Knowing what models don't know under data scarcity.
Reliable Model Adaptation
Reusing foundation models for new tasks effectively.
Reliable Model Evaluation
Certifying performance and safety claims of deployed models.
Applications
Real-world use cases where ML reliability matters.
Click a card to filter publications by topic.
Reliable Model Prediction.
Models trained on small datasets tend to be quietly overconfident, but commit to one answer
when several are equally plausible. I work on probabilistic meta-learning approaches that retain
meaningful uncertainty in their predictions, giving downstream users an honest signal of
when to rely on a model and when to seek additional evidence.
Reliable Model Adaptation.
Repurposing a foundation model for a new task usually involves ad-hoc choices (a hand-picked prompt,
an arbitrary label mapping) with little understanding of why one configuration works and another does not.
I study Neural Network Reprogrammability, hoping to reveal and exploit the latent
compatibility between a frozen model and a target task, replacing guesswork with informed design.
Reliable Model Evaluation.
A model in production often carries implicit claims about robustness, data provenance,
and evaluation coverage that nobody has actually checked. I build tools that make these
claims auditable: benchmarks that expose hidden trade-offs in model compression, statistical
frameworks that detect unauthorised data usage, and evaluation methods that deliver
trustworthy conclusions without exhaustive testing.
Applications.
The methods above are motivated by, and tested in, settings where getting reliability wrong has
real cost. These range from recommendation systems that must handle cold-start users, to brain-computer
interfaces that demand robust signal decoding, to trajectory forecasting where overconfident predictions
can compromise safety. Each application sharpens the methods as much as it benefits from them.