROBOTS
Approximations carried out by modern neural network algorithms have made it possible to solve a wide range of applied problems. The implementation of the complex tasks decomposition and the knowledge hierarchical representation in neural networks will significantly speed up the neural networks training and advance into the area of even more complex tasks. Analysis of the properties of the work of the hierarchical structure of knowledge representation based on approximation and decomposition allows to take a fresh look at such concepts as cognition, consciousness and emotions.
Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling
31. SHATS V.N.Independent investigator, St. Petersburg
Data structure approximation in the classification problem
The paper introduces the concept of proximity for a finite set of objects and proposes a classifier based on this concept. The data of the combined sample for each feature were approximated by mapping onto sets of ordered pairs containing the numbers of objects with reasonably close values of the feature. Then the information contained in the training sample for any feature for objects of individual classes is approximately represented as a set of lists of objects that form these pairs. The frequency of any element of the list is determined as a complex event, and classes of objects according to the simplest formulas of probability theory.
32. DEMCHEVA A.A., KORSAKOV A.M., BAKSHIEV A.V.
1Peter the Great St. Petersburg Polytechnic University
2Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
Pain Sensation Model based on the Compartment Spiking Neuron Model (CSNM)
The paper considers a pain sensation model based on the Gate control theory by R. Melzak and P.D. Wall. The model is defined through the use of CSNM (Compartment Spiking Neuron Model), which allows to describe and change the structure of a neuron online (cell body size, number and length of dendrites, number of synapses) depending on incoming spiking pattern. The paper provides a detailed description of proposed pain model and indicates the main differences between the model and the Gate Control scheme. The performance of proposed model was experimented and evaluated with solving the binary classification problem. The results of modeling have shown the 99.3% classification accuracy, which allows to conclude about the practical application. The proposed model can be used in neuromorphic systems as a part of system state control or as an Invert-er (NOT logic gate). In the future, the proposed model can be used to implement behavioral functions, as well as a part of control mechanism in neuromorphic systems.
33. SERGIN V.YA.
Institute of Mathematical Problems of Biology RAS
Nature and neural mechanisms of consciousness
The article discusses the nature and neurophysiological mechanisms of sensory awareness, the generation of thought and discursive thinking. It is shown that sensory awareness and the generation of thought are based on high-frequency cyclic processes of auto-identification. Discursive thinking is based on cyclic processes of sensory-motor rehearsal. The interacting mechanisms of sensory-motor rehearsal and auto-identification allow us to form images, scenes and dialogues, observe and change them, creating a moving and controlled world of conscious experience.
Neurobiology and neurobionics
34. BOZHOKIN SERGEY VALENTINOVICHPeter the Great St. Petersburg Polytechnic University
Neurocardiology: wavelet analysis of heart rate turbulence
To analyze the turbulence of the heart rhythm, the method of wavelet analysis of a frequency-modulated signal was used, in which heart beats occur at true times separated by different cardio intervals. Criteria are formulated for finding the local frequency, which is characterized by strong heterogeneity associated with extrasystole. The behavior of the local frequency is analyzed over the entire continuous time interval both before the extrasystole and after the extrasystole. Local rate averaging over different time intervals is compared with traditional heart rate turbulence parameters such as turbulence onset and turbulence slope.
35. VORONKOV G.S.
Lomonosov Moscow State University
UNKNOWN CHARACTERISTIC OF THE "NECKER CUBE» PHENOMENON: THE TWO 3D CUBE IMAGES ARE MIRRORED TO EACH OTHER
The paper describes a previously unknown characteristic of the "Necker Cube" visual phenomenon (NCph), namely, that two three-dimensional (3D) cube images, alternately visible in a "flat" NC, are mirrored in relation to each other. The analysis of NCph in this aspect allowed us to form an idea of the neural model representing 3D visual space in the monocular visual system; this made it possible to describe the NCph "mechanism" in neurophysiological terms. The possibility of a phenomenon similar to visual NCph in thinking is discussed.
36. NAZHESTKIN I.A.
The Moscow Institute of Physics and Technology (State University)
Approximation methods for calculation of integrated information coefficient for investigation of neuroplasticity during learning
The integrated information coefficient Φ is a useful metric to describe the degree of brain adaptation to the environment. But its calculation is hindered. In this work, three approximation methods were tested on the neural activity data from the rat hippocampus during the operant learning. It was shown that Φ calculated using all three methods successfully predicts a degree of lerning success.
37. TITOV VADIM EVGENYEVICH, DICK OLGA EVGENIEVNA
1Saint Petersburg State University of Aerospace Instrumentation
2Pavlov Institute of Physiology of the Russian Academy of Sciences, St.Petersburg
FREQUENCY ANALYSIS BASED ON SYNCHROSQUEEZED WAVELET TRANSFORM OF BRAIN AND HEART RHYTHMS IN BRAIN VASCULAR PATHOLOGY
The application of synchrosqueezed wavelet transform to assess the relationship between brain and heart rhythms in vascular pathology of varying severity before and during hyperventilation load is considered. The identified features of the frequency relationships between brain and heart rhythms during hyperventilation load can be useful in the search for neurophysiological correlates of the severity of cerebral vascular pathology.
Neuromorphic computing and deep learning
38. KRASNOV MIKHAIL MIKHAILOVICH, SMOLIN VLADIMIR SERGEEVICHKeldysh Institute of Applied Mathematics, Moscow
Differential properties using to improve the function approximation quality by neural networks
The BPE (backpropagation error) use Is described for the approximation of functions that are specified not explicitly, but in the properties described by a differential equation and initial values form. A method for forming neural network structures for calculating an approximating function partial derivatives is proposed, and a nonlinear transformation, based on the selected formal neurons initial nonlinear function is derived . The Hopf equation is given as an example.
39. SHAMIN ALEXANDER YURIEVICH, KARANDASHEV YAKOV MIKHAILOVICH
Scientific Research Institute for System Analysis, Moscow
ON A NEURAL NETWORK APPROACH TO SOLVING DIFFERENTIAL EQUATIONS
A generalization of the neural network method for the numerical solution of differential equations, both ordinary and partial, 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 is trained to solve not only one problem with fixed initial conditions, but a whole class with conditions in a certain range. Also, the paper considers several architectures of such networks with several training options.
SESSION 4
Tuesday, October 18 18:00 – 20:00
Lecture-hall Зал ученого совета
Chair: Prof. RATUSHNYAK ALEXANDER SAVELYEVICH
Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling
40. SILKIS I.G.Institute of Higher Nervous Activity and Neurophysiology of RAS, Moscow
Neural network providing the involvement of voluntary attention into the processing and conscious perception of sensory information
A hypothetical mechanism for processing and awareness of sensory information is proposed. Direction of voluntary attention to a stimulus evokes dopamine release that promotes dopamine-dependent plastic reorganizations of neural interactions in the network consisting of topographically connected areas in the neocortex, hippocampus, basal ganglia, thalamus and cerebellum. These reorganizations result in disinhibition of thalamic cells and increased firing of cortical and hippocampal neurons, thereby facilitating activity circulation and re-excitation of cortical areas required for the conscious perception of stimulus
41. VICTOR VVEDENSKY, IVAN FILATOV, KONSTANTIN GURTOVOY, MIKHAIL SOKOLOV
1National Research Centre "Kurchatov Institute", Moscow
2Lomonosov Moscow State University
3Child Technopark of Kurchatov University, Moscow
Alpha rhythm dynamics during spoken word recognition
We combined measurements of the spoken word recognition time with simultaneous EEG recording from the subjects. Alpha waves of all our subjects were sharp and we could accurately detect time instants of the tips of the waves even when the shape of the waves varied. Time intervals between the tips were constant during 5 to 10 alpha waves. We observed that these trains of oscillations change their period just at the moments of the word sound onset and when the subject recognized the word and confirmed this by pressing the button. This experiment shows that alpha rhythm is related to or even organizes the complex process of the spoken word recognition.
