Sample-Efficient Active Learning from Human Feedback
Event details
| Date | 07.01.2026 |
| Hour | 11:15 › 12:00 |
| Speaker | Belen Martin-Urcelay, PhD Giorgia Tech, USA |
| Location | |
| Category | Conferences - Seminars |
| Event Language | English |
Abstract
Aligning machine learning systems with human intent requires querying experts, but collecting high-quality feedback at scale is often prohibitively expensive. This talk presents active learning frameworks designed to minimize sample complexity by exploiting structure in the data and the feedback mechanism.
The first part of the talk focuses on interactive search over hierarchies, where the goal is to locate a target node using reachability queries. We introduce a query selection algorithm that is robust to oracle noise, and we bound its sample complexity. The second part addresses classifier learning within pre-trained embedding spaces. While traditional methods rely on simple labeling, we demonstrate that leveraging the embedding geometry allows for richer forms of feedback, such as rankings. We analyze how these richer queries significantly reduce the human feedback cost. The talk concludes by pointing out future avenues for more efficient human-in-the-loop learning.
Biography
Belén Martín-Urcelay is a Ph.D. candidate at Georgia Tech, advised by Matthieu Bloch and Christopher Rozell. She earned her B.Sc. and M.Sc. in Telecommunications Engineering from Universidad de Navarra before joining Georgia Tech for her doctorate. Her research focuses on active learning, human feedback, and reinforcement learning, with an emphasis on leveraging knowledgeable teachers (whether human experts or powerful machine learning models) to enhance learning algorithm performance. She has collaborated internationally, including a research visit to ETH Zurich with Andreas Krause. Outside of research, she is passionate about teaching, mentoring students, and promoting women in STEM fields.
Aligning machine learning systems with human intent requires querying experts, but collecting high-quality feedback at scale is often prohibitively expensive. This talk presents active learning frameworks designed to minimize sample complexity by exploiting structure in the data and the feedback mechanism.
The first part of the talk focuses on interactive search over hierarchies, where the goal is to locate a target node using reachability queries. We introduce a query selection algorithm that is robust to oracle noise, and we bound its sample complexity. The second part addresses classifier learning within pre-trained embedding spaces. While traditional methods rely on simple labeling, we demonstrate that leveraging the embedding geometry allows for richer forms of feedback, such as rankings. We analyze how these richer queries significantly reduce the human feedback cost. The talk concludes by pointing out future avenues for more efficient human-in-the-loop learning.
Biography
Belén Martín-Urcelay is a Ph.D. candidate at Georgia Tech, advised by Matthieu Bloch and Christopher Rozell. She earned her B.Sc. and M.Sc. in Telecommunications Engineering from Universidad de Navarra before joining Georgia Tech for her doctorate. Her research focuses on active learning, human feedback, and reinforcement learning, with an emphasis on leveraging knowledgeable teachers (whether human experts or powerful machine learning models) to enhance learning algorithm performance. She has collaborated internationally, including a research visit to ETH Zurich with Andreas Krause. Outside of research, she is passionate about teaching, mentoring students, and promoting women in STEM fields.
Practical information
- General public
- Free