Thesis supervisor: László Gulyás
Location of studies (in Hungarian): Eötvös Loránd University, Faculty of Informatics Abbreviation of location of studies: ELTE
Description of the research topic:
Modern artificial intelligence (AI) methods focus on mathematical optimization and generalization performance to make learning processes more widely available and useful. This requires smart solutions and a large amount of data. However, the use of data for learning processes is limited by existing legislation and EU regulations, such as the General Data Protection Regulation (GDPR) and the AI Act, which help protect individuals' personal data by limiting how it can be used in AI-based systems. For intelligent systems on the market and in continuous use, data limitation is a continual challenge, because the legitimacy of data used during training may be challenged later. These problems are being addressed by machine unlearning mechanisms. At the model level, machine learning can change predictions and eliminate data points and data subsets, while still maintaining generalization performance.
The field of machine unlearning is rapidly growing as evidenced by the increasing number of articles exploring the forgetting capabilities of models placed in static environments. These include various methods such as differential privacy standards, data perturbation methods, inverse data generation methods, data poisoning, and reinforcement forgetting techniques. Although the performance of these techniques has significantly improved over the years, there is still a lack of uniform validation capabilities for unlearning algorithms in the literature. One of the shortcomings of machine unlearning is its limited applicability in dynamic environments, where there are natural or artificial shifts in system-level variables. These shifts can be due to changes in the environment, in the goal(s) of the system, alignment, or system trends. Without a uniform forgetting standard with universal applicability, current AI-based methods are not equipped to handle potential unlearning requirements during their lifecycles.
The primary goal of this PhD research is to investigate the mechanisms of unlearning in models that operate in dynamic environments, and to explore new methods that can improve decision-making processes. In addition to developing new approaches, another key objective is to create guidelines for creating new models, with a focus on preparing them to handle unlearning. Dynamic environments are highly complex, and can pose additional validation challenges, so another topic expected to be an important aspect of the research project is the development of uniform and measurable metrics. Ultimately, the research aims to enhance the applicability of machine unlearning, thus contributing to explainability, legality and fairness of future models.
Required language skills: angol Recommended language skills (in Hungarian): B1 Further requirements: MSc in one of the following fields: informatics, mathematics, computer science or related areas.
Fluent in English.
Number of students who can be accepted: 1
Deadline for application: 2024-05-31
2024. IV. 17. ODT ülés Az ODT következő ülésére 2024. június 14-én, pénteken 10.00 órakor kerül sor a Semmelweis Egyetem Szenátusi termében (Bp. Üllői út 26. I. emelet).
Login
