Doctoral studies and the Univ
Hello! We are happy to welcome you to the new website of the University of Tartu. As it is still fresh new, problems may occur. We will fix them as soon as possible. Please use the feedback form to give feedback and let us know of any issues. Thank you!
First application period for a limited number of specialities will be from 1 to 15 February 2022.
Vacancies for February 2022 intake will be announced by 15 January.
Supervisor: Huber Flores
Abstract: Increased processing capabilities of smart, wearables and IoT devices, along with the emergence of services at the Edge of the network are enabling the execution of machine and deep learning in constrained environments. Indeed, distributed computation between devices can cope with the resource intensive requirements of AI-based models. More importantly, by moving the execution of AI models to the Edge, communication latency and energy consumption of the endpoint devices that rely on AI support is improved. Unfortunately, the execution of AI-based models is a black box -- even when distributed. Thus, tools are required to explain the behaviour of these models when training and executing in a distributed manner. In this project, we aim to develop an accountability toolkit that can be used to comprehensively illustrate the operation and decision making of AI during runtime. To achieve this, we are interested on analysing general root causes in a distributed computing infrastructure that can change the behaviour of AI models.
Supervisor(s): Angela Ivask, Vambola Kisand
Abstract: During this project the doctoral candidate will select and synthesize a variety of antimicrobial materials that are suitable for attachment to different types of surfaces, such as hard surfaces and textiles. Selected materials are expected to include materials of various shapes which, when attached to a surface, inhibit microbial adhesion and microbial membrane attack materials due to their structural geometry. Selected materials and surfaces, including hard surfaces and textiles, will be tested for their antimicrobial properties and durability under simulated real-life usage conditions.
Doctoral students admitted to these projects will have employment contracts with the junior researcher. The expected workload is 1.0, the expected work period is four years. The final workload will be set during the negotiations.
Supervisor: Meelis Kull
Abstract: The success of machine learning with deep neural networks has so far mostly been relying on huge training data sets. These data need to be gathered carefully from exactly the same setting where the networks are to be deployed to. Any differences between the training and deployment contexts typically results with over-confidence and poor predictive performance, known as the out-of-distribution (OoD) problem. The OoD problem is one reason why the perception and planning systems of autonomous cars are currently quite easily confused by novel situations.
The goal of the PhD project is to alleviate the OoD problem by developing methods to learn artificial neural networks wide hierarchical disentangled representations. Wideness supports better detection of OoD situations, disentanglement supports usage of only the relevant parts of the model in OoD situations, and the hierarchy supports paying attention to both the tiny details as well as the overall composition.
Supervisor: Raivo Kolde
Abstract: The amount of real-world medical data is growing every day and creating novel opportunities for obtaining clinical evidence. To facilitate such research, more and more of the electronic health records and insurance claims data sources are converting their data into a common research format called OMOP. The common data model enables large federated international clinical studies. However, running such projects requires novel methodology and software solutions.
One example of a poorly supported research question is the economic evaluation of health technologies that relies on simulating treatment outcomes using Markov models. To achieve realistic results, the model has to be trained on local data and also take advantage of knowledge generated elsewhere. The federated analysis approach across clinical sites has a lot of promise for this application, because it allows sharing model specifications and parameters and comparing results between countries and locales.
The goal for the PhD project is to create tools and methodologies to learn the models for economic analyses on OMOP formatted datasets. The OMOP platform has already a exciting software ecosystem that can be leveraged and the federated nature of the analysis raises interesting methodological challenges and opportunities. The method development will be complemented with applying the tools in actual studies together with clinical collaborators.
Supervisor: Arnis Paršovs
Abstract: Technology solutions are increasingly moving towards the wireless domain with technologies such as Bluetooth, Wi-Fi, NFC and remote keyless systems taking over conventional wired technologies. While the consumer’s reliance on these wireless technologies is significant, the security guarantees provided by the consumer products implementing these technologies have not been fully researched. The wireless solutions can be susceptible to various types of attacks such as: replay attacks, impersonation attacks, denial of service attacks, information disclosure attacks and others. The aim of this thesis is to research security guarantees and assumptions provided by these everyday technological solutions to understand risks and contribute to their security.
