Mikhail Tarkov | Novosibirsk State University (original) (raw)

Papers by Mikhail Tarkov

Lecture Notes in Computer Science, 2016

2013 International Siberian Conference on Control and Communications (SIBCON), 2013

2015 International Siberian Conference on Control and Communications (SIBCON), 2015

2013 International Siberian Conference on Control and Communications, Sep 1, 2013

Analytical and Quantitative Cytology and Histology the International Academy of Cytology and American Society of Cytology, May 1, 2007

To analyze smears of 197 thyroid follicular tumors (adenoma and carcinoma). Several types of arti... more To analyze smears of 197 thyroid follicular tumors (adenoma and carcinoma). Several types of artificial neural networks (ANN) of various designs were used for diagnosis of thyroid follicular tumors. The typical complex of cytologic features, some nuclear morphometric parameters (area, perimeter, shape factor) and density features of chromatin texture (mean value and SD of gray levels) were defined for each tumor. The ANN was trained by means of cytologic features characteristic for a thyroid follicular adenoma and a follicular carcinoma. At subsequent testing, the correct cytologic diagnosis was established in 93% (25 of 27) of cases. The morphometry increased the accuracy of diagnosis for follicular tumors in up to 97% (75 of 78) of cases. ANN correctly distinguished an adenoma or a carcinoma in 87% (73 of 84) of cases when using color microscopic images of tumors. The usage of ANN has raised sensitivity of cytologic diagnosis of follicular tumors to 90%, compared with a usual cytologic method (sensitivity of 56%). The automatic classification of thyroid follicular tumors by means of ANN is prospective.

An effective algorithm is proposed for balancing computation load in a heterogeneous multicompute... more An effective algorithm is proposed for balancing computation load in a heterogeneous multicomputer system by the self-organizing neural Kohonen network. The algorithm takes into consideration different performances of the system processors.

We propose algorithms for mapping a weight matrix of a neural layer onto distributed computer sys... more We propose algorithms for mapping a weight matrix of a neural layer onto distributed computer systems with a torus structure designed for parallel image processing. We show that the choice of the mapping technique depends upon the ratio between the number of neurons and the number of weight coefficients in a neuron (the number of pixels): if this ratio is sufficiently small, the column distribution is more efficient, otherwise the row distribution is more suitable. In particular, for the Hopfield network, the row distribution is always more efficient than the column distribution.

Optical Memory and Neural Networks

The problem of mapping a parallel program with weighted vertices (processes) and edges (interproc... more The problem of mapping a parallel program with weighted vertices (processes) and edges (interprocess exchanges) onto a weighted graph of the distributed computer system is considered. An algorithm for solving this problem based on the use of Hopfield networks is proposed. The algorithm is tested on mapping a number of graphs of parallel programs onto multicore computer. Experiments have shown that the proposed algorithm provides a well-balanced sub-optimal mappings.

Analytical and quantitative cytology and histology / the International Academy of Cytology [and] American Society of Cytology, 2007

Analysis of a number of dimensionality reduction techniques is carried out. It is shown that morp... more Analysis of a number of dimensionality reduction techniques is carried out. It is shown that morphological image analysis in conjunction with diffusion maps in a problem of diagnosing thyroid disease by images of cytological preparations gives us possibility to have recognition efficiency higher than early proposed method based on Fourier spectrum correction and distinguishing principal com ponents.

The problems of programming memristor array (crossbar) are considered. A voltage pulse duration t... more The problems of programming memristor array (crossbar) are considered. A voltage pulse duration to set the desired value of memristor resistance is evaluated. An algorithm for mapping weight matrix of the neurons layer onto memristor crossbar is proposed. A simulation of adaptive adder with memristor synapses is carried out in the LTspice using the proposed method of mapping. The possibility of using the adaptive adder for image recognition is demonstrated.

An algorith m is proposed for tracking objects in real time. The algorithm is based on neural net... more An algorith m is proposed for tracking objects in real time. The algorithm is based on neural network implemented on GPU. Investigation and parameter optimization of the algorithm are realized. Tracking process has accelerated by 10 times and the training process has accelerated by 2 times versus to the sequential algorithm version. The maximu m resolution of the frame for real-time tracking and the optimu m frame sampling from a movie are calculated.

Numerical Analysis and Applications, 2010

An algorithm based on a recurrent neural Wang’s network and the WTA (“Winner takes all”) principl... more An algorithm based on a recurrent neural Wang’s network and the WTA (“Winner takes all”) principle is applied to the construction of Hamiltonian cycles in graphs of distributed computer systems (CSs). The algorithm is used for: 1) regular graphs (2D- and 3D-tori, and hypercubes) of distributed CSs and 2) 2D-tori disturbed by removing an arbitrary edge. The neural network parameters for the construction of Hamiltonian cycles and suboptimal cycles with a length close to that of Hamiltonian ones are determined. Our experiments show that the iterative method (Jacobi, Gauss-Seidel, or SOR) used for solving the system of differential equations describing a neural network strongly affects the process of cycle construction and depends on the number of torus nodes.

ABSTRACT In this paper, an algorithm for mapping semigroup array operations onto distributed comp... more ABSTRACT In this paper, an algorithm for mapping semigroup array operations onto distributed computer systems with toroidal topology is analyzed. This algorithm is based on the well-known butterfly scheme and mapping this scheme onto hypercube with subsequent XOR -embedding of the hypercube onto torus. We show that the hypercube-onto-torus mapping algorithm provides the time of the semigroup operation implementation on torus less than the time provided by an algorithm using a sequence of cyclic data shifts.

Mapping one-measured (“line”, “ring”) and two-measured (“mesh”, “torus”) parallel program structu... more Mapping one-measured (“line”, “ring”) and two-measured (“mesh”, “torus”) parallel program structures onto a distributed computer system (DCS) regular structures (“torus”, “two-measured circulant”, “hypercube”) with faulty elements (processor nodes and links) is investigated. It is shown that: 1) onemeasured program structures mapped better than two-measured structures; 2) when failures are injected to the DCS structure the one-measured structures mapping aggravated very lesser than the two-measured structures mapping. The smaller a node degree of a program graph and the greater a node degree of a DCS graph the better the mapping quality. Thus, the one-measured program structure’s mappings are more fault-tolerant than two-measured structure’s one and more preferable for organization of computations in the DCS with faulty elements.

ABSTRACT An approach to a quantitative evaluation of heterogeneity ("motley") o... more ABSTRACT An approach to a quantitative evaluation of heterogeneity ("motley") of thermal images (thermograms) bated on the wavelet Haar transform is proposed and investigated. It is shown that the Haar transform allows getting the evaluation to be adequate to visual perception of the thermogram heterogeneity.

Parallel Computing, 2002

Algorithm of mapping adaptive fuzzy Kohonen clustering network (AFKCN) onto an image processing s... more Algorithm of mapping adaptive fuzzy Kohonen clustering network (AFKCN) onto an image processing system with distributed memory and torus topology is presented. A hypercube is used as a structure of the AFKCN parallel program for image segmentation. The impossibility of message congestions on torus links is proved for the hypercube-to-torus XOR-embedding. Expressions for the parallel AFKCN performance analysis are given. For distributed image processing systems having two- and three-measured torus topologies the analysis shows good estimation of the parallel AFKCN implementation.