Research

ISTC RESEARCH PROJECTS

The FPGA – Accelerated Time-Series (key-value) Database is a dedicated hardware for IoT (Internet of Things) data storage. Smart sensors’ data passes directly into FPGA from Ethernet adapter bypassing any type of operating systems and software layers. All data processing takes place under FPGA on the fly and streams into SSD for storage. Artificial Intelligence algorithms can be used to continuously evaluate interconnections between channels, classify and predict data. The device provides the general functionality of software-enabled time-series (key-value) database systems while implemented on the hardware layer. FPGA’s native parallel feature is heavily used to achieve data throughputs greater then all available software enabled time series database systems can offer while keeping server power consumption 3-4 times lower compared to CPU based solutions.

Team Members

  • Artavazd Khachatryan/YSU
  • Khachik Sahakyan/YSU

The main purpose of this project is to apply the latest developments in machine learning to medical databases. Although large databases of medical time-series data are freely available, they are largely ignored by the machine learning community. The main reason is that lots of medical expertise are required to understand and process these databases. During this project, we will define several standardized benchmarks on MIMIC-III clinical database1 that are known to be useful for doctors (e.g. in-hospital mortality prediction, phenotype classification, length of stay prediction etc.), will apply various state-of-the-art machine learning models on these benchmarks and will develop new models.

Team Members

The brain is the integrity of billions of computational units from the point of view of information processing. The communication between these units is realized via electrical signals and each unit shows the ability of experience integration, i.e. the ability of learning. The study of the mechanisms of information storage, encoding and conducting in this system is of great importance for application in brain-machine interfaces and neuroprosthetics. The recording of the electrical activity of a brain is an irreplaceable method for studying the behavior of neural populations and mathematical or physical patterns underlying this behavior. Recording of this signal makes it possible to detect the behavior of neural populations with high temporal and spatial resolution. An additional reason of interest in LFPs is that they provide a stable signal for a longer period and therefore, are useful for long‐term chronic experiments and for clinical applications such as Brain-Machine Interfaces (BMI). However, as LFP captures a multitude of neural processes, the extraction of essential features is among the main problems of Computational neuroscience. The investigation of neural activity underlying cognitive functions requires recording of activity in several brain regions and analytical methods that will give an opportunity to combine multiple features of spatially and temporally separated complex signals to achieve a robust classification of brain states. The extraction of cognitive functions requires complex processing of information. We will continuously execute these steps in a cycle and at each iteration, we are going to improve further the placement methods and positioning of the electrodes, the LFP signal conditioning circuit and signal analyzing algorithms. So, one of the outcomes of the project will be the developed complete methodology for LFP recording and analysis. Another expected outcome of this project is an analysis tool for LFP data which can also be used in the animal models of neurological and neurodegenerative disorders. As a result of our project, we will have a large amount of data on LFP recordings and video recordings, both in raw and processed versions. We are going to make this data available on the Internet. This community-driven data sharing may offer cross-validation of findings and refinement of interpretations of the data. After the achievement of solid results with LFP recording, we are planning to couple it with simultaneous recording of EEG in the same animal. This can lead to translation of our methods to human EEG data analysis, which is relevant as a non-invasive method.

Team Members

  • Serob Muradyan/RAU
  • Lusine Hovhannisyan/YSMU
  • Amalya Hakobyan/YSMU
  • Ani Alaverdyan/YSMU
DRAA project excels at managing complex and changing information, which is important for automating many kinds of analysis and policies. It provides new methods for rapidly authoring structured logical information (i.e., encoded knowledge) starting from user-supplied English documents, and for powerfully reasoning with that knowledge to answer queries and present the conclusions in English. In other words, users assert and query knowledge in DRAA. Automated decisions can then be made based on the results of the analysis. For example, compliance alerts can be generated automatically based on reasoning about policies and regulations that have been encoded into a DRAA knowledge base. The meaning of an English sentence is captured deeply and precisely. Unlike existing systems, almost anything one can say in English can be encoded and operationalized, relatively easily, largely by the subject matter experts themselves. In this connection, DRAA can perform truly deep reasoning based on the encoded knowledge statements. This makes the automation of many aspects of analysis and decision making become much more fordable and practical.

