CC BY
6Section 9
METHODS OF ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING
Russian-English dataset and comparative analysis of algorithms for cross-language embedding-based entity
alignment
V. A. Gnezdilova1, Z. V. Apanovich1,2
1Novosibirsk State University
2A. P. Ershov Institute of Informatics Systems SB RAS
Email: apanovich_09@mail.ru
DOI 10.24412/cl-35065-2021-1-02-58
The problem of data fusion from data bases and knowledge graphs in different languages is becoming in-
creasingly important. The main step of such a fusion is the identification of equivalent entities in different
knowledge graphs and merging of their descriptions [1]. This problem is known as identity resolution, or entity
alignment problem. Recently, a large group of new methods has emerged to look for so called "embeddings"
of entities and establish the equivalence of entities by comparing their embeddings [2]. This paper presents
experiments with embedding-based entity alignment algorithms on a Russian-English dataset for. Also, the
future directions of research are outlined.
References
1. Zequn., Qingheng Z., Wei H., Chengming W., Muhao C., M., Fartahnaz A., Chengkai L., A Benchmarking Study of
Embedding-based Entity Alignment for Knowledge Graphs, P.-15., 2019. URL: http://www.vldb.org/pvldb/vol13/p2326-
sun.pdf.
2. Apanovich Z.V., Marchuk A.G. Experiments on Russian-English identity resolution. LNCS -9469, 2015, pp. 12-21.
A lobster-inspired multi-robot control strategy for monitoring non-stationary concentration fields
I. V. Bychkov, A. A. Tolstikhin, S. A. Ulyanov
Matrosov Institute for System Dynamics and Control Theory SB RAS
Email: madstayler93@gmail.com
DOI 10.24412/cl-35065-2021-1-02-60
We propose a new lobster-inspired chemotaxis decentralized control strategy for monitoring a non-
stationary concentration field using a team of nonholonomic mobile robots. The task of the team is to locate
and trace the movement of the point (or points) with the highest field value (i.e. source), provided that the
robots are not aware of the dynamics of the field and can only periodically sample the field at their locations.
The proposed strategy combines the lobsters� plume localization behavior and flocking mechanisms to effi-
ciently solve the problem even with a small group of robots. Simulations and experimental works on physical
unicycle robots are performed to validate the effectiveness of the approach for the cases of stationary and
non-stationary fields.
Data-driven turbulence modelling using symbolic regression
A. Chakrabarty1, S. N. Yakovenko1,2
1Novosibirsk State University
2Khristianovich Institute of Theoretical and Applied Mechanics SB RAS
Email: s.yakovenko@mail.ru
DOI 10.24412/cl-35065-2021-1-02-61
The study is focused on the performance of machine learning (ML) methods applied to improve prediction
of main features in canonical turbulent flows by the Reynolds-averaged Navier�Stokes (RANS) equation mod-
els. A key issue here is to approximate the unknown term of the Reynolds stress (RS) tensor arising after Reyn-
olds averaging, which is needed to close the RANS equations.
Turbulent flows in channels with bumps on the bottom having the extensive LES and DNS data sets for
various Reynolds number cases are chosen to examine possibilities of GEP (gene expression programming) [1]
to formulate accurate RANS models. Such a symbolic regression technique allows us to get a new explicit
model for the RS anisotropy tensor. Results obtained by the new model produced using GEP are compared
with those from high-fidelity LES and DNS data (serving as the target benchmark solution during the ML algo-
rithm training) and from the conventional RANS model with the linear Boussinesq hypothesis for the RS tensor.
References
1. Weatheritt J., Sandberg R. D. A novel evolutionary algorithm applied to algebraic modifications of the RANS
stress�strain relationship // J. Comput. Phys. 2016. V. 325. P. 22-37.
Weakly supervised semantic segmentation of tomographic images in the diagnosis of stroke
A. V. Dobshik1, A. A. Tulupov1,2, V. B. Berikov1,3
1Novosibirsk State University
2International Tomography Center SB RAS
3Sobolev Institute of mathematics SB RAS
Email: a.dobshik@alumni.nsu.ru, taa@tomo.nsc.ru, berikov@math.nsc.ru
DOI 10.24412/cl-35065-2021-1-02-95
The early diagnosis of acute stroke is of primary importance for deciding on a method for further treat-
ment. In order to distinguish the brain areas affected by stroke, it is required to engage a highly qualified radi-
ologist. It is known that even a highly qualified specialist can make erroneous predictions with sufficiently large
probability (up to 10 %). Also, the process of manual annotation of a large number of computed tomography
digital images is difficult and time-consuming, thus a specialist may label areas affected by stroke inaccurately.
In this report, a method for automatic semantic segmentation of acute stroke using non-contrast comput-
ed tomography brain images is presented. Under the weakly supervised task we understand the scenario
when some images are labeled accurately and some images are labeled inaccurately. To solve this problem, we
use a convolutional neural network based on U-net architecture [1], since it is known that in the problems of
semantic segmentation of medical images fully convolutional neural networks show a better performance in
comparison with classical machine learning methods [2]. We introduce the model of inaccuracy that shows the
likelihood that the label is correct. The resulting values obtained from the inaccuracy model are used as
weights in the loss function. The proposed method improves the quality of segmentation; its effectiveness has
been tested on real computed tomography images.
This work was supported by the Russian Foundation for Basic Research, project 19-29-01175mk.
References
1. O. Ronneberger, P. Fischer, T. Brox. U-net: Convolutional networks for biomedical image segmentation // Medical
image computing and computer-assisted intervention, Springer, Cham. October 2015. P. 234-241.
2. Nedel'ko V., Kozinets R., Tulupov A., Berikov V. Comparative Analysis of Deep Neural Network and Texture-Based
Classifiers for Recognition of Acute Stroke using Non-Contrast CT Images // 2020 Ural Symposium on Biomedical Engi-
neering, Radioelectronics and Information Technology (USBEREIT). IEEE, 2020. P. 376-379.