42. * PUGAVKO M.M., MASLENNIKOV O.V., NEKORKIN V.I.
The Institute of Applied Physics of the Russian Academy of Sciences, Nizhny Novgorod
Dynamics of a recurrent spiking neural network in performing several tasks of cognitive neuroscience
The paper presents a recurrent spiking neural network, which is trained by machine learning methods to solve several different target tasks inspired by experiments in the field of cognitive neuroscience. The dynamic mechanisms of the performance of the target functions by the network are studied which consist in the appearance of certain trajectories in the space of population activity of neurons. The hierarchical structure of the network after training is revealed.
43. * ONUCHIN ARSENII ANDREEVICH, KACHAN OLEG NICKOLAEVICH
1Lomonosov Moscow State University
2Skolkovo Institute of Science and Technology (Skoltech)
Individual Topology Structure of Eye Movement Trajectories
Traditionally, extracting patterns from eye movement data relies on statistics of different macro-events such as fixations and saccades. This requires an additional preprocessing step to separate the eye movement subtypes, often with a number of parameters on which the classification results depend. Besides that, definitions of such macro events are formulated in different ways by different researchers. We propose an application of a new class of features to the quantitative analysis of personal eye movement trajectories structure. This new class of features based on algebraic topology allows extracting patterns from different modalities of gaze such as time series of coordinates and amplitudes, heatmaps, and point clouds in a unified way at all scales from micro to macro. We experimentally demonstrate the competitiveness of the new class of features with the traditional ones and their significant synergy while being used together for the person authentication task on the recently published eye movement trajectories dataset.
44. GANEEVA VERONIKA ALEKSANDROVNA, KLYSHINSKIY EDUARD STANISLAVOVICH
1National Research University "Higher School of Economics", Moscow
2
Investigation of Interpretability of Static Models for the Russian Language
In this paper, we investigate the issue of semantic interpretation of Word2Vec vector models. The LSA-based algorithm of extraction of semantically interpretable models was proved on several static models. We use a set of dictionaries for the sake of the model evaluation.
45. * SELEZNEV L.E., CHUPAKHIN A.A., KOSTENKO V.A., VARTANOV A.V., SHEVCHENKO A.O.
Lomonosov Moscow State University
RECOGNITION MENTALLY PRONOUNCED PHONEMES USING CONVOLUTIONAL NEURAL NETWORKS BASED ON ELECTROENCEPHALOGRAPHY DATA
The problem of binary and multiclass classification of mentally pronounced phonemes of the Russian language is considered on the basis of data obtained from a multichannel electroencephalography device. The methodology of the data collection experiment is described, as well as the methods of processing the data obtained. A classification algorithm based on convolutional neural networks is proposed, implemented and investigated.
46. * KAPUSTNIKOV ANTON A.
1Saratov State University
2
MODELING OF EPILEPSY USING SPECIALIZED NEURONS OF DIFFERENT TYPES
An ensemble of 14 coupled neurooscillators for modeling epilepsy was demonstrated, in which five types of neurons are present, appropriate physiological models and parameters were selected for each, in addition, due to the presence of excitatory and inhibitory connections in the structure, physiological synapse models for AMPA and GABA receptors were used, respectively. As a result, it is shown that the system is able to demonstrate transients that simulate epileptiform activity
47. SHEPELEV DANIL IGOREVICH, SAEVSKY ANTON IGOREVICH, SHEPELEV IGOR EVGENIEVICH, SHAPOSHNIKOV DMITRY GRIGORIEVICH, LAZURENKO DMITRY MIHAILOVICH
1Southern Federal University, Rostov-on-Don
2Scientific Research Center of Neurotechnologies, Southern Federal University, Rostov-on-Don
3
A Software System for Training Motor Imagery in Virtual Reality
This paper describes the hardware and software system which comprises the brain-computer interface system and the virtual reality environment. The software package includes a virtual reality game application, which is controlled by mental equivalents of real movements. The key aspect of our approach is a game inter-face in a virtual reality environment. The virtual reality contributes to deep immer-sion of the user in the process of ideomotor training. A neural network classifier, which was trained and tested using publicly available data, was implemented to solve the problem of recognition and classification of mental ideomotor com-mands and control in virtual reality environment.
SESSION 5
Wednesday, October 19 12:00 – 13:00
Lecture-hall Концертный зал
Chair: Prof. DOLENKO SERGEY
Neuromorphic computing and deep learning
48. ARIJ AL ADELThe Moscow Institute of Physics and Technology (State University)
Global memory transformer for processing long documents
This paper verifies the ability of previous introduced model to process chunked inputs on two new tasks; masked language modeling and question answering task and using the compressed chunk representations in autoregressive generation.
49. DMITRY ANTONOV, SERGEY SUKHOV
Mechanisms of catastrophic forgetting prevention in spiking neural networks
Artificial neural networks experience catastrophic forgetting during sequential learning. The peculiarities of spiking neural networks (SNNs) provide additional mechanisms for mitigating of catastrophic forgetting. In this paper, we study the effect of lateral inhibition and proper account of neurons’ importance for catastrophic forgetting prevention in SNNs. The experiments were performed on freely available datasets in the SpykeTorch framework.
50. KRASNIKOV V.V., CHEZHEGOV A.A
Lomonosov Moscow State University
Neuromorphic photoelectric synapses based on metal oxides nanocrystallites
This work present the implementation of a photoelectric synapse based on semiconductor nanocrystallite (ZnO, In2O3, WO3). By choosing the type of nanocrystallite and manufacturing parameters, we obtain neuromorphic elements with variable properties. In the form of a synaptic signal, light pulses are used at a wavelength of 405 nm. The change in the conductivity of a nanocrystallite upon irradiation with light corresponds to the change in the resistance on the postneuron.
51. * CHAPLINSKAYA N.V.
V. A. Trapeznikov Institute of Control Sciences of Russian Academy of Sciences, Moscow
Spike neural network training based on synaptic delays changing
The article describes a new learning rule for training a spiking neural network based on the biological effect of the synaptic delays plasticity. Using this rule, a spiking neural network of three neurons was successfully trained to recognize three contrast patterns.
SESSION 6
Wednesday, October 19 14:00 – 15:00
Lecture-hall Концертный зал
Chair: Prof. DOLENKO SERGEY
Applied neural systems
52. BEKSULTAN SAGYNDYK, DILYARA BAYMURZINA, MIKHAIL BURTSEV1The Moscow Institute of Physics and Technology (State University)
2London Institute for Mathematical Sciences, United Kingdom
DeepPavlov Topics: Topic Classification Dataset \ for Conversational Domain in English
This paper presents “DeepPavlov Topics”, a new dataset for topic classification in conversational domain. The dataset was collected and filtered automatically from web-sites and open datasets. We identify 33 topics, and present full (4.2M samples) and down-sampled (2.2M samples) versions of the “DeepPavlov Topics”. The proposed topics are aimed to cover conversational domain in details but maintain interpretability. We report baseline classification results trained in multi-label setup to allow multiple classes per text during inference. We also release pre-trained models for topic classification including distilled and multi-lingual versions.
