Přístupnostní navigace
E-application
Search Search Close
study programme
Faculty: FITAbbreviation: DIT-ENAcad. year: 2022/2023
Type of study programme: Doctoral
Study programme code: P0613D140029
Degree awarded: Ph.D.
Language of instruction: English
Tuition Fees: 4000 EUR/academic year for EU students, 4000 EUR/academic year for non-EU students
Accreditation: 8.12.2020 - 8.12.2030
Profile of the programme
Academically oriented
Mode of study
Combined study
Standard study length
4 years
Programme supervisor
prof. Ing. Lukáš Sekanina, Ph.D.
Doctoral Board
Chairman :prof. Ing. Lukáš Sekanina, Ph.D.Councillor internal :doc. Ing. Jan Kořenek, Ph.D.prof. Ing. Pavel Václavek, Ph.D.prof. Ing. Tomáš Vojnar, Ph.D.prof. Dr. Ing. Jan Černockýprof. RNDr. Milan Češka, CSc.prof. Ing. Adam Herout, Ph.D.prof. Ing. Jan M. Honzík, CSc.prof. Ing. Tomáš Hruška, CSc.prof. Dr. Ing. Pavel Zemčík, dr. h. c.prof. RNDr. Alexandr Meduna, CSc.prof. Dr. Ing. Zbyněk Raidaprof. RNDr. Josef Šlapal, CSc.Councillor external :prof. Ing. Jiří Sochor, CSc. (FI MUNI)prof.,RNDr. Jiří Barnat, Ph.D. (FI MUNI)
Fields of education
Study aims
The goal of the doctoral degree programme is to provide outstanding graduates from the master degree programme with a specialised university education of the highest level in certain fields of computer science and information technology, including especially the areas of information systems, computer-based systems and computer networks, computer graphics and multimedia, and intelligent systems. The education obtained within this degree programme also comprises a training and attestation for scientific work.
Graduate profile
Profession characteristics
FIT graduates in general and FIT doctoral graduates in particular do not have a problem finding employment at scientific, pedagogical or management positions both in Czech Republic and abroad.
Fulfilment criteria
The requirements that the doctoral students have to fulfil are given by their individual study plans, which specify the courses that they have to complete, their presupposed study visits and active participation at scientific conferences, and their minimum pedagogical activities within the bachelor and master degree programmes of the faculty. A successful completion of the doctoral studies is conditional on the following:
Study plan creation
The rules are determined by the directions of the dean for preparing the individual study plan of a doctoral student. The plan is to be based on the theme of his/her future dissertation thesis and it is to be approved by the board of the branch.
https://www.fit.vut.cz/fit/info/smernice/sm2018-13-en.pdf
Availability for the disabled
Brno university of technology provides studies for persons with health disabilities according to section 21 par. 1 e) of the Act no. 111/1998, about universities and about the change and supplementing other laws (Higher Education Act) as amended, and according to the requirements in this field arising from Government Regulation No. 274/2016 Coll., on standards for accreditation in higher education, provides services for study applicants and students with specific needs within the scope and in form corresponding with the specification stated in Annex III to Rules for allocation of a financial contribution and funding for public universities by the Ministry of Education, Youth and Sports, specifying financing additional costs of studies for students with specific needs.Services for students with specific needs at BUT are carried out through the activities of specialized workplace - Alfons counselling center, which is a part of BUT Lifelong Learning Institute - Student counselling section.Counselling center activities and rules for making studies accessible are guaranteed by the university through a valid Rector's directive 11/2017 concerning the status of study applicants and students with specific needs at BUT. This internal standard guarantees minimal stadards of provided services.Services of the counselling center are offered to all study applicants and students with any and all types of health disabilities stated in the Methodological standard of the Ministry of Education, Youth and Sports.
What degree programme types may have preceded
The study programme builds on both the ongoing follow-up Master's programme in Information Technology and the new follow-up Master's programme in Information Technology and Artificial Intelligence.Students can also, according to their needs and outside their formalized studies, take courses and trainings related to the methodology of scientific work, publishing and citation skills, ethics, pedagogy and soft skills organized by BUT or other institutions.
Issued topics of Doctoral Study Program
The topic concerns algorithms of image, video, and/or signal processing. Its main goal is to research and in-depth analyze existing algorithms and discover new ones so that they have desirable features and so that they are possible to efficiently implement. Such efficient implementation can be but does not necessarily have to be part of the work but it is important to prepare the algorithms so that they can be efficiently implemented e.g. in CPU, in CPU with acceleration through SSE instructions, in embeded systems, even in combination with FPGA, in Intel Xeon PHI, in extremely low power systems, or in other environments. It is possible to exploit algorithms of artificial intelligence, such as neural networks, especially CNNs The application possibilities of the algorithms are also important and the application can be but does not have to be part of the work. The algorithms/applications of interest include:
After mutual agreement, individually selected algorithms can be considered as well as soon as they do belong to the general topic.
Collaboration on grant projects, such as TACR, MPO, H2020, ECSEL (possible employment or scholarship).
Tutor: Zemčík Pavel, prof. Dr. Ing., dr. h. c.