Supervisor: Huber Flores
Abstract: Environmental monitoring is a key requirement for solving many grand challenges of our time as it offers information about the condition of the environment and feedback on how human activity is affecting it. Current methods for collecting environmental data, e.g., on air pollution, everyday trash, or ocean litter, are costly and suffer from limited spatiotemporal granularity. This restricts our understanding of the overall environmental impacts of human activity and how to mitigate their effects. The CityDrone project builds on the emergence of drone-based systems, including aerial, ground, and underwater-based drones, and low-cost sensors for environmental sensing to establish a foundation for massive-scale environmental monitoring using multi-drone systems. The idea is to have multiple drones, potentially even of different modalities, to work in coordination to acquire information about the state of the environment and the activities that impact it. Drones have the potential to significantly increase both the spatial and the temporal coverage of information and help to scale up environmental monitoring and offer information at finer resolution than what is currently possible. Realizing this vision requires solving foundational research challenges that currently limit the use of drones for environmental monitoring. These include challenges in establishing coordination among the drones collecting information, limitations in the quality of environmental data that drones can capture, and energy restrictions of drone-based systems. The project addresses these challenges contributing novel algorithms, systems, and experimental benchmarks that pave the way toward environmental monitoring at massive-scale.
Supervisors: Mark Fishel, Karl Kruusamäe
Abstract: Social robots and in particular physical conversational robots can significantly improve the quality of personalized healthcare. However, building them requires standalone components that are integrated as independent steps in a pipeline. Moreover, the solutions are typically language-dependent. The goal of this thesis is to explore the synergy between the conversational part of the task (i.e. speech and language processing) as well as the robotics part of it (vision, actions) and to experiment with knowledge transfer between different tasks as well as end-to-end solutions to task combinations. The practical aim is to develop a flexible platform for physical conversational robot creation with ease of adaptability and practical usefulness in mind. The main prerequisites are some experience with contemporary approaches to robotics and natural language processing tasks. Knowledge of Estonian gives a strong preference.
Supervisors: Piret Luik, Tauno Palts
Abstract: As the first choices for further studies are made after graduating from basic school, it is important to introduce the basics of programming for the students already at the lower-secondary school (age between 13-16). Thus the purpose of this study is to develop programming materials for lower-secondary school students to support development of computational thinking (CT) and enhance motivation to study ICT. First, based on the analysis of previous research and practice of several countries, different ways to learn programming and its materials will be analyzed. Secondly, a model for teaching programming will be created. Based on this model, materials to learn programming are developed, using the ADDIE design model. The materials will be in accordance with the national computer science curriculum. A survey for students and teachers will be conducted to evaluate the usage of the materials and to measure motivation to study ICT and decision to continue studies at the secondary school’s science or ICT class. Also, developed materials will be analyzed in the context of creating a digital project.
Supervisor: Kairit Sirts
Abstract: The accuracy of the natural language processing (NLP) systems is heavily dependent on the amount of annotated data. For many NLP tasks and languages, the annotated training sets are not large enough to train reliable models. While the data annotation is an ongoing process, there already exist several data resources, typically in the form of lexicons, ontologies and rule-based systems that are not exploited in modern deep learning models. This project will study methods for integrating external linguistic resources into neural networks in order to improve the accuracy of selected natural language processing tasks. Enhancing neural models with this external information can significantly improve the accuracy of predictions especially in the limited data settings. In addition to improving performance, the project aims to outline the relevant factors that influence the effect of the external information integration into neural models.
Supervisors: Eduard Ševtšenko, Ibrahim Oluwole Raji (Middlesex university)
Abstract: Due to globalisation, rapidly changing business environment, highly influenced consumers, fast-changing needs, and consumers’ behaviours, there is the demand to make transactions and interactions faster, more reliable, customer and consumer-friendly. Before we can satisfy all listed needs, we need to standardise some processes. Primarily there is heavy demand to standardise “consumer faced” processes. Standardisation helps also reduce the cost of service.
