Research Areas

Our work is motivated by a simple problem: high-performing AI systems are often trained and evaluated with imperfect information. We develop methods that make learning, evaluation, and reasoning more reliable in those conditions.

For several years the lab worked at the model-construction stage, on learning from supervision that is noisy, coarse, or drawn from outside the label space. As foundation models have made intervention at that stage progressively impractical — most deployed models are called rather than built — our focus has moved to the inference stage: how to quantify the risk of a given prediction, and what can still be guaranteed when the evidence used to assess it is itself imperfect.

Learning with Imperfect Supervision

We design algorithms for noisy labels, coarse labels, open-set noise, and other forms of imperfect supervision in real-world visual datasets.

Noisy labels Open-set noise Representation learning

AI Safety and Robust Evaluation

We study certification, adversarial robustness, and statistically reliable evaluation for models and LLM-based systems under imperfect judges.

Model certification Adversarial attacks LLM evaluation

Foundation Models for Visual and Multimodal Data

We adapt foundation models such as CLIP-style representations to downstream settings where data quality, annotation quality, or domain fit is limited.

Computer vision Multimodal learning Healthcare AI

Representative Questions

When labels are unreliable, what can still be learned?

We analyze failure modes of idealized noise correction and build practical robust learning methods for noisy supervision.

How should we evaluate models when the evaluator is imperfect?

We work on robust statistical evaluation protocols for LLMs and foundation models when automatic or human judges are noisy.

Can safety claims be made precise?

We develop certification methods that provide formal or statistical guarantees under adversarial perturbations and distribution shifts.