18.03.2025
Continual Model Adaptation: Methods & Applications
Abstract
Machine learning is becoming ubiquitous, influencing critical decisions across multiple sectors. However, most ML systems are one-model-fits-all solutions that are trained once with occasional updates, limiting their adaptability and performance over time. Continual ML approaches provide an adaptable solution that can handle distribution shifts, cater to rare events, and maintain robustness against adversarial attacks.
This talk explores the methodologies and applications of continual learning, emphasizing its importance in various domains such as automotive, robotics, and natural language processing (NLP). In the automotive domain, we explore incremental learning approaches for adapting gesture recognition systems to individual drivers, enhancing human-vehicle interaction. In robotics, we combine reinforcement and imitation learning to develop a curriculum learning approach to create a surgeon-in-the-loop ophthalmic robotic apprentice. In the NLP domain, we teach LLM agents to learn from their mistakes, adapt to contextually valid changes, and protect against privacy, security, and social engineering attacks.
Bio
Amr Gomaa is a final-year Ph.D. candidate in Human-centered AI & Applied ML at the German Research Center for Artificial Intelligence (DFKI). He is also a Visiting Researcher at the University of Cambridge, focusing on feasible ML (specifically LLMs) solutions for HCI design processes and interfaces, and ensuring the robustness of contextual LLM agents in collaboration with Microsoft Cambridge. His work focuses on developing adaptable machine learning systems, including incremental learning, reinforcement learning, and imitation learning, applied to domains such as automotive, robotics, health, and LLM agents.
His work is published in top-tier conferences such as NeurIPS, IROS, and IUI. Amr has led multiple research projects in collaboration with industry partners such as BMW and Zeiss, and secured multiple research grants for several projects such as "Secure Language Models for Knowledge Management" and "Hybrid Reinforcement Learning and Imitation Learnin
More: amrgomaaelhady.github.io
Photo provided by speaker
Machine learning is becoming ubiquitous, influencing critical decisions across multiple sectors. However, most ML systems are one-model-fits-all solutions that are trained once with occasional updates, limiting their adaptability and performance over time. Continual ML approaches provide an adaptable solution that can handle distribution shifts, cater to rare events, and maintain robustness against adversarial attacks.
This talk explores the methodologies and applications of continual learning, emphasizing its importance in various domains such as automotive, robotics, and natural language processing (NLP). In the automotive domain, we explore incremental learning approaches for adapting gesture recognition systems to individual drivers, enhancing human-vehicle interaction. In robotics, we combine reinforcement and imitation learning to develop a curriculum learning approach to create a surgeon-in-the-loop ophthalmic robotic apprentice. In the NLP domain, we teach LLM agents to learn from their mistakes, adapt to contextually valid changes, and protect against privacy, security, and social engineering attacks.
Bio
Amr Gomaa is a final-year Ph.D. candidate in Human-centered AI & Applied ML at the German Research Center for Artificial Intelligence (DFKI). He is also a Visiting Researcher at the University of Cambridge, focusing on feasible ML (specifically LLMs) solutions for HCI design processes and interfaces, and ensuring the robustness of contextual LLM agents in collaboration with Microsoft Cambridge. His work focuses on developing adaptable machine learning systems, including incremental learning, reinforcement learning, and imitation learning, applied to domains such as automotive, robotics, health, and LLM agents.
His work is published in top-tier conferences such as NeurIPS, IROS, and IUI. Amr has led multiple research projects in collaboration with industry partners such as BMW and Zeiss, and secured multiple research grants for several projects such as "Secure Language Models for Knowledge Management" and "Hybrid Reinforcement Learning and Imitation Learnin
More: amrgomaaelhady.github.io
Photo provided by speaker