The Scope of current research work is to develop a framework for consumer-faced process standardisation, which is an essential part of the Digital Sustainable Partner Network.
The research’s main idea is to apply customer satisfaction level Key Performance Indicators (KPIs) for supply chain reliability improvement. The Supply Chain Operations Reference (SCOR) model-based KPI metrics enable increasing the quality of product/service by monitoring, and further digitalisation of directly involved processes. In the long run, the solution will ultimately reduce/eliminate the number of customer reclamations in the supply chain.
Supply Chain Digitalisation (SCD) framework analyses the companies’ current reliability level based on reclamations and processes quality data provided and points out the improvement needed in business processes. The business process KPI-s are developed based on the SCOR model, and the efficiency of digitalised business processes is predicted by BBN tool. The next step is to apply the SCD framework to improve the companies’ reliability by adjusting to the customer needs.
Supervisor: Kuldar Taveter
Abstract: Positive user experience and usability are important for any digital product or e-service. The main method for evaluating usability is prototyping by mockups at the level of a user interface. However, prototyping has a number of deficiencies such as presuming that a boundary between the software system and its users has already been defined. On the other hand, various predictive technology acceptance models (TAM) have been proposed. Such models aim to predict technology acceptance through analysing user intentions but have not been overly successful because of the gap between user intentions and behaviour. Against this background, the research hypothesis for the PhD project is that user adoption of a digital product or service can be predicted based on analysing the usage of other analogous products in similar usage situations. The purpose of the proposed research work is to put forward a new novel TAM that would be based on the usage of the analogous digital products or services in similar usage conditions, as well as on the other existing technology acceptance models. We intend to answer this research question by applying the research method of action design research. By using this method, we will iteratively develop two artefacts: a novel technology acceptance model and a tool supporting the application of the model by digital product design practitioners. Important aspects of the tool would be obtaining and analysing the usage data about similar applications, and mining from social media and app stores for user feedback on the desired features of the planned new product or service and its analysis. We will validate the model and the tool in in predicting acceptance of the applications designed for older adults and healthcare, related to an ongoing H2020 project, and in two other application domains.
Supervisor: Chinmaya Kumar Dehury
Abstract: It is predicted by Gartner that 50% of the enterprise-generated data will be created and processed in the edge infrastructure comprised of ever-increasing billions of edge devices. Instead of implementing the intelligence in the cloud environment giving the full responsibility to manage a large number of edge devices, a minimal and required intelligence is imposed on the edge devices, enabling the edge intelligence that works in a master-worker approach. The major problem with this approach is that an edge device relies mostly on the data collected by the onboard sensors and the built-in or cloud-instructed intelligence, resulting in little scope for cooperation and collaboration among peers and hence no device-to-device knowledge transfer.
Unlike the traditional approach to cluster the IoT devices based on the location and other meta information, a new research direction focuses on clustering the intelligence that the edge infrastructure is equipped with, also known as Clustered Edge Intelligence (CEI). Intelligence discoverability and observability are some of the primary challenges that need to be addressed to realize the full capacity of CEI. This Ph.D. project will focus on designing and developing dynamic intelligence clustering methodology with the discoverability and observability feature for large-scale edge infrastructure.
Supervisor: Mozhgan Pourmoradnasseri, Amnsir Hadachi
Abstract: Active mobility promotes regular physical activity, such as walking and cycling, as a means of transport. It is associated with several public benefits, particularly reducing noise and CO2 emissions and individual health benefits.
With the emergence of smart cities, many technological and infrastructure developments are undergoing. Moreover, an enormous amount of collected spatiotemporal data presents the immense opportunity to provide vision and guidelines for moving towards “smart sustainable cities.” This thesis aims to employ the data and analysis of foot and bike network topology to understand pedestrians’ and cyclists’ preferences and study the effect of urban planning on mobility patterns of active modes of transport. This analysis can be tailored with the technological shift into smart cities by prioritizing accessibility, connectivity, and sustainability of the network. Urban planners and policymakers can acquire the derived measurements and factors to improve the safety and proximity of the urban transportation network and citizens’ quality of life.
