Research program
Research
Our research focuses on interactive machine learning, with an emphasis on active learning and adaptive experimental design. We study how learning systems can actively choose what information to acquire, how to represent it, and how to use it to improve subsequent decisions. This leads to work in active learning, Bayesian optimization, reinforcement learning, and machine learning for scientific discovery.
A common theme is that data acquisition, representation learning, and decision-making should not be treated as separate stages. Instead, we study methods that adapt these components jointly as learning and interaction proceed.
Active Learning & Representation Learning
How can a learner choose its own training data? A long line of our work studies active learning and machine teaching with rigorous guarantees: near-optimal information acquisition under noise, batch-mode and deep Bayesian active learning, and formal models of teaching. More recently we focus on active representation learning — jointly learning representations and acquisition strategies so that each supports the other — and on learned acquisition functions that improve from experience rather than relying on fixed heuristics.
Representative publications
-
Near-optimal Bayesian Active Learning with Correlated and Noisy Tests
AISTATS 2017 · Oral presentation
Bayesian Optimization & Adaptive Experimental Design
Many scientific and engineering problems reduce to optimizing an expensive black-box function. We develop Bayesian optimization methods for the settings practitioners actually face — multi-fidelity observations, unknown constraints, multiple objectives, high-dimensional or structured domains, and inexact acquisition optimization — and study when and why these methods provably help.
Representative publications
-
Direct Regret Optimization in Bayesian Optimization
ICML ExAI 2025
Reinforcement Learning & Sequential Decision-Making
Beyond one-shot queries, we study agents that act over time. This includes robust policy improvement that blends imitation with exploration, model-based reinforcement learning under approximate Bayesian inference, agents that actively refine their own training curriculum, and continual learning.
Representative publications
-
Model-based Policy Optimization under Approximate Bayesian Inference
AISTATS 2024 · Oral presentation
-
Blending Imitation and Reinforcement Learning for Robust Policy Improvement
ICLR 2024 · Spotlight presentation
Machine Learning for Scientific Discovery
With collaborators in physics, chemistry, and materials science, we put these methods into real experimental loops: reinforcement-learning-based trigger systems at the Large Hadron Collider, a self-driving thin-film deposition system that makes sample-specific decisions during synthesis, machine-learning-driven electrocatalyst discovery, and adaptive mesh optimization for accelerating PDE surrogates.
Representative publications
-
Learning to Trigger: Reinforcement Learning at the Large Hadron Collider
ICML AI4Physics 2026 · Best Paper Award
-
A Self-driving Physical Vapor Deposition System Making Sample-specific Decisions on the Fly
npj Comput. Mater. 2025
Current and emerging directions
We are also extending these ideas to foundation models, including questions around data quality and scaling laws, preference learning and alignment, and the optimization of agentic workflows.
Representative publications