| Topic title |
Possible scientific supervisors |
Source of funding |
|
Research and implementation of methods for brain-computer interface on the base of deep neural networks
|
prof. dr. Vacius JUSAS |
state non-funded |
Research Topic Summary.
Human-computer interface is the form of communication between a user and computer, when the user issues commands for computer in natural form. Such an interface can be implemented using various forms depending on the purpose and user requirements. Natural and user-friendly forms of the implementation shorten time of the user learning to use the implemented tool. Brain – Computer interface is a system that converts the brain activity into the control signals for a computer or some other electronic device. Such a system has large field of application ranging from the disabled persons to playing games. The clear relationship between the brain activity and the commands to muscles is not disclosed yet. Nevertheless, it is possible using mathematical methods to analyse the brain activity and to form the control signals for the external electronic devices. For this purpose, the method is needed to scan and to save the brain activity. The formation of electroencephalogram (EEG) is one of such methods. This is a non-invasive investigation of bioelectrical brain activity. The measured signals of brain activity are non-stationary and non-linear and they are directly dependent of the particular person. Consequently, it is difficult to distinguish the effective control signals. For this purpose, the feature extraction and classification methods are used. The methods of neural networks are dominant among classification methods. The deep neural networks present very good results for image recognition and classification. The signals of brain activity can be presented in the form of image. The purpose of the investigation is to apply the deep neural networks to recognize and classify the signals of brain-computer interface.
|
| Artificial intelligence based model for predicting early symptoms of Alzheimer ‘s |
prof. dr. Rytis MASKELIŪNAS |
state non-funded |
| Deep learning based algorithm for generating musical melodies for more than one instrument songs |
prof. dr. Rytis MASKELIŪNAS |
state non-funded |
|
Multicriteria decision making in finance using explainable artificial intelligence
|
doc. dr. Audrius KABAŠINSKAS |
state non-funded |
Research Topic Summary.
Artificial intelligence is one of the hot topics of the early 21st century in the physical and technological sciences. Large-scale investors such as pension funds, banks and investment companies have been using artificial intelligence principles for some time, however, most of techniques they use are limited to machine learning or artificial neural networks. The resulting solutions are difficult to explain and interpret. Moreover, the worst thing is that it is often it is nearly impossible to verify if the resulting solution is exactly what we aspired and whether it is optimal. Because investment decisions usually depend on many factors, such as expected return, risk, ethics, investor's risk profile and personal characteristics, all of these must be taken into account when developing an autonomous system or robo-advisor. Quite often such problems are quite easily solved by conventional methods, however, as the amount of information and number of criteria increases, the complexity of the task can become unbeatable even for supercomputers. It is therefore necessary to develop new methods based on artificial intelligence that are understandable and explainable and that their results can be validated. Therefore, the main task is: to create artificial intelligence based methods that are explainable and interpretable for most of the investment decision makers.
|
|
Research of methods for development of explainable artificial intelligence algorithms
|
doc. dr. Agnė PAULAUSKAITĖ-TARASEVIČIENĖ |
state non-funded |
Research Topic Summary.
Artificial Intelligence (AI) has achieved a wide range of capabilities from image recognition to language processing and data analysis. But, the automatic decision made by “black box”- based algorithm is hard to explain and even their designers cannot explain why the algorithm arrived at a specific decision. Quit often, this kind of algorithm for generalizing of training data creates rules that may not fit to another sample of data, i.e., it not generalizing properly or partly generalizing a given data. Therefore, the widespread application of artificial intelligence techniques is limited by the lack of trust of people in algorithmic decisions. Especially when there are no clear and objective criteria and rules for explaining one or the other decision made by the algorithm (e.g. Convolutional Neural Network – a deep learning algorithm). The absence of clear connection between accuracy and interpretability creates a barrier to a human consistently predict the model results.... For further information please contact the supervisor of the topic.
|
| Image analysis using deep learning for seabed coverage evaluation |
prof. dr. Evaldas VAIČIUKYNAS |
state non-funded |
|
Design and analysis of long series dependence estimators
|
doc. dr. Tomas RUZGAS |
state non-funded |
Research Topic Summary.