53. CHIZHIKOVA A.P., KONOVALOV V. P., BURTSEV M.S.
1The Moscow Institute of Physics and Technology (State University)
2London Institute for Mathematical Sciences, United Kingdom
Multilingual Case-insensitive Named Entity Recognition
In this paper, we address the task of Named Entity Recognition on inconsistently capitalised data in English and Russian. We demonstrate that using multilingual BERT trained on a concatenation of original and lowered datasets is the most effective way to solve the task. Our model achieves the highest average result on CoNLL-2003 and Collection 3 datasets while being robust to missing casing.
54. KNIAZ VLADIMIR, KNYAZ VLADIMIR, PETR MOSHKANTSEV
1State Research Institute of Aviation Systems, Moscow
2The Moscow Institute of Physics and Technology (State University)
IQ-GAN: Instance Quantized Image Synthesis
For humans, it is natural to decompose an image into objects and background scene. Still, modern generative models usually analyze image at the scene level. Hence, it is challenging to control the style and quality of individual object instances. We propose an instance-quantized conditional generative model for the synthesis of images with high-fidelity instances of multiple classes. Specifically, we train two generators simultaneously: a scene generator that synthesizes the background environment and an instance generator that synthesizes each object instance individually. We design a differentiable image compositing layer that assembles the resulting image and allows effective error back-propagation. For our generators and we developed a new architecture leveraging modulated convolutional blocks. We evaluate our model and baselines on ADE20k, MHPv2, and Cityscapes datasets to demonstrate that our instance-quantized framework outperforms baselines in terms of FID and mIoU scores. Moreover, our approach allows us to separately control the style of each object and learn fine texture details. We demonstrate the effectiveness of our framework in a wide range of image manipulation tasks. We made our code publicly available.
55. * YULIYA A. TSYBINA, ALEXEY A. ZAIKIN, AND SUSANNA YU. GORDLEEVA
1N.I. Lobachevsky State University of Nizhni Novgorod
2University College London
Information processing in spiking neuron-astrocyte network in ageing
Recent experimental works revealed that astrocytes are the critical players involved in dynamic interaction with neurons in information processing and storage. Several studies provided direct evidence of the involvement of abnormal astrocytic signaling in cognitive impairment. Astroglial atrophy is observed in many age-dependent neurodegenerative diseases. Astroglial morphological atrophy underlies the decrease in synaptic coverage, which in turn contributes to a decrease in the cognitive abilities of the brain. However, in many cases, the understanding of astrocytic involvement is still incomplete. Now that we better understand the molecular components of astrocyte-neuron interactions, the new challenge is to investigate how they integrate at the network and physiological levels. To address this question, novel, bio-inspired, and detailed computational models should be developed to simulate the information processing through the coordinated activity of both astrocytes and neurons. In this work, we studied the effect of astroglial atrophy on information processing in model of spiking neuron-astrocyte network that implements the functions of short-term memory.
SESSION 7
Wednesday, October 19 15:00 – 16:00
Lecture-hall Концертный зал
Chair: Prof. MAIOROV VLADIMIR IVANOVICH
Neurobiology and neurobionics
56. MAIOROV VLADIMIR IVANOVICHLomonosov Moscow State University
ABCD OF NEUROMORPHIC INTELLIGENCE
The elements of the neuromorphic core of intelligence are outlined. These are a) “attractors” – attractor groups of neurons in recurrent neural networks; b) "binding" - functional associations of groups based on bidirectional synaptic plasticity; c) “context/conditional” – the formation of contextual combinations and conditional sequences; d) "dendrites" - local integration in dendrites, "dopamine" - modulation of connections by dopamine, consistent with the dopamine theory of the origin of intelligence.
57. ALEXANDER RYLOV, TATIANA LEVANOVA, SERGEY STASENKO
N.I. Lobachevsky State University of Nizhni Novgorod
Classification of neuron type based on average activity
Neuronal activity recorded in experiments is the basis for the processing, storage and transmission of information in the brain, as well as functional states. It can also carry information about the structure and type of neuronal cells involved in these processes. This paper proposes a new approach to the analysis of this type of data using modern methods of data classification. It was found that the average activity representation of spike sequences carries information about the type of neuronal cells and allows to effectively classify the initial data.
58. ANTON I. SAEVSKIY, IGOR E. SHEPELEV, IGOR V. SHCHERBAN, DMITRY G. SHAPOSHNIKOV, DMITRY M. LAZURENKO
1Scientific Research Center of Neurotechnologies, Southern Federal University, Rostov-on-Don
2
COMPARATIVE ANALYSIS OF STATISTICAL AND NEURAL NETWORK CLASSIFICATION METHODS ON THE EXAMPLE OF SYNTHETIZED DATA IN THE STIMULUS-INDEPENDENT BRAIN-COMPUTER INTERFACE PARADYGM
In the present study, a comparison of approaches to the motor imagery clas-sification on the example of simulated signals is given. Data were simulated according to conventional consideration of motor imagery as channel-specific desynchronization at the µ-rhythm (about 10 Hz). Moreover, the is-sues of the need to compress the feature space at various stages of complex approaches and of extracting informative signal segments are considered. An approach to searching for informative segments based on the Fisher criterion is also presented. Finally, the results of convolutional neural networks appli-cation to the problem are given both for using whole signals and extracted informative shorter segments. All accuracies were obtained and analyzed at different comparatively small sample sizes and signal-to-noise ratios for sim-ulated data of three motor imagery classes and three data channels (~leads).
59. A.S. RATUSHNYAK, T.A. ZAPARA, A.L. PROSKURA, A.N. SKLYAROV, E.D. SOROKOUMOV,
1
2The Institute of Computational Technologies of SB RAS (ICT SB RAS), Novosibirsk
Analysis of Appearances, Formation and Evolution of Biological Functional Systems
The purpose of the work was to search for approaches to obtain fundamentally significant results when modeling biological information functional systems. In order to promote this issue, it is necessary to understand the physical principles that lie at the heart of their appearances, evolution and function. The analysis of the preconditions and principles of the emergence of these molecular machines, the evolution of primary prebions, the formation of them on the basis of all subsequent levels of living systems, neurons and the brain.
POSTER SESSION 2
Wednesday, October 19 16:30 – 18:00
Lecture-hall Фойе 2 этажа
Chair: Prof. KAGANOV YURI
Applied neural systems
60. GLEB S. BRYKINBauman Moscow State Technical University
DeOldify.NET: cross-platform application for coloring black and white photos
The problem of image coloring is an old problem that researchers have been solving for many years. The article is devoted to a cross-platform desktop application for automatic coloring of black and white images, which is a port of the well-known DeOldify development.
61. TATIANA LAZOVSKAYA, DMITRIY TARKHOV, ALINA DUDNIK, ELENA KOKSHAROVA, OLGA MOCHALOVA, DANIL MURANOV, KSENIA POZHVANYUK, ANASTASIA SYSOEVA
Peter the Great St. Petersburg Polytechnic University
Investigation of Pareto Front of Neural Network Approximation of Solution of Laplace Equation in Two Statements: with Discontinuous Initial Conditions or with Measurement Data
The paper uses a general neural network approach to solving partial differential problems. The solution is the output of a neural network with one hidden layer and linearly and nonlinearly adjustable input parameters (weights) during training. Network training is considered as a multi-criteria optimization problem solved by constructing a Pareto front and then choosing the optimal solution in some sense. An evolutionary algorithm for constructing solutions based on the Pareto front is proposed. The method is tested on two reference problems: the Laplace equation in the unit square with a discontinuous Dirichlet boundary condition (Motz's problem) and without initial conditions, using measurement data of varying accuracy. Optimal solutions were obtained in accordance with two different selection criteria.