The project is concerned with advanced methods of computational photography. The aim is to research new computational photography methods, which comprises software solutions potentially supported by new optics and/or hardware. Our interest is on HDR image and video processing, color-to-grayscale conversions, spectral imaging, and others.
Tutor: Čadík Martin, doc. Ing., Ph.D.
The project is concerned with advanced rendering and global illumination methods. The aim is to research new photorealistic (physically accurate) as well as non-photorealistic (NPR) simulations of interaction of light with the 3D scene. Cooperation and research visits with leading research labs are possible (Adobe, USA, MPII Saarbrücken, Německo, Disney Curych, Švýcarsko, INRIA Bordeaux, Francie).
Machine learning is in the centre of research of artificial intelligence. Many researchers worldwide are dealing with the topics related to machine learning, both in academia and industry. This very dynamic field is characterized with fast transfer of solutions into practical use.
The topics in this domain are defined by premier scientific conferences, where top-class researchers meet, for example ICML (International Conference on Machine Learning), NeurIPS (Advances in Neural Information Processing Systems), IJCAI (International Joint Conference on AI), COLT (Conference on Learning Theory).
This thesis will be advised by an external mentor, who will also define its particular topic.
Interesting research challenges are contained within (but are not limited to) these topics:
General Machine Learning (e.g., active learning, clustering, online learning, ranking, reinforcement learning, semi-supervised learning, time series analysis, unsupervised learning)
Deep Learning (e.g., architectures, generative models, deep reinforcement learning)
Learning Theory (e.g., bandits, game theory, statistical learning theory)
Optimization (e.g., convex and non-convex optimization, matrix/tensor methods, sparsity)
Trustworthy Machine Learning (e.g., accountability, causality, fairness, privacy, robustness)
There are many application domains, where advanced machine learning methods can be deployed.
The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with researchers from highly respected research units. A combined (external) form of study and full employment at KInIT is expected.
Tutor: Bieliková Mária, prof. Ing., Ph.D.
Dissertation focuses on the security of wireless local area networks. As part of the solution, student should become familiar with selected wireless networks and their security. The goals of this work: studying the theory of wireless networks, their properties and possibilities of attacks, testing the basic types of attacks, designing a new method of protection, experiments, evaluating the results and proposing the direction of further research.Co-supervised by dr. Kamil Malinka.
Tutor: Hanáček Petr, doc. Dr. Ing.
SCADA (Supervisory Control And Data Acquisition) and Industrial Control Systems (ICS) communication is intended to control and monitor industrial processes in smart factories, smart grids, etc. Disruption of communication may interrupt important industrial processes or cause a blackout of critical supplies like water or energy. To protect SCADA/ICS communication, we can apply advanced SCADA/ICS monitoring that creates normal behavior profiles and detects deviations (anomalies). Communication profiles can be described using statistical methods. We can model communication using automata or create models using machine learning. The research aims to describe the behavior of industrial communication and propose a suitable representation using multidimensional profiles. It is then necessary to verify what types of anomalies are covered by these profiles and how the accuracy of detection can be increased. The topic is a part of the research project Security monitoring of ICS communication in the smart grid (Bonnet). Co-supervisor: Matoušek Petr, Ing., Ph.D., M.A.
Tutor: Ryšavý Ondřej, doc. Ing., Ph.D.
Current internet traffic is mostly encrypted using TLS or DTLS. Detection of network applications communicating over the local network is important for network monitoring and cyber security. There are several approaches how to detect an application with the encrypted network traffic. One approach is based on TLS fingerprinting using the JA3/JA3S method. Another approaches include statistical analysis of encrypted traffic or the application of machine learning. A student is expected to research and evaluate current methods for detecting applications in encrypted traffic and to propose a new approach for fast and reliable application detection using advanced methods.Co-supervisor: Matoušek Petr, Ing., Ph.D., M.A.
The topic aims an the reseach of advanced architectures of neural networks for the tasks of speech and speaker recognition. Although empirically, the results of such NNs are often excellent, our knowledge and understanding of such representations is insufficient. This PhD topic has an ambition to fill this gap and to study neural representations for speech and text units of different scopes (from phonemes and letters to whole spoken and written documents) and representations acquired both for isolated tasks and multi-task setups.The topic is related to the GACR project of excellence in fundamental research "Neural Representations in multi-modal and multi-lingual modeling", on which we cooperate with colleagues from UFAL MFF Charles University in Prague. The assignment requires an interest in mathematics, statistics, machine learning and speech processing; experience with Python and its libraries for machine learning is an advantage.
Tutor: Burget Lukáš, doc. Ing., Ph.D.
The rising trend in artificial intelligence usage brings novel cybersecurity approaches on both sides - attacker and defender. The most prominent examples are deepfake usage to counterfeit biometric systems or security analytics, using deep learning for cyber-attacks detection. The goal of this work is to analyze all existing approaches, their properties, and potential applications. The work should then propose novel applications of AI for the problems that were not resolved before while also implementing the most interesting application.
Co-supervised by dr. Kamil Malinka.