Supervisor: Anna Aljanaki
Abstract: Music similarity is a central concept in music recommendation. Content-based approaches have always been very important for musical data with their potential to improve music recommendation experience on “beyond metrics” of serendipity and diversity, increase novelty and alleviate cold start problem. In this PhD project the potential of personalized music similarity embeddings would be explored. As most concepts in Music Information Retrieval, music similarity is a highly subjective and difficult to define concept, it’s definition is only possible when taking into account personal experience, cultural aspects and musical proficiency of the listener.
Disentangling both acoustic and metadata-related aspects of music similarity in trainable embeddings would allow to isolate the factors that influence music similarity for the particular music listener, which would lead to more relevant recommendations. The research on personalized embedding spaces for music similarity will hopefully allow for content-based recommendation to be closer in performance to collaborative filtering based ones. For music, where the amount of items is very large and the consumption time is short, unlike with film, books, and retail commerce, improving content-based recommendation is of crucial importance.
Supervisors: Kallol Roy, Osamu Shimmi
Abstract: Rapid progress in human genome sequencing during the last decades has shown that successive genetic mutations of oncogenic genes increase the risk of cancer. Further development of the human genome project is expected to facilitate establishing personalized medicine. Recent studies in biology reveal that clusters of cells with oncogenic mutations do not always result in uncontrolled growth of the malignant tumors, but multiple mechanisms rather suppress tumor formation. Tumor suppressive mechanisms not only rely on the specific genes but also on the cellular environment. This indicates that reading gene sequencing is not sufficient for predicting human cancers. To understand the correlation between gene mutations and uncontrolled tumor growth, research on model organisms of Drosophila melanogaster is crucial. Though this experimental approach shows some promise, it has a restriction to understand the dynamic process of cancer risk. The goal of the Ph.D. project is to build an Explainable Artificial Intelligence System (XAI) that predicts the cancer risk from microscopy images. The Explainable Artificial Intelligence System builds an internal model of tumor growth by training on microscopy images features. The XAI learns to assign attention to important parts of images that contribute to cancer risk. The XAI learns the experience of what it is like to be cancer through a supervised manner by seeing examples of millions of microscopy images and also explain the cause of cancer.
Supervisor: Raimundas Matulevičius
The digital transformation necessitates a well-functioning nationwide information security management. In 2021, the Estonian Information Security Standard (E-ITS) based on the new paradigm was published in Estonia, which replaced the database-based ISKE (Three-Level Security Baseline of Information Systems). The ISKE methodology was based on the translation of the BSI standard of the German Information Security Agency. Although the new standard is based on BSI renewed version, to Estonian version it is thoroughly reworked, simplified, and adapted to Estonian needs and culture. However, the E-ITS and its management require the development of the standard document itself and creating the entire surrounding ecosystem so that organizations can implement the standard effectively. For that, a method of indicators needs to be developed to understand the current state of information security implementation, identify more general bottlenecks and support them centrally. In addition, it is necessary to find the means to obtain repeatable and comparable results for supervisory authority, which would help to ensure that the implementation of security measures is effective in practice. So far, no systematic measurement of the level of information security has been performed in Estonia. It is also necessary to find methods for updating the content of the standard while keeping it following the requirements of Estonia and the requirements of international standards. The creation of such an information security standard ecosystem model can also be re-used in information security management in other countries.
Supervisor: Dominique Unruh
Abstract: The security of cryptographic systems is normally ensured by mathematical proofs. Due to human error, however, these proofs often contain errors, limiting the usefulness of said proofs. This is especially true in the case of quantum protocols since human intuition is well-adapted to the classical world, but not to quantum mechanics. To resolve this problem, we need methods for verifying cryptographic security proofs using computers (i.e., for “certifying” the security).
The goal of this thesis is to: (a) Formalize post-quantum security proofs of relevant cryptosystems with computer-verified proofs using tools developed in our group. (b) Develop new automation techniques for doing such security proofs and to use them to improve the verification tools.