62. VLADIMIR B. KOTOV AND GALINA A. BESKHLEBNOVA
Scientific Research Institute for System Analysis, Moscow
Specifics of Crossbar Resistor Arrays
Parallel neuromorphic computing is based on matrix-vector multiplications which can be realized using resistor arrays. The arrays should consist of variable resistors to enable the multiplication by vari-ous matrices. The elements that can change their resistance under the action of the current flowing through them are most interesting in this case. Among different sorts of resistor array the regular crossbar struc-tures are most manufacturable. The use of crossbar resistor arrays as storage elements or matrix-vector multipliers faces considerable diffi-culties. The biggest problem is how to record a needed conductance matrix with the limited number of control signals and strong inter-resistor connections. In parallel recording only specific conductance matrixes can be formed. In consecutive recording it is impossible to localize the action, not only targeted resistors change their conductance. In the paper the mathematical modelling is used to analyze the methods allowing us to generate a great variety of conductance matrices, though not permitting the formation of matrices of arbitrary types.
63. ENGEL EKATERINA ALEXANDROVNA, ENGEL NIKITA EVGENYEVICH
Katanov Khakass State University, Abakan
The intelligent times series forecasting framework
Most successful applications of machine learning to real-world forecasting tasks have been achieved using deep networks of rather large size disregard to task’s complexity and overfitting problems. This paper presents the effective intelligent times series forecasting framework which provides an automatic creation of an optimum architecture of the modified fuzzy neural net with regards to a forecast-ing task’s complexity. We verified the developed intelligent times series forecast-ing framework by solving the PV array power forecasting task. The validity and advantages of the developed intelligent two days ahead hourly PV array power forecasting system are demonstrated using numerical simulations. Based on the fuzzy hour’s state of cloudiness, the created modified fuzzy neural net provides an effective two days ahead hourly PV array power forecasting under various uncertainties. The simulation results show that the proposed intelligent two days ahead hourly PV array power forecasting system based on a modified fuzzy neu-ral net achieves competitive performance, as compared to a classical recurrent neural net trained by the Levenberg-Marquardt algorithm
64. MAKAROV M.V., SEMENOV I.A., DEMIDOV A.A., TRANTINA N.S.
Researching a new type of heuristic decisions for adaptive control of a mobile robot in a dynamic environment
The paper presents theoretical information that reveals the essence of a new type of heuristic decisions synthesized inside the components of a mobile robot control system. The methodology of experimental research aimed at substantiating the feasibility of one of the aspects of the synthesis of such decisions has been developed and implemented. It is revealed that the implementation of this aspect is possible within the neural network architecture and leads to an increase in the adaptive abilities of the robot in a dynamic environment of existence.
65. DMITRY M. IGONIN AND YURY V. TIUMENTSEV
Moscow Aviation Institute (National Research University)
Multi-Input Convolutional Neural Networks in Real-Time Semantic Segmentation Tasks
When processing the information obtained during remote sensing of the Earth’s surface, a number of tasks arise, including the task of semantic segmentation of the corresponding images. An approach to solving this problem directly in the process of obtaining information of remote sensing of the Earth’s surface as the observing object moves is considered. Such an approach is necessary for work in real-time systems. This problem is solved using multi-input convolutional neural networks, nine variants of such networks are considered, including variants with links between encoder and decoder blocks, as well as without such links. When comparing results of segmentation for these variants it is revealed, that the most suitable variant is a network without the mentioned connections between blocks. At increase in quantity of the coders, perceiving and processing sequences of the input patterns, accuracy of an extraction of the prescribed classes of objects raises. The optimum number of inputs (coders) for a given task is determined by the complexity of segmented images and the size of the training set. It is shown that multi-input networks make it possible to solve the task of semantic segmentation of satellite images with an increased accuracy. However, the quality of solving the problem of semantic segmentation significantly depends on the size and informativity of the training set. The minimum allowable size of the training and test sets in relation to the problem to be solved has been determined.
66. ARTAMONOV I., ARTAMONOVA Y.
1Moscow Aviation Institute (National Research University)
2NeuroCorpus
Multilevel separation pipeline for similar structure data
Text classification model based on multi-level filtering is considered. The features of social network messages that have to be taken into account for language processing are determined. Based on a large dataset of real data, it is shown that in case of high proximity between relevant and irrelevant messages widely spread methods of machine learning, ANN and NLP show unsatisfactory results. The idea of preclassification using context data structuring is proposed. The solution proposed in the article involves a combination of traditional methods, context-structuring and clustering in subsets of records. Their joint use makes it possible both to noticeably improve the quality of classification and to make the assessment of the quality of classification more controlled. We demonstrated that that the combined use of a large number of wide filters provides, on average, better classification quality than a smaller number of narrower ones. The proposed solution has been tested on data from social networks, but can be effectively applied to other text data with similar features.
67. ОLEG S. LITVINOV, АLEKSANDR N. ZABELIN, ANASTASIA A. RAKOVSKAYA
Bauman Moscow State Technical University
Modeling of a neural network algorithm for suppressing non-stationary interference in an adaptive antenna array
The existing adaptation algorithms do not provide quite deep suppression of non-stationary interference in time in some situations. Recurrent neural networks can improve the efficiency of suppressing such interference by predicting time series. This paper investigates the characteristics of the suppression of non-stationary blinking jamming in an adaptive antenna array (AAA) on a neural network amplitude-phase control. We used the example of a simulation model developed in the Matlab/Simulink software environment. The results of comparing the characteristics for the adaptation algorithm based on a recurrent neural network (LSTM) and for the traditional algorithm according to the criterion of the minimum power of noise and interference at the AAA output are presented. For the simulated case, the proposed algorithm provides an increase in the SNIR, which is important for radar, radio navigation and communication systems
68. V.GANCHENKO, A.DOUDKIN, А.INYUTIN YA.MARUSHKO
United Institute of Informatics Problems of the National Academy of Sciences of Belarus, Minsk
Neural network classification model based on the use of an autoencoder and built on the architecture of an ensemble of multilayer perceptrons
This report presents a neural network model for identifying and classifying objects in images obtained using UAV and orbital-based imaging equipment. The model is based on the use of an autoencoder and is built on the architecture of an ensemble of multilayer perceptrons. The model has an accuracy above 99% when classified into four classes: "Fire", "Smoke", "Vegetation" and "Building".
69. LITINSKII LEONID BORISOVICH
Scientific Research Institute for System Analysis, Moscow
ON PROCESSING OF EXHALED BREATH SPECTROGRAMS OF TEST GROUP OF PATIENTS BY PRINCIPAL COMPONENT METHOD
In this paper we give an account of our results on processing of a test fragment of a large medical dataset. The main medical problem is to be able to recognize on the exhaled breath spectrograms the medical conditions of patients or their predispositions to the given disease. We used the principal component method to process a test dataset and estimated the degree of separability between different classes.
70. LOBANOV ALEXEY VLADIMIROVICH, LAZAREVA GALINA GENNADIEVNA
Peoples’ Friendship University of Russia, Moscow
PRIMARY PREPARATION OF DATA FOR THE ANALYSIS OF BRAIN MSCT IMAGES FOR THE TASK OF SEARCHING THE PRIMARY SIGNS OF BRAIN STROKE
As part of the task of searching for areas with primary signs of cerebral stroke, there is a need for initial filtering of data from multislice computed tomography (MSCT) images of the brain in order to contrastly highlight potential abnormal areas. An algorithm for data preparation within the framework of the problem under study is presented.
71. SOBIANIN K.V., RUSAKOVA E.I.
Application of combined loss function and metric in U-Net training in the task of Automatic Segmentation of Acute Ischemic Stroke Lesions
The problem of semantic segmentation of acute ischemic stroke lesions on MRI images is considered. A combined loss function and metrics were proposed and applied. The UNet model was trained and tested with the pre-trained EfficientNet B4 encoder. The model showed a Dice coefficient of 66% on the test dataset.