Problem Statement: Mental stress and anxiety are two of the mental health conditions that often occur together. In such a case, the person is stressed and is not able to control the worry, and it correspondingly affects his/her social and occupational activities. Hence, proper assessment and diagnosis for mental stress and anxiety is required in order for a person to effectively keep taking part in his/her normal daily tasks and activities.
Issues with Current Solutions: Unfortunately, conventional assessment and diagnostic measures are subjective in nature and are used only when the symptoms are already evident due to advanced stages of mental stress and anxiety. However, mental stress and anxiety do not occur overnight, rather it is a long process. Hence, detection of symptoms is required at early stages of mental stress and anxiety because that may result in a cure or at least it will delay the onset of serious mental health issues associated with them, for example, depression, Generalized Anxiety Disorder etc.
Challenges: Unlike other diseases where the symptoms like fever and cough allow people to seek help, symptoms at early stages of mental stress and anxiety are not easily identifiable. Hence, the brain needs to be continuously monitored for any sign of change or deterioration in order to detect the symptoms at early stage.
Solution: The solution lies in the development of an objective and quantitative method that can detect mental stress and anxiety at an early stage. Perception of mental stress and anxiety originates in the brain; therefore, this research investigates the neurophysiological features extracted from brain electroencephalogram (EEG) signal to measure mental stress and anxiety at early stage. This will require development of method for extraction of features as well as pattern recognition approach to provide a solution. The EEG dataset is already available for this project.
Few Words About Supervision: I have recently moved to FIT at Brno University of Technology. I have decade long experience of working in the field of neuro-signal and neuroimage processing and I am currently in the process of setting up a research group in this area. This is a multidisciplinary project and it will involve working with clinicians. However, the core of the project is related to IT in terms of development of a new method. Please feel free to contact me at malik@fit.vutbr.cz
Tutor: Malik Aamir Saeed, prof., Ph.D.
Automatic speech recognition (ASR) is among popular machine learning tasks already well exhibited by industry in various applications. Recent advances in deep learning enabled for the use of speech technologiesin areas where very also low error-rates are expected, such as in recognition of air-traffic communication (i.e. an intensive use of spoken language between controllers and pilots). The objective of this thesis is to focus on unsupervised training of ASR components by exploiting auxiliary information provided by other resources, such as radar data or ADS-B communication (surveillance technology in which an aircraft determines its position via satellite navigation or other sensors and periodically broadcasts it). Possible application of ASR can be expected to (i) support real-time voice communication, or (ii) to automatically process and transcribe large data audio archives.
Tutor: Motlíček Petr, doc. Ing., Ph.D.
Problem Statement: Biometrics technology enables recognition of a person based on certain unique characteristic of human body. It is the primary tool for identification of a person and it has various applications including banking, security, immigration etc. Hence, it is critical to have a foolproof biometrics method.
Issues with Current Solutions: Fingerprints and face identification are commonly used in today's world for verifying a person's identity. Some other biometrics technologies include voice recognition, gait recognition, iris recognition and retina-based pattern recognition. However, issues like ageing affects the recognition of a person based on these technologies. In addition, some of them are easy to replicate and hence a person's identity can be hacked.
Challenges: Ageing is one of the main challenges for any biometric method. In addition, physical damage to the body as well as replication of the biometric identifier are the main challenges.
Solution: This research will investigate the brain biometrics using the human electroencephalogram (EEG) brain signals. It will involve identifying a person's unique brain activity patterns and it has the advantage that it is invisible from outside and hence cannot be replicated or cloned. Development of a brain biometric method will involve identification of a unique pattern from the brain signals that can be used as an identifier for the person.
Tutor: Herout Adam, prof. Ing., Ph.D.
The aim of the thesis is to create 3D face model from 2D photos of diverse origins, namely:
Tutor: Drahanský Martin, prof. Ing., Ph.D.
Majority of network communication today is encrypted. One of the remaing plain-text communication is the DNS that is essential for most network operations. In the past years the attackers misuse DNS communication to redirect internet connections or for DDoS attacks. DNS communication is also misused by malware for transporting stealed data over a hidden channel. Domains generated by DGA algorithms are also misused by botnets for C&C communication. In this project, a student is expected to study typical classes of DNS cyber attacks based on MITRE ATT&CK classification and to propose a method for DNS cyber threats detection using machine learning and artificial intelligence techniques. Co-Supervisor: Ing. Petr Matoušek, Ph.D., M.A.References:[1] Daihes Y., Tzaban H., Nadler A., Shabtai A. (2021) MORTON: Detection of Malicious Routines in Large-Scale DNS Traffic. In: Bertino E., Shulman H., Waidner M. (eds) Computer Security - ESORICS 2021. ESORICS 2021. Lecture Notes in Computer Science, vol 12972. Springer, Cham. https://doi.org/10.1007/978-3-030-88418-5_35[2] M. Grill, I. Nikolaev, V. Valeros and M. Rehak, "Detecting DGA malware using NetFlow," 2015 IFIP/IEEE International Symposium on Integrated Network Management (IM), 2015, pp. 1304-1309, doi: 10.1109/INM.2015.7140486.[3] S. Torabi, A. Boukhtouta, C. Assi and M. Debbabi, "Detecting Internet Abuse by Analyzing Passive DNS Traffic: A Survey of Implemented Systems," in IEEE Communications Surveys & Tutorials, vol. 20, no. 4, pp. 3389-3415, Fourthquarter 2018, doi: 10.1109/COMST.2018.2849614.
