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:
The use of deep neural networks for solving particular problems has proven to be extremely effective when a large amount of high-quality training data is available. Mathematical operations and gradient-based stochastic optimization are used in deep learning models in order to minimize or maximize an objective function. One of the major advantages of deep learning compared to other machine learning methods is its ability to automatically learn representations. Nevertheless, learning an optimal representation is not trivial.
Deep learning has recently demonstrated remarkable results for representation learning in natural language processing, resulting in models which can be applied in a wide variety of downstream tasks, and achieving state of the art. For automotive purposes, for example, object detection, semantic segmentation, and direct vehicle control, similar general computer vision models have not yet been developed. In the automotive industry, an ideal deep representation would be robust to distribution shifts (e.g., different environments, lighting conditions, etc.) and could be applied to a wide range of related downstream tasks, including object detection and semantic segmentation as well as classification. Furthermore, data- and energy-efficiency play a critical role both in training and inference of large deep learning models.
The goal of this Ph.D. research is to elaborate novel deep representation learning models and training algorithms capable of effectively and efficiently learning visual features that could be utilized in various downstream applications in the automotive industry.
The foreseen research tasks of the Ph.D. student are the following:
• Overview of related scientific papers, including the basics of neural networks and deep learning, representation learning for automotive applications, and data-efficient representation learning and inference.
• Design and implement baseline methods from the fields of supervised, semi-supervised, and/or self-supervised representation learning, focusing on automotive use-cases.
• Conduct research in deep representation learning algorithms to achieve improvements (in terms of accuracy, data-efficiency, task-efficiency, for instance) in multiple downstream tasks, compared to previous approaches.
• Elaborate novel deep representation learning methods for at least two automotive use-cases, e.g. object detection, semantic segmentation, vehicle control.
• Consider at least two of the following aspects / constraints in the proposed methods: interpretability, explainability, label quantity, label quality, data quantity, data diversity, inference latency and throughput.
• Demonstrate the effectiveness of the results in at least one application scenario.
The research can be conducted both in English and Hungarian.
Required language skills: angol Number of students who can be accepted: 1