Improving Semantic Anomaly Detection through Prompt Tuning of LLMs

Institut

Professur für autonome Fahrzeugsysteme

Typ

Semesterarbeit / Masterarbeit /

Inhalt

 

Beschreibung

Background: As the complexity and capabilities of robots expand, navigating through diverse and intricate environments, they increasingly face the risk of encountering unforeseen or anomalous situations that could lead to system failures. Notable instances include Tesla vehicles experiencing unexpected behaviors, such as autopilot deactivation when encountering traffic lights on moving trucks or unintended braking in response to billboard images of stop signs. These incidents highlight not the failure of individual components, but rather a broader issue with the systems' semantic understanding capabilities. Such situations, termed "semantic anomalies," present challenges that are easily resolved by humans but require sophisticated reasoning from automated systems.

Research Focus: This project aims to explore the potential of leveraging large language models (LLMs) for their extensive contextual comprehension and reasoning abilities to enhance the detection of semantic anomalies within vision-based systems. Through the development of a monitoring framework, this research will apply to autonomous driving policies, seeking to align the anomaly detection capabilities of LLMs with human judgment.

Objective: The successful candidate will contribute to pioneering research on refining the detection of semantic anomalies by employing prompt tuning techniques with foundation models - note that prompt tuning differs from prompt engineering. This involves integrating open-source Image-to-Text models with open-source LLMs to conduct experiments in the CARLA simulation environment, analyzing the effectiveness of the LLM-based monitoring in various scenarios. Afterwards, the LLM-based monitor shall be improved by the use of prompt tuning. Comparisons should be made between both approaches, engaging in comprehensive discussions to identify both the strengths and potential improvements of the latter.

Offer: We invite applications for a master's student position at our university, where you will have the opportunity to work on cutting-edge research aimed at advancing the field of semantic anomaly detection. This project not only offers the chance to contribute to groundbreaking work but also to explore the further potential of foundation models in enhancing system reliability and reasoning capabilities in complex environments.

 

Your Benefits (What we offer):

• Fascinating and exciting future-oriented research field
• Working with a state-of-the-art software development
• Possibility of publication as a scientific paper if suitable

Voraussetzungen

Requirements (What you should bring with you):

• High degree of initiative and a creative mindset
• Excellent English and programming (Python or C++) skills
• Foundational Knowledge in Machine Learning
• Initial experience with autonomous vehicles is an advantage
• Experience in software development is an advantage (Git, Ubuntu)

Tags

AVS Brusnicki

Möglicher Beginn

sofort

Kontakt

Roberto Brusnicki
roberto.brusnicki@tum.de