Deep convolution networks have been a clear trend of machine learning for image analysis in recent years. However, in tasks with a very small and specific data set, where it is not enough to use data augmentation or GAN concepts, their usage is still problematic.The goal of the dissertation thesis is to explore, analyze and design new architectures of deep convolutional networks and approaches to their learning for image analysis tasks in which the size of the annotated data set is extremely small or is gradually growing. For learning neural networks it is possible to use unannotated data or partially annotated data in the form of a limited user input.Proposed methods will be applied in the projects on which the supervisor participates.
Tutor: Španěl Michal, Ing., Ph.D.
The aim of the work is to create a detection mechanism of moving objects (lasers, drones) with revealing their position or broadcast sources, namely:
Problem Statement: Learning and memory are two fundamental inter-related cognitive processes. Learning is the change in behavior because of an experience; while memory is the ability to store and recall learned experience (which deals with Long Term Memory). Therefore, it is vital to assess the capabilities of long-term memory of an individual and provide relevant classification (ability to remember, speed of recall).
Issues with Current Solutions: Currently, the existing studies focus on short term memory. However, the corresponding studies related to long-term memory are much smaller in number. The main reason is that an experiment involving short term memory can be completed very quickly but an experiment involving long-term memory requires considerable time and effort.
Challenges: Over the last few decades, researchers and educationalists have put in considerable efforts in developing content that can provide effective learning. Both 2D and 3D audio-visual content and now virtual reality content is being developed. The challenge lies in relating such content with the long-term human memory that could lead to effective learning.
Solution: This research plans to utilize 2D and 3D educational content and develop a corresponding long-term memory classification and grading method. The method will be based on learning and measuring human memory performance by analyzing the neurophysiological effects (from brain EEG data) using 3D and virtual reality contents during learning and memorization process. The EEG dataset is already available for this project.
Problem Statement: This research deals with depression or to be more specific - unipolar depression or Major depressive disorder (MDD). MDD is a common mental illness that may impede daily life functioning, hence creating problems during individual social and personal activities. Hence, it is critical to have the right diagnosis as well as treatment selection for MDD patients.
Issues with Current Solutions: Currently, subjective questionnaires are used by clinicians to diagnose MDD. However, there are chances of misdiagnosis because of similar symptoms of bipolar disorder and schizophrenia. In addition, the treatment selection is done based on the experience of a clinician who prescribes an anti-depressant and then checks its efficacy for 4 to 6 weeks before deciding to either continue with same medicine or change to a new medicine. This is a repetitive process till the right medicine is found for the patient.
Challenges: Treatment management for major depressive disorder (MDD) has been a challenge including diagnosis and treatment selection because of similarities in symptoms of various diseases. Further, MDD is gender specific with different brain patterns for males and females.
Solution: This research deals with the development of diagnostic and treatment selection method for MDD (Unipolar Depression). It will be based on objective measures of brain activities from EEG signal that can detect unipolar depression. The solution will be based on a machine learning approach inherently based on feature extraction from the recorded EEG data of the patients. This method will complement and assist psychiatrists in diagnosis of unipolar depression. The EEG dataset is already available for this project.
The topic focuses embedded image, video and/or signal processing. Its main goal is to research capabilities of "smart" and "small" units that have such features that allow for their applications requiring smyll, hidden, distributed, low power, mechanically or climatically stressed systems suitable of processing of some signal input. Exploitation of such systems is perspective and wide and also client/server and/or cloud systems. The units themselves can be based on CPU/DSP/GPU, programmable hardware, or their combination. Smart cameras can be considered as well. Applications of interest include:
A possibility exists in collaboration on grant projects, especially the newly submitted TAČR, H2020, ECSEL ones (potentially employment or scholarship possible).
The aim of this work is to generate various damages into synthetic fingerprints and analysis of their quality. The work will consist of:
Deep convolution networks have been a clear trend of machine learning for image analysis in recent years. Neural networks can also be used for 3D image analysis, where the network works with 3D convolutions. However, this approach is problematic because of its huge memory and computational requirements.The aim of the dissertation thesis is to explore, analyze and design new architectures of Graph Neural Networks and approaches to their learning for 3D object shape recognition tasks, where the data set consists of various 3D data representations - e.g. 3D meshes, voxel representation, etc.Proposed methods will be applied in the projects on which the supervisor participates.
The use of drones in a number of specific situations is an important trend today not only in industry, but especially in security operations. This trend brings new problems in the effective communication of man with the drone, ie how to increase the pilot's efficiency, reduce his cognitive load, increase his orientation in a complex environment or when cooperating with multiple drones, etc.The aim of this work is to explore new possibilities in the field of effective and intuitive use of sensory data and their fusion with other available data sources in the interaction of the pilot with the drone. The solution requires:
Tutor: Beran Vítězslav, doc. Ing., Ph.D.