Team Members

  • Kristina Sargsyan/MSUIECS
  • Arman Grigoryan/ERIICTA
  • Eduard Khachatryan/ERIICTA

Electronic Health Information Management System (e-HIMS) is an essential element of healthcare improvement and a vital component of strategies for healthcare system reform. A health information system that provides reliable, timely, high-quality information is a key prerequisite for healthcare system improvement in nowadays. Since 2010 Armenian government is working on development and implementation of integrational Integrated Health Information System of Armenia (IHISA). Currently, the government is highlighting the need for e-HIMS implementation in the healthcare facilities. This fact enforcing and expanding the market for e-HIMS in Armenia. Having this vision and perspectives we developed Electronic Health Information Management System (e-HIMS) for the healthcare facilities which is an intelligent instrument that supports health care services to compile and use health information for better clinical and financial performance. It has multiple modules that allow computerizing clinical workflow throughout the healthcare facility and adapts the system to reflect the way healthcare facility works. Moreover, one of the key components of our e-HIMS product is electronic quality control management system. The Quality of healthcare is a comparatively new concept in healthcare and it is defined as the level of value delivered by any healthcare resources. According to the US Institute of Medicine, it is a degree to which health services for people to increase the probability of targeted health outcomes and are consistent with existing medical knowledge. In other words, quality of health care services could be defined as doing the right thing, at the right time, in the right way, for the right person – and having the best possible results. Effective and efficient quality systems can promote good practice. The studies and scientific publications are showing that is very difficult to implement described e-Health projects. For doctors, patients, and the healthcare managers the implications of Electronic Health Information Management Systems are usually hard to understand, as are the likely pace and extent of adoption of e-HIMS.

Team Members

  • Karapet Davtyan/AUA
  • Artashes Tadevosyan/YSMU
  • Hayk Davtyan/AUA

The Project aims to make a deep analysis of three business processes in telecommunication sector companies of Armenia and to define efficiency growth measures via application of IBM® products with sufficient automation and control. The following business processes shall be considered: Training & Development, Compensation & Benefits (Reward Management), and Performance Management. The projects aim to link science with practice via connecting theoretic research with practical applications and output analyses. Purpose 1: Analyze telecommunication sector development trends and BPM challenges from customer excellence and shareholder expectations perspective. Purpose 2: Integrate IBM® products into the Training & Development, Reward Management, and Performance Management processes. Purpose 4: Test and prove the efficiency of IBM® products application in telecommunication sector from a BPM perspective.

Team Members

  • Marine Ghahramanyan/UJML
  • Lilit Poghosyan/NUACA
  • Irina Poghossian/AUA

Research-based decision making is a rare procedure for Armenian regional development. Technology-based research and innovative solutions also are far from the development strategy. In this context, the proposed interdisciplinary research (including statistics, economics, regional development and public administration) cloud integrate academic approach, business needs, and community development through new technological solutions through development of computer-human interaction and knowledge-rich intervention. Economics, statistics, and public administration are core components of the research that will be developed through SPSS modeling, GIS spatial management and Cloud infrastructure management tools. The objectives of the project are regional economic modeling and spatial development strategy based on inter-municipal cooperation and social entrepreneurship for the target region, as well as knowledge transfer and education. The expected results of the project will be (1) a developed course for higher education, (2) a five-person research group, (3) a manual, (4) economic models and relevant academic papers, (5) a spatial development strategy and relevant academic paper.

Team Members

  • Sos Khachikyan/ASUE
  • Arthur Dolmajian/École Polytechnique de Montréal
  • Armen Ktoyan/ASUE
  • Davit Shahnazaryan/YSU

Intrinsically disordered proteins (IDPs) are associated with various diseases and have been proposed as
promising drug targets. However, conventional structure-based approaches cannot be applied directly to
IDPs, due to their lack of ordered structures. Nowadays the processes of designing new drugs by using
bioinformatics tools have opened a new area of drug research and development. Computational
techniques assist in searching drug target and in designing a drug in silico. As many protein-protein and
ligand-protein interactions can be used as key targets for drug design, one of the solutions is to design
protein drugs based directly on the protein complexes or the target structure. Recently the several
strategies for structure-based protein drug design have been developed.
The project is dedicated to the development of the neural networked algorithms and software for the intrinsically disordered proteins under the different conditions of the environment. The project will be focused mainly on the coarse-grained models of the intrinsically disordered proteins and their improvement with neural – network based algorithms. The planned results will be demanded in bioinformatics and related areas.

