Thesis supervisor: Bálint Gyires-Tóth
Location of studies (in Hungarian): Department of Telecommunications and Media Informatics Abbreviation of location of studies: TMIT
Description of the research topic:
Self-supervised learning (SSL) methods are a solution for datasets with no or only a few manual annotations. In SSL the input and the corresponding supervisory target samples are formed from unlabeled data. SSL methods utilize different approaches, including contrastive learning, reconstruction and compression methods. Contrastive learning is a discriminative SSL method, that learns to classify, whether two measurements from a stochastic measurement function comes from a process with the same distribution. For instance, predicting the assumed correlation of the data, based on its spatial or temporal relations, is such an approach. Reconstruction SSL methods form the target variables from the data itself, and a mapping function is learned by the model that restores the target from the input. Compression SSL methods are a special kind of reconstruction, where the same data is used as input and output, and the mapping function has fewer parameters than the dimensions of the original space. Self-supervised methods have the perspective of learning representation when labels are not available; however, these methods are not general, are data-intensive, the efficiency is not consistent among datasets, and scalability is also limited.
Though SSL methods have boosted natural language processing, and the results in common computer vision tasks are also promising, addressing the described shortcomings and applying SSL for complex data is still a challenge. Such complex data are graph data structure, multidimensional sequences, heterogeneous data, or images with diverse content, for instance.
The goal of this research is to elaborate novel SSL and deep learning methods for complex data structures. Besides performance improvements (in terms of accuracy, speed and/or robustness) the research should also focus on the stability and scalability of the elaborated methods. The effectiveness of the proposed methods must be proven, at least in one application scenario. Such an application scenario can be (1) graph neural networks, (2) biotechnology, (3) medical imaging, (4) robust handwriting recognition, etc.
The research can be performed in English and Hungarian. For training the models public and private databases and high-performance GPUs are available.
The possible research tasks of the PhD candidate are the following:
• Overview of the related scientific papers, including fundamentals of deep learning, novel results of self-supervised learning and domain-specific scientific works.
• Design and implement baseline and existing state-of-the-art complex data modeling approaches for benchmark.
• Elaborate novel deep learning and/or self-supervised learning models for complex data.
• Investigate the performance of the proposed models in terms of accuracy, speed, robustness, stability and/or scalability.
• Demonstrate the effectiveness of the proposed methods, at least in one application scenario.
• Objective and (if possible) subjective evaluation of the proposed methods.
Required language skills: English or Hungarian Further requirements: Programming and mathematical skills.
Number of students who can be accepted: 2
Deadline for application: 2021-06-14
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
