Our research connects foundational machine learning with systems that work reliably in the world. Explore four directions and the projects through which we test new ideas.
Representation Learning
Learning compact, transferable representations for data, models, and retrieval systems.
Representation learning as mutual information maximization
A first-principles view that derives two families of self-supervised learning methods from a common information-theoretic objective.
Read the visual explainer →Learning useful representations from randomness
Random frozen projectors replace handcrafted augmentations, enabling one learning principle across images, time series, and tables.
Read the visual explainer →Human Preferences
Understanding how intelligent systems can reflect changing goals, needs, and contexts.
Tell the recommender what you want right now
Natural-language requests become sparse parameter masks that steer a frozen sequential recommender in real time.
Read the visual explainer →When a recommendation is almost right
Say what is wrong—“not tacos,” “less smoky,” “not pop”—and let the next suggestion respond without forgetting your broader taste.
Read the visual explainer →Trustworthy AI
Building uncertainty-aware systems whose outputs can be checked, explained, and acted upon.
From reliable RAG answers to self-correcting agents
How calibrated prediction sets can filter unsupported claims, localize errors in agent trajectories, and support principled rollback.
Read the visual explainer →Minimally invasive machine unlearning
A learned posterior controller identifies a sparse parameter mask for each forget request—without retraining or request-time optimization.
Read the visual explainer →Explain a prediction with the data that taught it
Kernelized Stein Discrepancy retrieves the training examples that provide model-specific support for an individual prediction.
Read the visual explainer →Make an edit fire only when it should
A contrastively learned projection space separates true paraphrases from lexically similar but unrelated prompts.
Read the visual explainer →Remove data without losing what the model learned
PUMA subtracts a marked example’s distinctive influence, then sparsely reweights retained data to preserve model performance.
Read the visual explainer →Can We Trust AI?
An interactive lecture for law students on legal search, generative AI, agents, verification, and accountability.
Open the interactive slides →Agentic Intelligence
Agents that coordinate, use tools, remember context, and recover when execution goes wrong.
From reliable RAG answers to self-correcting agents
Calibrated prediction sets localize decisive errors in multi-step agent trajectories and identify where execution should roll back and restart.
Read the visual explainer →