Team Members

  • Yevgeniy Mamasakhlisov/ YSU
  • Shushanik Tonoyan/ YSU
  • Vardges Mambreyan/ YSU

Due to its importance in many application areas such as video surveillance, content-driven retrieval, robotics, and human-computer interaction research, human action recognition in videos is an important task in computer vision research. Despite the rapid development of the research in this field, the human action recognition still remains a very challenging task, due to the complexness of capturing information from still frames and motion between frames on the one hand, and the combination of these two streams in a way that optimizes class prediction, on the other. The research proposal implies a two-fold contribution. Firstly, the team is working on the promotion in Armenia the computer vision research in general and the video classification research in particular. Secondly, the team determines development to a condition of contributing to the international research on this topic.

Team Members

  • Hayk Sargsyan/ University of Zurich
  • Sona Hunanyan/ EPFL Lausanne, Switzerland
  • Arsen Yeghiazaryan/ Yerevan Physics Institute

There are many existing standards in the financial sector, yet the documents available in the industry and in its different areas are usually inconsistent. Therefore, the available data is unstructured and cannot easily be analyzed. The vision of this project is to develop a product that will ensure the uninterrupted realization of the whole process of matching, extraction, transformation, analysis, and reporting of financial data.

Team Members

  • Hrant Davtyan/ AUA
  • Nare Banduryan/ YSU

An original logistics optimization algorithm is being developed. It considers the whole vehicles & orders the team have shown results in qualitative optimization for Chicago taxi historical data (up to 46%). A further validation of real-time data is required. The proposed research project addresses that gap.

Team Members

  • Hayk Ishkhanyan/ Moscow Institute of Physics and Technology
  • Sergey Kovalenko/ Moscow Institute of Physics and Technology

The purpose of this research project is to design and develop a bionic hand which can be controlled by simultaneous EEG data analysis using Python machine learning algorithms.

Team Members

  • Lusine Hovhannisyan/ YSMU
  • Marina Davtyan/ YSMU
  • Sose Berberyan/ AUA

The main purpose of this project is to build high quality and open source part-of-speech taggers, morphological analyzers and sentence parsers for multiple languages. Part-of-speech tagging and dependency parsing of sentences are the two fundamental building blocks of NLP systems. Morphological analyzers are more important for morphologically-rich languages (including Russian and Armenian). The task is motivated by the success of the CoNLL 2017 Shared Task which was about creating POS taggers and parsers for 60 languages. This project follows two goals. First is the participation in the CoNLL 2018 Shared Task, which is expected to start in April 2018 and last for six months. It will be the continuation of the CoNLL 2017 Shared Task and will bring researchers from all over the world to build the next generation of POS taggers, parsers, lemmatizers and morphological analyzers based on UD corpora. The second goal is to continue building Eastern Armenian UD Treebank and make it available for CoNLL 2018 Shared Task.

Team Members

  • Hrant Khachatryan/YerevaNN
  • Marat Yavrumyan/ University of Bayreuth, Germany;

One of the most important areas of genomic medicine is identifying and describing the genetic markers, which are associated with the predisposition or involved in the development of complex diseases, such as cancers, cardiovascular diseases, neurodegeneration, etc. The important hallmark of the complex disease is that there is no single gene responsible for it development, but rather set of dozen genes, as well as gene-environment interaction, are implicated in disease development. There are hundreds of diseases, which are classified as complex, polygenic diseases, moreover, multiple studies have already determined hundred or thousand genetic markers i.e. single
nucleotide polymorphisms (SNPs), associated with them, and their identification and development of genome-based diagnostics and prognostic tools are important questions of modern personalized and precision medicine. Genome-wide association study (GWAS) is one of the most effective methods, which allows determining SNPs, across the whole genome, which is associated with a predisposition to complex diseases.

