Georgia Institute of Technology
Sample Efficient Knowledge Transmission in Interactive Machine Learning
Abstract
dc:description.abstractThis thesis investigates how to make knowledge transfer more efficient in learning settings where data acquisition is costly, expert feedback is uncertain, and observations are noisy. Learning can be substantially accelerated when learners receive guidance from experts, provided that the interaction is designed effectively. When the expert is a machine, we relax the assumptions of standard machine teaching algorithms by developing a greedy algorithm that allows for uncertainty about the learner while guaranteeing the same exponential speed-up as the omniscient baseline. When the expert is a human, we reduce annotation burden by designing information-rich query modalities and algorithms that exploit them effectively. The first human modality studied is natural language feedback in reinforcement learning, which yields strong reductions in sample complexity in experiments with human evaluators. Second, we design ranking and exemplar selection queries to learn classifiers in pre-existing embeddings spaces, where the additional problem structure also enables theoretical guarantees on sample complexity. Lastly, we introduce the first algorithm for interactive tree search that poses multiple-choice queries while explicitly accounting for human noise. Together, these contributions demonstrate that knowledge transfer can be accelerated by designing learner-teacher interactions that account for uncertainty and extract more information per interaction.
Degree
thesis:*- Name thesis:degree_name
- Electrical and Computer Engineering, PhD
- Department thesis:degree_department
- Electrical and Computer Engineering
- Grantor
- Georgia Institute of Technology
- Year dc:date.issued
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Martin Urcelay, Belen
- Advisors dc:contributor.advisor
-
- Bloch, Matthieu R.
- Rozell, Christopher J.
- Committee members dc:contributor.committeemember
-
- Davenport, Mark
- Pananjady Martin, Ashwin
- Romberg, Justin
- Xie, Yao
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/1853/82160