The trend of the manufacturing industry is the introduction of collaborative robots into production, which allows for closer human-robot cooperation. The aim is to streamline production by using robots for repetitive activities and workers for complex activities, their robotization would be too expensive and not very scalable. This trend brings new problems in how to communicate effectively with robots: to have an idea of the state of the robot and its understanding of the situation, and to control and program the robot easily and naturally.The aim of this work is to explore new possibilities of human-robot communication using modern technologies and devices. The solution requires:
The project deals with image and video quality assessment metrics (IQM). The aim is to explore new ways how to incorporate human visual system properties into IQM. In particular, we will consider perception of HDR images, and utilization of additional knowledge (in form of metadata, 3D information, etc.) about the tested scenes using machine learning (e.g. neural networks).
The project is concerned with advanced methods of image processing. The aim is to research new methods using machine learning, in particular deep convolutional neural networks.
This topic focuses on the process and techniques used for cyberthreat detection, obtaining and representation of threat data. The goal of this dissertation thesis is to study the current techniques and propose a solution that effectively improves the process and its results usable for malware detection and threat intelligence.
The topic of identifying and extracting specific information from documents on the Web has been the subject of intensive research for quite a long time. The basic obstacles that make this problem difficult are the loose structure of HTML documents and absence of meta-information (annotations) useful for recognizing the content semantics. This missing information is therefore compensated by the analysis of various aspects of web documents that include especially the following:
A background knowledge about the target domain and the commonly used presentation patterns is also necessary for successful information extraction. This knowledge allows a more precise recognition of the individual information fields in the document body.
Current approaches to information extraction from web documents focus mainly modeling and analyzing the documents themselves; modeling the target information for more precise recognition has not yet been examined in detail in this context. The assumed goals of the dissertation are therefore the following:
Experimental implementation of the proposed methods using existing tools and experimental evaluation on real-world documents available on the WWW is also an integral part of the solution.
Tutor: Burget Radek, doc. Ing., Ph.D.
Thanks to the deployment of new technologies in our daily lives, a huge amount of data of various types is constantly being generated. The created datasets need to be processed efficiently - the information contained in them often supports the correctness and accuracy of decision-making processes. The development of data analysis methods is therefore an important part of IT research. For various reasons, traditional processing methods are not generally applicable, so new approaches need to be sought. They are mostly based on artificial intelligence and machine learning.
Intelligent data analysis has great potential for solving current research tasks in many areas, including the energy domain. This field is undergoing great changes - renewable energy sources, batteries and electromobility convert the nature of the electricity grid and the classic one-way centralized network is becoming a two-way distributed network. This fact raises a number of research questions and problems to solve. Research topics include:
- optimal management of microgrids (small energy networks with renewable energy sources) enabling energy sharing between customers,
- predicting customer consumption and renewable production needed for efficient microgrid management,
- disaggregation of energy consumption into consumption of individual appliances, which will allow a better understanding of the nature of consumer consumption,
- investigation of anomalies and / or extreme values in the consumption or production of energy from renewable sources.
However, data analysis in a broader sense can be of interest - it is possible to focus on the tasks of prediction, clustering, classification or detection of anomalies in different domains.
Relevant publications:
Rozinajova, V., Bou Ezzeddine, A., Grmanova, G., Vrablecova, P., Pomffyova, M.(2020): Intelligent Analysis of Data Streams, Published in: Towards Digital Intelligence Society: A Knowledge-based Approach, Publish date 22 December 2020
Kloska M., Rozinajova V. (2021): Towards Symbolic Time Series Representation Improved by Kernel Density Estimators. In: Hameurlain A., Tjoa A.M. (eds) Transactions on Large-Scale Data- and Knowledge-Centered Systems L. Lecture Notes in Computer Science, vol 12930. Springer, Berlin, Heidelberg.
Tutor: Rozinajová Věra, Doc., Ph.D.
Dissertation focuses on the security of IoT systems. The goals of this work: studying the theory of IoT systems, their properties and possibilities of attacks, testing the basic types of attacks, designing a new method of protection, experiments, evaluation of results and design of further research.Co-supervised by dr. Kamil Malinka.
Although there are technologies that allow publishing data on the WWW in machine-readable form (such as JSON-LD, RDFa, etc.), a large amount of structured data is still published on the web in the form of plain HTML/CSS code, which greatly limits the possibilities of their further use.Recently, new machine learning methods (especially deep learning methods) are gaining importance, which show interesting results, e.g., in recognizing important entities in weakly structured or unstructured data (e.g., text or images). However, the area of web document processing has not received much attention from this perspective. Existing works deal with the identification of simple data items and neglect structured data and more complex usage scenarios.The goal of this topic is to analyze and develop web content models suitable as input for machine learning and, at the same time, machine learning methods suitable for recognizing structured data in web documents.
Množstvo problémov týkajúcich sa bezpečnosti je potrebné riešiť prostredníctvom metód založených na umelej inteligencii a s pomocou strojového učenia. Medzi aktuálne výskumné problémy v oblasti informačnej bezpečnosti patrí analýza údajov veľkých bezpečnostných logov, detekcia širokej škály anomálií pri sieťovej komunikácii, detekcia škodlivého správania.
