Research

Instead of performing the embedding step of a chunk of a document after the chunks get calculated, record all token-level embeddings of a document and chunk post-embedding. In essence, you pool the token embeddings of the document according to the chunk structure, and this means you don't lose contextual level information from chunk to chunk.

Differential inference in time-varying parametric probabilistic models. Essentially, models that have changing structures for which we need to estimate the derivative of the parameter (with respect to time) rather than the parameter itself. Uses the time score function rather than typical score matching for such estimation.

Developed a data-driven solution to truncated probability density estimation by solving a constrained optimisation problem and minimising a modified kernelised Stein discrepancy. I showed theoretical guarantees for this estimator and ran numerical experiments demonstrating improvements over state-of-the-art methods.

Extended the theoretical approach of truncated score matching to the manifold case, using an application of Stokes' theorem to rederive the objective function. I constructed alternative weighting functions for a generic spherical boundary and demonstrated strong numerical performance.

Developed an estimator of truncated probability densities for the case where computing the normalising constant is infeasible, using score matching. Our estimator involves a weighting function derived from heuristics and analytical results, with demonstrable numerical advantages across a series of benchmark experiments and applications.

Modelled the relationship between extremes of rainfall and their spatial properties, for both gridded model output and point-level observations, using extreme value theory. By downscaling, I predicted rainfall extremes at a high resolution via a spatial generalised additive model.