72. EGOR MIKHAYLOVICH SYSOYKIN, IGOR DENISOVICH SHPAK, ARTEM ILYICH ANTONOV
Bauman Moscow State Technical University
COMPETITION NUMBER RECOGNITION USING AN ENSEMBLE OF DEEP LEARNING MODELS
The article deals with the problem of recognizing competition numbers of sports participants on images. A method for solving this problem has been proposed. It uses an algorithm that contains four steps: detection of a region of interest, detection of a polygon containing a number, recognition of numbers on the found areas and the integration of the results. An implementation based on the one-step detector YOLOv5 has been developed. The advantage of YOLO is the availability of tools to train the model on one’s own data and its high speed of inference. For model training have been prepared datasets which are based on open data. The results of trained models are also presented in this paper.
73. MARIA O. TARAN, GEORGIY I. REVUNKOV, AND YURIY E. GAPANYUK
Bauman Moscow State Technical University
Creating a Brief Review of Judicial Practice Using Clustering Methods
Based on the results of previous research in the field of legal text processing, this article discusses a number of high-level modules of the “hybrid intelligent information system for analysis of the judicial practice of arbitration courts” that can be used in the work of lawyers. The development of the clustering module “Court Clustering” is discussed. A brief overview of related works based on clustering approach is given. The clustering experiments result are discussed, it is shown that the proposed algorithm solves the clustering problem better than the considered analogues for all the main clustering quality metrics. The module for creating a brief review of judicial practice is discussed. Module output include certain statistical information on the analyzed documents as well as paragraphs that can help in preparing for the dispute. The analytical module is discussed. Module can generate various reports based on data extracted from court documents.
74. CHERNYCHEV L.S.
Prediction of time series values using a neural network when solving the problem of dynamic pricing
The paper gives an example of forecasting the values of the simulated time series, built by analogy with the time series of daily number of purchases of goods on the marketplace, using a neural network specially developed for this purpose. The obtained forecast will serve as the basis for building a dynamic pricing model. Comparison of the obtained forecast with the results of the standard time series forecasting model - the integrated moving average autoregressive model (ARIMA), the implementation of which is taken from the standard python library.
Neural network theory, neural paradigms and architectures
75. KARANDASHEV I.M., MIKHALCHENKO E.V., MALSAGOV M. YU., NIKITIN V.F.Scientific Research Institute for System Analysis, Moscow
APPROXIMATION OF CHEMICAL KINETICS BY ARTIFICIAL NEURAL NETWORK
Adding the pressure of the system as a new variable significantly complicates the task of approximating changes in the concentrations of substances in a mixture during chemical reactions. The value of the initial pressure affects both the value of the data and the rate of the reactions taking place. In this work, we managed to build a single neural network capable of describing the model at various pressures without sacrificing either the accuracy of the solution or the speed of work.
76. ALEXANDER EFITOROV, SERGEY BURIKOV, TATIANA DOLENKO, AND SERGEY DOLENKO
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
Use of Conditional Variational Autoencoders and Partial Least Squares in Solving an Inverse Problem of Spectroscopy
In this article, the solution of an inverse problem of Raman spectroscopy of water-ethanol solutions by neural network models is presented. Since the process of training a neural network requires a large dataset, which often cannot be collected by laboratory measurements, we propose the approach of generating synthetic pairs of patterns represented by continuous vectors. This was achieved by using the partial least squares (PLS) model as an additional embedding model of latent space generated by conditional variational autoencoder (cVAE). Random sampling procedure generates vectors in PLS score space, which allows assigning a vector in the cVAE latent space and an output vector of the inverse problem. The generated patterns have a shape similar to real spectra, and they are used to train the neural network on the multi-output regression task, along with the original patterns. Applying this approach results in significant improvement of the quality of solving the inverse problem on real patterns that were not used in the training process.
77. ANTSIPEROV V., PAVLYUKOVA E.
Neuromorphic image coding based on partitioning of counts by a system of receptive fields
The article is devoted to the synthesis of image encoding methods based on the images data themselves. The proposed approach is based on a previously developed special representation of images by samples of counts (sampling representations). Since the sampling representations are essentially random constructions, the synthesis of encoding methods is carried out strictly within the framework of the generative paradigm. In essence, the approach proposed treats the image coding within generative model as a special case of the classical statistical problem of probability distribution density estimation. In the paper, we restrict ourselves to the class of parametric estimation procedures, which imply some parametric family of probability distributions. Namely, we propose to use the model of a parametric mixture of simple distribution components. Accordingly, a set of component weights estimates calculated from a sampling representation, considering as input data, is interpreted as an encoded image – output data. In this context, optimal coding is synthesized with the maximum likelihood method. For the algorithmic implementation of the coding procedure, the mixture model is equipped with the structure of receptive fields that is a well-known organizing principal for receptors in the human eye retina. On this basis, we synthesize a relatively simple recurrent coding algorithm, which turned out to be close to the popular in machine learning EM algorithm. The paper presents interpretation of several features of the algorithm from the point of view of well-known facts about the image processing in the periphery of the visual system, discusses options for the algorithm implementation, and presents the results of numerical simulation of its operation.
78. PAVEL A. KOLGANOV, YURY V. TIUMENTSEV
Moscow Aviation Institute (National Research University)
An Attempt to Formalize the Formulation of the Network Architecture Search Problem for Convolutional Neural Networks
The paper deals with the problem of searching for a neural network architecture. The paper presents a mathematical formulation of the problem of searching a neural network model, optimal from the point of view of a predefined criterion. The analysis of components of this problem is given. Some difficulties encountered by researchers when solving the NAS problem are described. A computational experiment is conducted, which consists in the search of a neural network architecture on the MNIST dataset.
79. VLADIMIR B. KOTOV AND ZAREMA B. SOKHOVA
Scientific Research Institute for System Analysis, Moscow
Using a resistor array to tackle optimization problems
Resistor array allows us to realize vector-matrix multiplication. The use of variable resistors makes it possible to vary the matrix-multiplier. A neuromorphic system incorporating this sort of vector-matrix multiplication can be optimized with array resistor conductances. The optimization process needs the resistor array to allow concurrent formation of the conductivity matrix and reading this array (that is, multiplication by a proper matrix). It is also necessary to have an algorithm of finding an optimal conductivity matrix which can be easily realized and provide good chances of achieving a global optimum. The paper considers the concurrent write/read design of the resistor array. An algorithm that enables us to browse the space of realizable conductivity matrices and provides a sufficiently representative trajectory in the matrix space is described. The tuning of optimal matrix and the use of stationary conductivity matrices are discussed.
SESSION 8
Wednesday, October 19 18:00 – 18:45
Lecture-hall Концертный зал
Chair: Prof. DMITRY YUDIN
Adaptive behavior and evolutionary modelling
80. VLADIMIR B. KOTOV AND ZAREMA B. SOKHOVAScientific Research Institute for System Analysis, Moscow
On the importance of diversity
The structure optimization of a heterogeneous community with different objectives and values can level the distinctions between community members. In advanced self-sufficient communities, the objectives and values change in development. This fact and environmental changes cause the value ranking of different groups of a community to change. For this reason, the diversity is a necessary condition for steady progress of a society with ever-changing social values and environment. The paper takes simple population models of a heterogeneous community as examples to consider the issues of the diversity preservation. Different resource distribution principles yielding different relations between optimization and diversity preservation are presented. Contradictions between optimization and diversity preservation are particularly acute in segregated social groups. The possibility of intergroup transitions smooths the contradictions and increases the flexibility of optimization and the ability of the community to adapt to the changeable environment.