Jednou z aktuálnych výziev umelej inteligencie, je učenie bez potreby označenej dátovej sady. Dnešné modely väčšinou dosahujú vynikajúce výsledky, ale veľa krát ich dosahujú vďaka veľkým množstvám označených trénovacích vzorov. Existujú ale problémy, pri ktorých by vytvorenie takejto dátovej sady bolo veľmi nákladné, alebo sa vytvoriť vôbec nedá. Prístupy učenia posilňovaním (reinforcement learning) umožňujú učiť agenta vykonávať akcie tak, aby dosiahol svoj cieľ, pričom sa snaží maximalizovať svoje odmeny. Takýto agent nepotrebuje označenú dátovú sadu, postačí mu len vonkajšie ohodnotenie jeho správania (odmena za dobré rozhodnutia, trest za zlé).
Výzvou je interpretovateľnosť strojového učenia pri riešeniach bezpečnosti a dôvery.
Súvisiace publikácie:
CHUDÁ, Daniela - KRÁTKY , Peter - BURDA, Kamil. Biometric Properties of Mouse Interaction Features on the Web. In Interacting with Computers : The Interdisciplinary Journal of Human-Computer Interaction. Vol. 30, iss. 5 (2018), s. 359-377. https://doi.org/10.1093/iwc/iwy015
P. Lacko, V. Kvasnička. Mixture of Expert Used to Learn Game Play. In: Kůrková V., Neruda R., Koutník J. (eds) Artificial Neural Networks - ICANN 2008, Lecture Notes in Computer Science, vol 5163. Springer. https://doi.org/10.1007/978-3-540-87536-9_24
Výskum bude doktorand vykonávať v rámci Kempelenovho inštitútu inteligentných technológií (KInIT, https://kinit.sk) v Bratislave v spolupráci s priemyselnými partnermi alebo výskumníkmi zo svetovo uznávaných výskumných skupín. Predpokladá sa kombinovaná (externá) forma štúdia a pracovný pomer na plný úväzok v KInIT.
Tutor: Chudá Daniela, doc. Mgr., Ph.D.
The models created in machine learning can only be as good as the data on which they are trained. Researchers and practitioners thus strive to provide their training processes with the best data possible. It is not uncommon to spend much human effort in achieving upfront good general data quality (e.g. through annotation). Yet sometimes, upfront dataset preparation cannot be done properly, sufficiently or at all.
In such cases the solutions, colloquially denoted as human-in-the-loop solutions, employ the human effort in improving the machine learned models through actions taken during the training process and/or during the deployment of the modes (e.g. user feedback on automated translations). They are particularly useful for surgical improvements of training data through identification and resolving of border cases. This is also directly related to explainability and interpretability of models.
Human-in-the-loop approaches draw from a wide palette of techniques, including active and interactive learning, human computation, crowdsourcing (also with motivation schemes of gamification and serious games), and collective intelligence. Each of these fields (or combination thereof) presents opportunities for new discoveries. They border on computer science disciplines such as data visualization, user experience (usability in particular) and software engineering.
The domains of application of machine learning with human-in-the-loop are predominantly those with a lot of heterogeneity and volatility of data. Such domains include online false information detection, online information spreading (including spreading of narratives or memes), auditing of social media algorithms and their tendencies for disinformation spreading, support of manual/automated fact-checking and more.
M. Tomlein, B. Pecher, J. Simko, I. Srba, R. Moro, E. Stefancova, M. Kompan, A. Hrckova, J. Podrouzek, and M. Bielikova. An Audit of Misinformation Filter Bubbles on YouTube: Bubble Bursting and Recent Behavior Changes. In Fifteenth ACM Conference on Recommender Systems (pp. 1-11). 2021, September. https://dl.acm.org/doi/10.1145/3460231.3474241
J. Šimko and M. Bieliková. Semantic Acquisition Games: Harnessing Manpower for Creating Semantics. 1st Edition. Springer Int. Publ. Switzerland. 150 p. https://link.springer.com/book/10.1007/978-3-319-06115-3
The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT, https://kinit.sk) in Bratislava in cooperation with industrial partners or researchers from highly respected research units. A combined (external) form of study and full employment at KInIT is expected.
Tutor: Šimko Jakub, doc. Ing., PhD.
Ubiquity of computing is nowadays generating vast datasets potentially usable for training machine learned models. Yet, these data mostly come not annotated, i.e. without labels necessary for their use for supervised (most conventional) training methods (e.g. classification). Conventional acquisition of proper labels using human force is a costly endeavour. The result is that proper labels are either unavailable, few or of poor quality.
To circumvent the issue, many approaches are emerging and are presently researched by many researchers: meta-learning, transfer-learning, weakly supervised learning, zero/one-shot learning, semi-supervised learning. Each of these fields (or combination thereof) presents opportunities for new discoveries. Orthogonal to this, explainability and interpretability of models is an important factor to consider and advances in this regard are welcome (generally in AI and particularly in the mentioned approaches).
There are several domains of applications, where research of methods and models for addressing small labelled data can be applied. These include (but are not limited to) false information detection, auditing of social media algorithms and their tendencies for disinformation spreading, and support of manual/automated fact-checking.