By nowadays, researchers all over the world have done multiple GWA studies and have accumulated huge data of SNPs-diseases association (4,000,000 markers in average per single subject, several project initiatives have databases with thousand individuals
studied). These data are used in population genetics to determine ancestry based on population-specific genetic markers and provide for individuals their percentage map of genetic belongingness to distinct populations. Similarly, this information can be (and is
being) used to study the genetic association with diseases usually by comparing diseased and healthy subjects.
Nevertheless, it is very important to take into consideration that disease-associated SNPs are also very distinguishing across different populations, so SNP, which is associated with particular disease in one population, could be neutral or even protective in another.

Team Members

  • Maria Nikoghosyan/ RAU
  • Ani Sakhlyan/ AUA

The main goal of this project is investigating the possibility of applying non-orthogonal multiple access technologies with the compound modulation scheme for future 5G networks to improve the overall spectral and spatial efficiency. Develop calculation method of channel budget and BER estimation for NOMA with combined SM and amplitude phase modulation schemes. Carry out numerical calculations and realize a simulation model of NOMA for 5G networks. Implementation and testing of the proposed scheme in existing standards (for instance for WI-FI with MIMO).

Team Members

  • Hovhannes Haroyan/ Radiophysics Research Institute
  • Gor Hovsepyan/ The Institute for Physical Researches Ashtarak, Armenia
  • Sargis Sargsyan/ Institute of Radiophysics and Microelectronics

The goal of this project is to develop a system for closed-loop deep brain stimulation in vascular dementia animal model using machine-learning packages. We are going to design and implant microelectrode arrays within subcortical brain regions, which are the most
vulnerable areas seen in VD. The array concept consists of three logical parts: the sensing part will give us the ability to record the electrical activity of neural cells in a relatively large spatiotemporal pattern [5]. Stimulation will be made with the same microelectrode array. Recovery time will be calculated to remove noise after stimulation. As for the third part, drug delivery microcannula will be attached to this system for injection of a neuroprotective compound as a stimulation positive control. The custom-made program will
be developed in the LabVIEW environment to analyze gained data. Recordings will be done from both anesthetized and behaving animals to validate “proof of concept”. Behavioral patterns will be correlated with gained data to implement effective machine learning algorithms for this system. Two-photon microscopy station of Candle SRI will be used for Ca2+ imaging of selected sites. Image processing algorithms will be used to detect the spikes. This will be a two-fold validation of electrophysiological studies.

Team Members

  • Senik Matinyan/ Candle Scientific Research Institute
  • Liana Hayrapetyan/ YSMU
  • Lusine Hovhannisyan/ YSMU

Biomolecular signaling cascades or pathways are represented as directed graphs with nodes corresponding to biomolecules and edges to physical interactions between them. These graphs have input nodes corresponding to signal receptors and output or sink nodes corresponding to final targets.
The latters usually represent realizations of biological processes (e.g. cell death, cell division, or migration). The set of all possible interactions within a cell forms one global pathway, which theoretically can reach a size of up to 109 interactions. However, for comprehension and analytical purposes, this global pathway is artificially divided into functional units or sub-pathways, which are
stored separately in pathway databases (KEGG, Reactome, Wikipathways, SMPDB, etc.). These databases have two major shortcomings: they do not account for interactions between sub-pathways (sharing of nodes, branches) and they do not cover all the information about the global pathway. On the other hand, databases storing protein-protein interaction (PPI) data are relatively rich and cover the vast
amount of accumulated knowledge about PP interactions. However, they also have disadvantages: first, they do not contain nodes of type “biological process” (BP) and, thus, do not link the networks of interactions to cellular responses, and second, they do not store directionality, which is important for recovery of signal flow from inputs to outputs. Owing to the mentioned limitations, currently available databases are not suitable for studies aimed at drug target identification or repositioning.
Overall, the PPI and pathway databases each store separate pieces of the puzzle. Our project’s objective is to combine these pieces together to create a model of the global cellular pathway. We, thus, aim at generating a distributed database that stores a graph representing the global pathway, which keeps not only PPI information, but also protein-BP links, and has directionality. We will develop a front-end tool that will allow for retrieving any arbitrarily defined path or sub-pathway between user-defined input and
output nodes, at the same time accounting for custom constraints, such as the tissue type, gene expression signature, etc. An example use case will assume a user to supply a drug or immediate drug targets, and our platform to retrieve all the biological processes potentially affected by the drug, and displaying the pathways linking the drug to predicted BPs in an interactive GUI. The user may thus
predict drug side effects, or identify potential targets for drug repositioning. Later, we plan to incorporate additional modules, allowing the user to run various types of analyses on retrieved pathways.