81. * 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
ADAPTIVE ALGORITHM FOR REAL-TIME TASKS TRACKING
The article examines the construction of an adaptive tracking algorithm. The solution involves the accumulation and updating of knowledge about the multidimensional movement of objects in the process of controlling their recognition in real time. The accumulation of knowledge takes into account the assessment of the quality of cognition, which is used in the management process. The algorithm forms new multilevel images from a set of already recognized ones. The simulation is based on a real video stream.
82. * KUPRIYANOV GAVRIIL ALEXANDROVICH, ISAEV IGOR VIKTOROVICH, DOLENKO SERGEY ANATOLYEVICH
1Lomonosov Moscow State University
2Skobeltsyn Institute of Nuclear Physics Lomonosov Moscow State University
A Gender Genetic Algorithm and its Comparison with Conventional Genetic Algorithm
This study presents a new variety of gender genetic algorithm (GGA). Using the example of five test optimization problems, its superiority over conventional genetic algorithm (GA) with an elitism operator is shown. The analysis of the novel GGA on multi-extreme optimization functions confirms the effectiveness of the idea of gender separation of the population. Based on the results obtained from experimental data and on their analysis, a more advanced modification of GGA is proposed, using determination of the offspring's gender based on the combination of parental chromosomes in it.
SESSION 9
Wednesday, October 19 18:45 – 20:00
Lecture-hall Концертный зал
Chair: Prof. VVEDENSKY VIKTOR
Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling
83. IRINA K. MALASHENKOVA, VADIM L. USHAKOV, SERGEY A. KRYNSKIY, DANIIL P. OGURTSOV, NIKITA A. KHAILOV, EKATERINA I. CHEKULAEVA, EKATERINA A. FILIPPOVA, VYACHESLAV A. ORLOV, NIKOLAY A. DIDKOVSKY, NATALIAAssociations of morphometric changes of the brain with the levels of IGF1, a multifunctional growth factor, and with systemic immune parameters reflect the disturbances of neuroimmune interactions in patients with schizophrenia
The development of neuroimaging methods allows to obtain new data on the brain functioning to find neuroimaging patterns of psychopathological conditions, creating a basis for further application of neuroscience research results in practical medicine. The involvement of insulin-like growth factor 1 (IGF1), which has a wide range of biological functions that include stimulation of glucose metabolism and transport, stimulation of cell proliferation and differentiation, inhibition of cell apoptosis and immune process regulation, is of great interest in the pathogenesis of schizophrenia. The aim of the work was to study possible associations of structural brain changes according to fMRI data with a multifunctional growth factor, IGF1, and systemic inflammatory immunity parameters for multilateral assessment of the state of neuro-immune interrelations in schizophrenia. 52 schizophrenic patients and 67 healthy volunteers were enrolled into the study. An interdisciplinary approach was used, including psychological testing (clinical method), neurophysiological and immunological testing, and cross-methodological calculations. Patients with decreased IGF-1 levels had the reduced average cortical thickness in a number of cortical areas in the left and right hemisphere. Patients with low IGF1 levels had higher systemic inflammation and signs of autoimmune reactions. Assessment of association of IGF-1 levels with clinical symptoms has shown that decreased IGF-1 levels were associated with motor disorders. The results show that the level of IGF1 growth factor reflects the features of immune-inflammatory disorders, the severity of brain morphometric disorders and extrapyramidal disorders in patients with schizophrenia. The identified associations show the presence and significance of impaired neuroimmune interactions in schizophrenia.
84. BARTSEV S.I., MARKOVA G.M.
1
2Siberian Federal University, Krasnoyarsk
Does a recurrent neural network use reflection during a reflexive game?
We tested the ability of simple recurrent neural networks to use the reflection of playmate during the “Even-Odd” reflexive game. Reflection is understood as an internal representation of an external environment. To determine if subject uses reflection in the game, three criteria were proposed and applied: formal, neuronal, behavioral. The formal criterion is a win in the game, the neuronal one is a dynamic attractor’s formation of the player’s neural activity, the behavioral one is the deviation of a move sequence from the natural strategy “win-stay, lose-shift”. Three configurations of simple recurrent neural networks were used: 1) the basic one; 2) the one with an additional input which receives information about the success of the last game move, 3) the one which has off-game cycles for processing the received stimulus and making instant decision. We considered the following game conditions: 1) round-robin tournaments of various neural network configurations with modifiable and fixed synapses, 2) playing against fixed sequences of moves. It was found that the simplest neural network model is able to play as good as and, in some cases, better than neural networks of a more complex configuration. The off-game cycles for processing the received stimuli contribute to the non-reflexive behavior of the neural net-work. The performance of neural networks in a reflexive game correlates 1) with the presence of a self-sustaining dynamic attractor and 2) with the use of moves other than the natural strategy.
85. * POLEVAIA A.V., LOSKOT I.V., POLEVAIA S.A., PETUKHOV A.Y., PARIN S.B.
1N.I. Lobachevsky State University of Nizhni Novgorod
2
Experimental diagnostics of the emotional state using affective audiovisual stimuli
The model for assessment of the influence of external audiovisual stimuli on the emotional state of individuals is proposed with the use of heart telemetry and subjective self-assessment tools like SAM and desadaptation level assessment quesstionnaire. The novelty of the work is manifested in the assessment of the influence of stimuli, which takes into account the conflict of their valence and modality. Presented results prove it possible to assess the influence of information on the emotional state of the individual.
86. AFONIN ANDREY NIKOLAEVICH, GLADYSHEV ANDREY ROMANOVICH, GLADYSHEVA ANASTASIA VLADIMIROVNA
Belgorod State Technological University named after V.G. Shukhov
Control system for a bionic hand prosthesis based on an adaptive non-invasive muscle activity analyzer
A control system for a bionic hand prosthesis has been developed, as well as a machine learning model and its software implementation for analyzing signals of muscle activity of various nature from a sensor array. The project is based on a new approach to the implementation of a combined human-machine interface for controlling a bionic upper limb prosthesis based on EMG sensors and a matrix of piezo sensors.
87. A.V. SAMSONOVICH
National Research Nuclear University (MEPhI), Moscow
ON THE POSSIBILITY OF DETERMINING THE PSYCHOLOGICAL CHARACTERISTICS OF THE USER'S PERSONALITY IN HIS ROUTINE INTERACTION WITH AN ARTIFACT
The task of predicting the results of a "Big Five" psychological test based on user behavior data in a routine process of virtual hotel check-in is considered. The participants of the experiment engaged in a dialogue with an imaginary virtual agent, the Registrar. The participants' behavioral data allowed them to predict, with reasonable accuracy, two of the five characteristics of their personality type. The obtained result allows us to propose the concept of "Psychometer" based on a virtual agent or robot capable of discreetly determining the psychological characteristics of clients in the process of performing its main task, without the use of additional means. The method can also be used to assess the behavior of artificial social agents. Translated with www.DeepL.com/Translator (free version)
SESSION 10
Thursday, October 20 10:00 – 11:00
Lecture-hall Концертный зал
Chair: Prof. VLADIMIR DOROKHOV
Neurobiology and neurobionics
88. I.V. MAKUSHEVICH, N.G. BIBIKOVN.N. Andreyev Acoustics Institute, Moscow
Dynamics of background and evoked activity of neurons in the auditory cortex of the unanaesthetized cat
The variability of the impulse activity of neurons in the temporal cortex of an unanesthetized cat was studied in the absence of special sound dampening and in the presence of reverberation. Registration was carried out both in silence and under the influence of various intense sounds coming from the frontal direction. Almost all cells had background activity. Both background and evoked impulse activity were characterized by high bursting and properties inherent in fractal point time processes. The dynamics of impulses over long time intervals of background activity and the nature of variations in the number of impulses per presentation were characterized by the Fano and Hurst methods. A high variability was revealed not only in the background, but also in the evoked activity, which was especially pronounced outside the primary zone A1.