B. Pecher, I. Srba, and M. Bielikova. Learning to Detect Misinformation Using Meta-Learning. Presented in PhD Forum by B. Pechcer at the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases - ECML-PKDD 2020.
M. Pikuliak, M. Simko, M. Bielikova. Cross-Lingual Learning for Text Processing: A Survey Expert Systems With Applications or its open access mirror. Expert Systems with Applications. Vol. 165, 2021. https://doi.org/10.1016/j.eswa.2020.113765
The topic concerns algorithms of computer graphics and image synthesis. Its main goal is to research new algorithms so that their features and application possibilities are better understood so that they are improved or newly created. If suitable, it is possible to work on various platforms, includeing parallel CPUs, such as x86/64, ARM, Xeon PHI, GPU, etc. or other cores in CUDA, OpenCl, VHDL, etc. Algorithms of interest include:
Collaboration on grant projects, such as TACR, H2020, ECSEL possible (employment or scholarship).
Problem Statement: The importance of mental health has increased significantly over the past decade. However, the methods for the assessment of mental health issues at early stages are still in their infancy compared to the availability of corresponding methods for early assessment of physical health issues. Hence, it is required that due research is done to develop methods for early assessment of abnormalities leading to mental health problems.
Issues with Current Solutions: Unlike physical health parameters, the mental health is assessed through a number of subjective parameters. Hence, there is lack of objective and quantitative methods for mental health assessments. In addition, the patients seek help when their mental health problem is at advanced stage. So, there is lack of continuous monitoring for mental health issues.
Challenges: Many of the abnormalities related to mental health issues are subtle in nature and are related to behavior and other changes in facial expressions, speech and handwriting. In addition, there are changes in cortisol levels, skin conductance, heart rate variability and breathing rate. Hence, there are multiple modalities that should be included for measuring and quantifying any abnormalities related to mental health.
Solution: Every modality has its pros and cons. For example, in neuroimaging, functional magnetic resonance imaging has high spatial resolution (in mm) and low temporal resolution (in seconds) while electroencephalogram has low spatial resolution (in cm) and high temporal resolution (in milliseconds). Combining both of them will result in high spatial as well as high temporal resolution. This research deals with the assessment of abnormalities leading to mental health problems by utilizing multimodal approach. The various modalities may include, but not limited to, electroencephalogram (EEG) brain signals, facial videos, speech audios, handwriting and text from social media. The physiological parameters from various modalities include, but not limited to, the heart rate, breathing rate, dominant emotion, fatigue and stress. Dominant emotion can be classified as positive or negative and then sub-classified as sad, happy, angry etc. Data mining and data fusion techniques will be developed for this multimodal analysis. The corresponding multimodal data is available for this project.
Continuous increase of data available world-wide emphasizes the need of its automatic processing and understanding. Particular challenges are posed by the heterogeneous and unstructured nature of text content provided in natural language. Natural language processing (NLP) ranks among the most prospective subfields of artificial intelligence with great potential for innovative applications affecting everyday life.
Recent advances in neural networks and machine learning allowed to push efficiency and scope of natural language understanding and generation forward. Yet, there remain many research challenges related to particular subtasks, application domains and languages. Further research and various resulting phenomena exploration is necessary. Special attention is drawn by the issues of interpretability and transparency of NLP models or by novel paradigms of learning addressing the problem of low-resource languages.
Particularly interesting challenges include, but are not limited to:
Low-resource language processing
Transfer/multilingual learning
Fairness, interpretability, transparency, explainability and/or robustness for NLP
Domain-specific information extraction, text classification
Visual grounding of natural language, image captioning, multimodal data processing
Deep learning for NLP
M. Pikuliak, M. Šimko, M. Bieliková. Cross-lingual learning for text processing: A survey. Expert Systems with Applications, 2021. https://doi.org/10.1016/j.eswa.2020.113765
P. Korenek, M. Šimko. Sentiment analysis on microblog utilizing appraisal theory. World Wide Web, 2014. https://doi.org/10.1007/s11280-013-0247-z
The research will be performed at the Kempelen Institute of Intelligent Technologies (KInIT) in Bratislava in cooperation with industrial partners or researchers from highly respected research units from abroad. A combined (external) form of study and full employment at KInIT is expected.
Tutor: Šimko Marián, doc. Ing., Ph.D.
The aim of the work is to create a reliable detection mechanism of the presentation attack on the technology of fingerprints, faces and handwriting, namely:
The recommender systems are an integral part of almost every modern Web application. Personalized, or at least adaptive, services have become a standard that is expected by the users in almost every domain (e.g., news, chatbots, social media, or search).
Obviously, personalization has a great impact on the everyday life of hundreds of million users across many domains and applications. This results in a major challenge - to propose methods that are not only accurate but also trustworthy and fair. Such a goal offers plenty of research opportunities in many directions:
Novel machine learning approaches for adaptive and recommender systems
Off-policy learning
Explaining recommendations
Fairness and justice in recommendations
Biases in the recommendations
There are several application domains where these research problems can be addressed, e.g., search, e-commerce, news, and many others.