The research goal is to create knowledge about autocorrelation methods of long series that are effective in the case of non-stacionarity. In order to achieve this goal, it is necessary to study mathematical estimators based on the non-parametric analysis. The goal of this study will be achieved by four main tasks: A) to create a probabilistic model for the statistical relationship of series and to create model identification procedures; B) to design constructive autocorrelation methods based on this model and its its identification procedures; C) to offer mathematical estimators for taking account of additional information related to series context in autocorrelation methods; D) to study the proposed solutions, to perform comparative analyzes of alternative statistical relationship estimators on the basis of real data.
|
|
Modelling of Financial Network and Their Topology Using Graph Theory
|
doc. dr. Kristina ŠUTIENĖ |
state non-funded |
Research Topic Summary.
Recently, graph theory methods are gaining a high interest in finance and economics. The reason is that the financial entities and markets are strongly linked in a globalized world, and graph theory as the framework could be used for describing and modelling the structure of complex financial network. As such, the model to be developed might be used to investigate financial network characteristics such as market stability, network contagion, spillover, risk sharing among network entities, network shock propagation, etc. The solution to these individual questions depends on the topology of the network and its possible evolution over time. Therefore, the purpose of this work is to develop methods for modelling the financial network and its topology, focusing on financial risk assessment by introducing network characteristics and proposing stress-testing methods.
|
|
Mathematical models of pension system
|
doc. dr. Audrius KABAŠINSKAS |
state non-funded |
Research Topic Summary.
The idea is to propose the methods/methodology that provides recommendations for the pension system participant (person) in choosing the efficient pension fund with respect to various criteria including stochastic dominance in time by increasing his possibilities to accumulate optimal amount of money in long term perspective. The main goal is to develop the dynamic risk assessment model based on stochastic dominance principles that serves as a tool for the performance analysis of pension funds. The tasks to be implemented include: 1. To identify a set of risk factors for quantifying the systemic and non-systemic risk that arises during the accumulation for the retirement in long term; 2. To analyse the performance results of 2nd and 3rd pillar pension funds in Lithuania over time using technical analysis; 3. To develop the dynamic multi-criteria pension fund risk and performance assessment model by integrating stochastic dominance principles, data envelopment analysis and other approaches allowing additional gains comparing to static risk assessment techniques; 4. To develop a Markov chain model based on phase type distributions for estimation of expected number of PHs participants depending on funds risk. 5. To propose the methodology for pension funds’ ranking based on dynamic risk and efficiency model developed. The achievement of results will be judged on validation period for various risk type pension funds (PF) working in Lithuania. It is expected that results could be applied in Lithuanian PF market. The ideas implemented in the model will be applicable for other pension systems with few pillars in other countries (e.g. Slovakia, Italy) as well. For more details please contact the supervisor.
|
| Hierarchical heuristic algorithms for combinatorial optimization problems |
prof. dr. Alfonsas MISEVIČIUS |
state non-funded |
|
Machine learning and information integration research for disease prediction and health care
|
prof. dr. Robertas ALZBUTAS |
state non-funded |
Research Topic Summary.
The objective of the research is creation of methodology and test calculations based on machine learning and information integration for the most accurate disease prediction and health care. Tasks: 1. Review and compare possibilities of information integration methods and algorithms as well as application of related software and machine learning methods intended for health care. 2. Define information integration and smart systems accuracy criteria and their evaluation procedures in the context of data relevant for health care. 3. Develop and demonstrate the information integration tools which usage increase the accuracy and / or reduce the risk of incorrect decisions. 4. Perform the pilot studies of smart systems intended for health care and create methodology for effective application of these systems. For further information please contact the supervisor of the topic.
|