89. O.E. DICK
Pavlov Institute of Physiology of the Russian Academy of Sciences, St.Petersburg
SEARCH FOR MARKERS OF MODERATE COGNITIVE DISORDERS THROUGH PHASE SYNCHRONIZATION BETWEEN RHYTHMIC PHOTOSTIMULUS AND EEG PATTERN
The task is to find differences in the degree of synchronization of the rhythmic photostimulus and brain response in different forms of vascular genesis and the presence or absence of moderate cognitive impairment. To solve the problem, the method of synchrosqueezed wavelet transform of electroencephalogram (EEG) has been used with subsequent analysis of instantaneous frequencies and phases of the EEG pattern and photostimulus and estimation of the phase synchronization index. The method of joint recurrences is also used to analyze phase synchronization. It has been shown that phase synchronization parameters, such as the duration of phase synchronization and the cross-correlation coefficient between the probabilities of photostimulus recurrence and EEG pattern, can serve as markers of moderate cognitive impairment in EEG patterns.
90. MARIE R. KOTIKOVA, ANTON V. CHIZHOV, MICHAEL DRUZIN
1Ioffe Physical Technical Institute, Russian Academy of Sciences, St Petersburg
2
Shunting effect of synaptic channels located on presynaptic terminal
Experiments show that GABAA receptors located on an axon can affect synaptic transmission, presumably through modulation of calcium influx on a presynaptic terminal which determines release of mediator. In this paper, a mathematical mod-el that describes the shunting effect of these receptors on the presynaptic calcium influx is proposed.
91. NATALYA V. LIGUN, VLADIMIR B. DOROKHOV, ARCADY A. PUTILOV, VLADIMIR I. TORSHIN
Institute of Higher Nervous Activity and Neurophysiology of RAS, Moscow
Sleep of poor and good nappers under the afternoon exposure to weak 2-Hz/8-Hz electromagnetic fields
Human sleep might respond to the fields emitted by such natural sources as the earth’s magnetic fields and magnetic activity of the sun. However, the experiments aimed on testing such response are scarce. In two napping experiments, we examined a possibility of beneficial response of sleep charac-teristics in good and poor nappers (n = 14 and 13, respectively) to the exposure to low-level (0.004μT) electromagnetic fields of frequencies 1Hz or 8Hz. A significant increase in amount of slow wave sleep and power density in delta frequency range of the electroencephalographic (EEG) spectrum was found during the napping attempts of only good but not poor nappers. We concluded that the effects of these fields might be beneficial for sleep intensity (e.g., due to prolongation of slow wave sleep), but they might not be additionally effective against the falling asleep problems. We also concluded that these findings should be considered exploratory, and the demonstration of their replicability in larger samples of poor and good sleepers is required.
SESSION 11
Thursday, October 20 11:00 – 12:15
Lecture-hall Концертный зал
Chair: Prof. TEREKHOV SERGE
Applied neural systems
92. MIKHAIL S. TARKOVRzhanov Institute of Semiconductor Physics, Siberian Branch of Russian Academy of Sciences, Novosibirsk
SPICE Model of Analog Content-Addressable Memory Based on 2G FeFET Crossbar
Based on the SPICE model of a double-gate ferroelectric transistor (2G FeFET), a crossbar model was built that can be used as a content-addressable memory (CAM). A new approach is analog computing in memory on matrices (crossbars) of non-volatile memory cells based on 2G FeFET. This architecture eliminates the need for digital-to-analog converters and ensures low power consumption. The CAM model is investigated as an image classifier.
93. SHIBZUKHOV Z.M.
Moscow State Pedagogical University
About One Generalized Neural Network Method of Regression Clustering
An approach to the construction of cluster regression is proposed based on the application of the principle of minimizing differentiable averaging aggregating functions from losses that are insensitive to outliers. An iterative reweighing algorithm is proposed to find optimal values of regression model parameters. The stability of the proposed approach and algorithm to outliers is shown by illustrative examples.
94. ANTON NUZHNY, EVGENIY LEVCHENKO, ALEKSEY GLUKHOV
1The Nuclear Safety Institute of the Russian Academy of Sciences
2
CONSTRUCTION OF NEURAL NETWORK MODEL FOR OPTIMAL CONTROL OF OIL REFINING PROCESSES
The problem of optimal refinery management is considered on the example of a tar hydrocracking plant. A model for economic efficiency prediction of the plant is being built by training a neural network on historical data. Optimization of the predictive economics of the plant by control actions allows us to obtain their values that maximize potential marginality. The correctness of the model recommendations is evaluated by experts, as well as by testing directly on the plant.
95. SOROKIN DMITRY IGOREVICH, BABAEV DMITRY LEONIDOVICH
Learning various locomotion skills from scratch with deep reinforcement learning
Proficiency in locomotion skills will help robots to navigate over challenging terrains. This task is hard to solve programmatically due to the wide variety of terrains and motion patterns. Here we present a framework to learn an agent capable of solving the task of moving with desired linear and angular velocity. The agent learns the task in a curriculum which gradually increases the difficulty of the learned task. We carefully tune the reward function for the agent. The training process is performed in a simulator with domain randomization which forces the agent to learn a robust policy. We tested the proposed framework on the quadruped robot and achieved competitive results.
96. PILGUN M., GABDRAKHMANOVA N.
1Peoples’ Friendship University of Russia, Moscow
2Institute of Linguistics RAS, Moscow
CONVOLUTIONAL NEURAL NETWORKS IN THE TASKS OF SITUATION ASSESSMENT AND FORECASTING ON NUMERICAL DATA
We consider the problem of dynamic description of conflict situation in the society in real time on the basis of the content generated by users and their digital footprints on the example of the Grand Ring Railway Station (South Precinct) project. Integration of neural-network and mathematical models allowed to reveal semantic negative accents, to determine the project positioning in the media-space, the segments of the greatest information attention, social tension around the construction, as well as to forecast the situation development.
SESSION 12
Thursday, October 20 19:00 – 19:45
Lecture-hall Концертный зал
Chair: Prof. DOROFEEV VLADISLAV PETROVICH
Artificial intelligence
97. ARTEM ZHOLUS, YAROSLAV IVCHENKOV, ALEKSANDR I. PANOV1The Moscow Institute of Physics and Technology (State University)
2Federal Research Center "Informatics and Control" RAS, Moscow
Addressing Task Prioritization in Model-based Reinforcement Learning
We propose a data-centric approach to sample efficient model-based reinforcement learning with multitask information. We build a controllable optimization process for MBRL agents that selectively prioritizes the data used by the model-based agent to improve its performance. We show how this can favor implicit task generalization in a custom environment based on MetaWorld with a parametric task variability. We frame the approach within the scope of methods that have followed this direction.
98. KAPELYUSHNIK D.M., BAYMURZINA D.R., KUZNETSOV D.P., BURTSEV M.S.
1National Research University "Higher School of Economics", Moscow
2The Moscow Institute of Physics and Technology (State University)
3DeepPavlov
4London Institute for Mathematical Sciences, United Kingdom
Automatic Generation of Conversational Skills from Dialog Datasets
Modern dialog agents are complex modular systems which, among others, may contain several natural language generation systems named skills. Having enough functionality to be independent chatbots, they often become building blocks for more complex systems able to converse on a range of topics. Such multiskill dialog agents are the closest thing to fully functional open-domain conversational systems. This work proposes an approach to generate conversational skills that, using open-source dialog data and a context of several utterances, can generalize to new domains and generate responses in a controlled and interpretable way. Although it is designed for the open domain, it may be used for automatic skill generation if a more domain-specific dialog skill is required.
