Trustworthy and Interpretable AI

Before foundation models can be deployed in real settings, people must be able to rely on them. We study how to make them robust to distribution and domain shifts, fair across groups, safe, and transparent in how they reach decisions. This includes analysing the structure of vision-language representation spaces, such as the compositionality of visual embeddings. It also includes data-free interpretability methods that explain model behaviour from the weights alone.

Benchmarking of Foundation Models

As multimodal models become more capable, measuring what they can and cannot do becomes a research problem in its own right. We design benchmarks and evaluation protocols that test models beyond standard accuracy: on compositional reasoning, open-ended recognition and the semantic similarity of videos. We also develop automatic, programmatic pipelines that probe large multimodal models with iterative experiments.

Video Understanding

Video brings challenges of its own: long temporal context, scarce annotations and large shifts between domains. We work on video understanding in low-supervision regimes, such as zero-shot and test-time temporal action localisation, source-free video domain adaptation and training-free video anomaly detection. Several of these combine vision-language models with large language models.

Efficient Adaptation and Continual Learning of Multimodal Models

Retraining large models whenever data or requirements change is expensive and often impractical. We develop methods that let vision-language and multimodal models be adapted, specialised and continually updated cheaply. Our work covers test-time adaptation, prompt tuning, including in federated settings, and task-agnostic pruning and compression. It also covers data-free continual learning, so that models keep learning new tasks without forgetting earlier ones.