Thesis supervisor: László Gulyás
Location of studies (in Hungarian): Faculty of Informatics, Eötvös Loránd University Abbreviation of location of studies: ELTE
Description of the research topic:
In the last two decades Artificial Intelligence and Deep Learning (DL) technologies underwent a significant evolution and became one of the most researched topics. Neural Networks (NN) and DL thus became an integral part of many industrial applications, including safety critical domains, such as robotics or healthcare. This industrial adoption of the technology drew attention to the pitfalls of black box models and invoked (1) a set of explainability techniques that provide insights of the inner workings of the model, and (2) validation techniques that ensure safety and robustness.
Autonomous driving is an important domain for industrial DL applications, as in many areas the problem complexity does not allow for rule-based solutions. DL already dominates the perception part of the self-driving technology stack, mostly via Convolutional Neural Networks (CNNs). Recently, graph-based representations and Graph Neural Networks (GNNs) opened a potentially new area of DL in self-driving, in scene understanding and in behavior planning and prediction. GNNs also dominate the current scientific literature due to their natural representation of a wide variety of topics. They are powerful in learning interactions and connections that are present in the data, but they are not inherently explainable. As some of the largest automotive players are already working on industrial standards for validating CNNs, GNNs might be the next step in the downstream validation process and need the same attention in the upcoming years.
The aim of the dissertation is to investigate the existing techniques for DL model validation, and to adapt them to the GNN use case. In this process, both publicly available data sources of many domains (traffic datasets, chemical interactions, molecule structure, social networks, etc.), and industry data can be used. The goal of the dissertation is a validation strategy for GNNs in the self-driving domain, focusing on model explainability.
Required language skills: English Recommended language skills (in Hungarian): Hungarian Further requirements: Python, basic data processing libraries (Numpy, Pandas), Deep Learning frameworks (TensorFlow-Keras or Pytorch), Graph processing libraries (Pytorch Geometric, NetworkX, etc.)
Deep learning, Graph Neural Networks
Number of students who can be accepted: 1
Deadline for application: 2023-05-31
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
