01
Domain Adaptation
Adapting models from a labeled source domain to a shifted target domain.
Research
Reliable learning under distribution shift.
I study adaptive and deployment-aware deep learning methods for models that must remain dependable when deployed in environments different from their training data.
Research Focus
Core adaptation settings.
02
Source-Free Domain Adaptation
Adapting target models when the original source data is unavailable.
03
Domain Generalization
Training on multiple source domains to perform well on unseen target domains.
Research agenda
From source-free adaptation to clinical robustness.
My work is motivated by practical deployment constraints: source data may be unavailable due to privacy, proprietary ownership, or storage limitations, yet deployed systems still need to adapt to new domains.
I develop methods for source-free and black-box adaptation, domain-aware initialization and label calibration. These approaches are especially relevant in medical imaging, where scanner, hospital, device, and population shifts can sharply reduce model reliability.
Domain Adaptation and Generalization
Designing methods that adapt to target distributions or generalize to unseen environments while minimizing assumptions about source-domain access.
Source-Free and Black-Box Adaptation
Building deployment-aware algorithms for settings where source data, gradients, or model internals are inaccessible.
Medical Image Analysis
Applying robust learning to clinical tasks such as diabetic retinopathy grading, where subtle severity levels and domain shifts make reliability essential.
Benchmark Domain Shifts
Visual shifts across datasets, sensors, and clinical severity.