Team Members

  • Lilit Nersisyan/ Institute of Molecular Biology NAS RA
  • Vahe Momjyan/ AUA
  • Arpine Manukyan/ AUA
  • Shushan Astanyan/ YSU

A brain-computer interface (BCI) is a hardware and software communications system that permits cerebral activity alone to control computers or external devices. The primary goal of BCI research is to give disabled people with different diseases such as amyotrophic lateral sclerosis (ALS), brainstem stroke, spinal cord injury etc. the ability to communicate using only cerebral activity. One of the reasons why BCI research is gaining popularity is ever growing number of above-mentioned diseases, but also more accessible and smaller technology. As an example we can mention, that the processors in nowadays smartphones can perform machine learning
tasks in an acceptable time, which wasn’t possible until recently. To get proper results standard BCI usually has to be able to perform different tasks at once such as signal acquisition, preprocessing or signal enhancement, feature extraction, and classification. Different BCI uses different methods for signal acquisition such as electroencephalography (EEG), electrocorticography (ECoG), magnetoencephalography (MEG), functional magnetic resonance (fMRI) and near-infrared spectroscopy. But as you can imagine EEG is the widest spread and most accessible method for brain activity measurement, thanks to its non-invasive application, relatively small size and easy to use interface. The EEG recording system consists of electrodes, amplifiers, A/D converter, and a recording device. The electrodes acquire the signal from the scalp, the amplifiers process the analog signal to enlarge the amplitude of the EEG signals so that the A/D converter can digitalize the signal in a more accurate way. Finally, the recording device, which may be a personal computer or similar, stores, and displays the data. We are also planning to research ways of making EEG devices more accessible(cheaper and
portable) to the public, which plays a key role to spread BCI.

However, this is something for the future, for this research we will use commercial EEG devices.

Team Members

  • Davit Ghukasyan/ RAU
  • Tigran Ghukasyan/ RAU

The team plans to develop methods for parameter learning in non-identifiable models, where the number of parameters is so large that their precise estimation is impossible even with infinitely large data (i.e. for a large number of samples). A straightforward example of a non-identifiable model is when the data (prior to observation) passed through a noise with unknown magnitude, e.g. Hidden Markov
Models. The team wants to develop parameter learning methods that generalize the maximum likelihood, provide an automatic Occam’s razor and deal with the overconfidence problem.

Team Members

  • Armen Allahverdyan/ Yerevan Physics Institute
  • Narek Martirosyan/ Yerevan Physics Institute

Nowadays transportation and health-care companies face real optimization problems. They need to optimize asset usage (e.g. trucks, airlines), scheduling (e.g. surgeon, staff planning) etc. Many real-world optimization solutions use algorithms and techniques developed by Operation Researches discipline. These include Large-scale Neighborhood Search algorithm, Discrete Event Simulation, Mixed Integer Programming etc. In real-world solutions, these algorithms and techniques become a performance bottleneck. Also, Operation Research algorithms and techniques do not consider customizations and preferences (e.g. driver’s favorite routes).
The objective of this research is to solve real-world optimization problems with Machine Learning discipline.

Team Members

  • Gevorg Martirosyan/ RAU
  • Ashkhen Hovhannisyan/ YSU

Credit scoring remains one of the main tools used during the loan provision decision-making process. Yet, the traditional system used by local financial institutions is still anchored on the opinion of loan specialists or in some rare cases on a basic scoring system. This study aims to leverage machine learning techniques to automate evidence-based decision-making and allow for adaptive learning from new observations. At the same time, the econometric approach will be applied to compensate for the uninterpretability of artificial neural networks.

Team Members

  • Vardan Baghdasaryan/ AUA
  • Aleksandr Grigoryan/ AUA
  • Hrant Davtyan/ AUA
  • Knar Khachatryan/ AUA