99. LUKIA MISTRYUKOVA, IRINA KNYAZEVA, ANDREY PLOTNIKOV, ALEKSANDR KHIZHIK, MIKHAIL HUSHCHYN, AND DENIS DERKACH
1National Research University "Higher School of Economics", Moscow
2Sanct-Petersburg State University
3The Central Astronomical Observatory of the Russian Academy of Sciences at Pulkovo, Saint-Petersburg
Towards Reliable Solar Atmospheric Parameters Neural-Based Inference
Methods of Stokes profile inversion based on spectral polarization analysis represent a powerful tool for obtaining information on magnetic and thermodynamic properties in the solar atmosphere. However, these methods involve solving the radiation transport equation. Over the past decades, several approaches have been developed to provide an analytical solution to the inverse problem, but despite its advantages, in many cases it requires large computing resources. Neural networks have been shown to be a good alternative to these methods, but in general they tend to be overly confident in their predictions. In this paper, the uncertainty estimation of atmospheric parameters prediction is presented. It is shown that deterministic networks containing partially-independent MLP blocks allow one to estimate uncertainty in predictions achiving the high accuracy results.
SESSION 13
Friday, October 21 16:00 – 17:00
Lecture-hall Концертный зал
Chair: Prof. USHAKOV VADIM
Cognitive sciences and brain-computer interfaces. Adaptive behaviour and evolutionary modelling
100. * PROKOFIEV A.A.,IVANOVA D.V.,ZURAVSKA A.1
2Peter the Great St. Petersburg Polytechnic University
“MYO-chat” – a new computer control system for people with disabilities
A new computer control system using electromyographic signals is presented in this work. The system has been designed for the people with disabilities who cannot use the keyboard and/or mouse of computer. The hardware of the system registers the contraction of the human muscle and transmits this data to the computer with the software of the virtual keyboard installed, which performs the "selection action" of the key. The work was inspired by the development of “NeuroChat” and the ideas of the “P300-speller”, but these technologies use an electroencephalographic signal, which in some cases requires more time for processing the signal and special preparing the user. In our case the user only needs one working muscle to use “MYO-chat”. Using “MYO-chat” in computer with Windows operating system, you can fully control the computer without using the mouse. The hardware and software of the current development of “MYO-chat” are described. It is planned to make “MYO-chat” wireless in future versions.
101. STANKEVICH LEV ALEXANDROVICH, GUNDELAKH FILIPP VIKTOROVICH
1Peter the Great St. Petersburg Polytechnic University
2Russian State Scientific Center for Robotics and Technical Cybernetics, Saint-Petersburg
Control of robotic devices based on neuromorphic classifiers of imaginary commands
The work is devoited to problems of development and application of brain-computer interfaces (BCI) in contactless robotic device control systems. BCIs based on EEG signals classifires, arising at imaginary of differend movments are considered. Description and comparison of current the classifires are presented and it is shown that they can provide accuracy 60-80% at recognition for 4 classes of movements. A new type of classifier based on neuromorphic network is proposed which gives accuracy of classification about 90% for 4 classes of imaginary commands. Example of application of the BCI for controlling mobile robot is presented.
102. ZHARIKOV I.N, KRIVOROTOV I., MAXIMOV E., KORVIAKOV V., LETUNOVSKIY A.
The Moscow Institute of Physics and Technology (State University)
A Review of One-Shot Neural Architecture Search methods
Neural network architecture design is a challenging and computational expensive problem. For this reason training a one-shot model becomes very popular way to obtain several architectures or find the best according to different requirements without retraining. In this paper we summarize the existing one-shot NAS methods, highlight base concepts and compare considered methods in terms of accuracy, number of needed for training GPU hours and ranking quality.
103. ZHARIKOV ILIA NIKOLAEVICH
The Moscow Institute of Physics and Technology (State University)
Low-bit Quantization of Transformer for Audio Speech Recognition
The automatic speech recognition is a challenging deep learning problem and transformer architectures have gained an immense improvement in the performance on that task. However, transformer-based models are computationally expensive and comparatively large, which creates issues on deploying them on the memory-constrained devices. Quantization is one of the most promising approaches in reducing the neural network's size and latency. In this paper, we mainly focus on the optimization of the ASR transformer model by applying quantization and knowledge distillation. We apply the SotA quantization methods on the baseline ASR model and examine the sensitive layers which make significant contribution to the performance drop. We've come up with the improvements to accelerate the convergence of quantization methods and to enhance the quantization representation quality. Our modified 2-bit model has shown less than 1% drop in WER in comparsion to the float model on the LibriSpeech dataset.
SESSION 14
Friday, October 21 17:00 – 18:00
Lecture-hall Концертный зал
Chair: Prof. USHAKOV VADIM
Neurobiology and neurobionics
104. SMIRNITSKAYA I.A.Scientific Research Institute for System Analysis, Moscow
The reinforcement learning theory, value function, and the nature of value function calculation by the insular cortex
The Reinforcement Learning theory is a powerful tool for building recognition systems. This theory has long been used in the construction of computational models of neural networks of the brain. However, the validity of its use for these purposes is not unequivocally recognized. One of the reasons for this is the significant differences between the variables used in the theory and the characteristics of the brain that determine behavioral choice. In this paper, the possibility of applying the theory of reinforcement learning for modeling the process of behavioral choice is evaluated. It is argued that insular cortex may be considered displaying state value of RL theory. Such an attempt turns out to be useful, as it allows us to formulate new questions concerning the way control structures interact and the nature of control in the brain, which in turn will allow us to make further progress in understanding the mechanisms of its work
105. STASENKO SERGEY VICTOROVICH, KAZANTSEV VICTOR BORISOVICH
N.I. Lobachevsky State University of Nizhni Novgorod
Astrocytes Enhance Image Representation Encoded in Spiking Neural Network
In this paper we applied a set of metrics estimating the qual- ity of image representation by a spike neural network. The image was encoded in the form of dynamic spiking pattern preserving image’s spa- tial shape in the form of spatio-temporal distribution of spikes. We pre- sented the techniques based on different standard metrics capable to identify the image in the spiking pattern.
106. LYAHOVETSKII, MOROZOV, MUSIENKO, MERKULYEVA
1Pavlov Institute of Physiology of the Russian Academy of Sciences, St.Petersburg
2
3Sanct-Petersburg State University
Nonlinear dynamic model of a system of pacing rhythm generators
The problem of modeling the synchronization of the rhythms of the locomotor central pattern generators that occurs during one-time bidirectional walking caused by electrical epi-dural stimulation is considered. A nonlinear dynamic system of walking rhythm generators based on Van der Pol oscillators is proposed. The simulation results are discussed in the context of knowledge about the neuronal structures of the mammalian spinal cord.
107. MAIOROV VLADIMIR IVANOVICH
Lomonosov Moscow State University
DOPAMINE FUNCTIONS IN REINFORCEMENT LEARNING
Recent experimental data refute the hypothesis of the dopamine signal as a carrier of “reward prediction error” (RPE). The "dopamine drive" is the driving force behind the incentive value that the goal of an action acquires through the Pavlovian association with reward or punishment. The dopamine signal at the same time activates the motor system of behavior and creates the necessary conditions for the induction of plastic rearrangements in synapses (“synaptic tag”), which are “reinforced” if the dopamine drive is reduced.