M. Kompan, P. Gaspar, J. Macina, M. Cimerman, M. Bielikova. Exploring Customer Price Preference and Product Profit Role in Recommender Systems. In IEEE Intelligent Systems int. Journal, 2021. https://doi.org/10.1109/MIS.2021.3092768
M. Svrcek, M. Kompan, M. Bielikova. Towards Understandable Personalized Recommendations: Hybrid Explanations. In Computer Science and Information Systems. Vol. 16, No. 1, 2019. https://doi.org/10.2298/CSIS171217012S
Tutor: Kompan Michal, doc. Ing., PhD.
This thesis is actually not supported by any grant. Anyway, this thesis will be dealt with in cooperation with the Police of the Czech Republic. Thus, we can expect support to get financed by a granted project.Moreover, foreign internship is possible and strongly supported. The aim of the thesis is reconstruction of damaged CD/DVD/BR/HDD surfaces, consisting of:
Tutor: Kolář Dušan, doc. Dr. Ing.
The aim of this work is to create a reliable matching of 2D face with 3D face projection, namely:
Internet of Things is a communication platform that interconnects different types of devices in smart homes, smart buildings, smart hospitals, traffic control systems, etc. IoT devices produce a large amount of monitoring data that include both network status data and application data. By longitudinal monitoring we can create a profile of the IoT device and then observe deviation in its behavior which is important for failure detection or identification of cyber attacks. The topic focuses on security analysis of IoT communication. The goal of this project is to propose a automated method for creating IoT device profiles based on monitoring data and demonstrate how these profiles can be exploited for anomaly detection. Co-supervisor: Matoušek Petr, Ing., Ph.D., M.A.
The idea of developing a smart method supporting a network design, monitoring, and management is not new. However, recent advancements in the AI domain and the introduction of P4 for network device programmability creates a promising environment to achieve this goal.In this context, self-awareness means the ability of a network to observe its state and improve its behavior to meet the stated objectives [1]. Though the primary goal of each network is to deliver data packets, there are also other important non-functional requirements, notably, QoS, security, reducing energy consumption. Project GoalThis work aims to design a network system able to observe the situation and by applying an AI approach to decide about the next behavior to keep providing reliable function even when under cyberattack. The project assumes P4 as the technology enabler to achieve the expected goals.Research ApproachThe research consists of the following major steps:
Semantic web technology allows the representation of information and knowledge for the purpose of its further sharing, for example, in computer applications. Available knowledge databases, such as DBPedia or Wikidata, contain a great amount of useful information and facts. On the current web, however, most of the new information is published in the form of documents most often in HTML, whose further processing is problematic mainly due to their free structure and the absence of explicit information about the meaning of individual parts of the content. There exist two ways to overcome this gap between the classical and the semantic web:
To achieve both these goals, it is necessary to analyze the capabilities of existing ontological models regarding the modelling of the target domains and mapping these descriptions on the content of real-world web pages and documents. Possible applications include, but are not limited to:
Experimental implementation of the proposed methods using the existing tools and experimental evaluation on real-world data and documents is also an integral part of the solution.
The topic of the work is focused on speech recognition and data processing for ATC-pilot communication in aviation. It will cover all components of automatic speech recognition (ASR), i.e. data processing, acoustic model, vocabulary (including special aviation terminology) and language model.The topic is related to the European project Horizon 2020 ATCO2, and will be elaborated in cooperation with project partners. The assignment requires an interest in mathematics, statistics, machine learning and speech processing, the advantage is proficiency in Python and its libraries for machine learning.
Tutor: Černocký Jan, prof. Dr. Ing.
The project deals with geo-localization of mobile devices in unknown environments using computer vision and computer graphics methods. The aim is to investigate and develop new image registration techniques (with geo-localized image database or 3D terrain model). The goal is an efficient implementation of proposed methods on mobile devices as well as search for additional applications in the area of image processing, computational photography, and augmented reality.
The objective of this PhD thesis focuses on the voice biometry, i.e. advanced person identification from voice content. Recent advances in deep learning give a potential to significantly increase performance of person identification systems from various type of biological signals. This thesis primarily aims to advance speaker identification technologies for applicability in law enforcement, while taking into account apriori knowledge provided by end-users. Typical case is a combination of intercepted voice calls with call detail records (i.e. network built from telephone numbers). PhD thesis will also address known problems of machine learning algorithms such as "bias" provided by training data.
Recent advances in Deep Learning enabled for the automatic generation of hyper-realistic fake media content known as Deepfakes. While this technology has positive potential for entertainment applications, its malicious use has been already reported to harm citizens and the society at large by producing obscene content, broadcasting fake news to undermine elections or de-stabilise politics and improving social engineering to commit financial fraud. The severity of the problem calls for the urgent development of Deepfake detection technologies for media companies, multimedia news providers and Law Enforcement Agencies. Also as Deepfakes encompasses several categories of synthetically altered or generated media, a secondary objective is to identify the type of manipulation and the specific technology employed for that purpose, this is referred to as attribution. The objective of this PhD thesis focuses on the "voice" part of the media content and aims to develop "voice" Deepfake detection and attribution algorithms. Novel research directions in one class modeling, spatio-temporal learning (i.e. combination with semantic knowledge), few-shot learning, and adversrial training will be explored.