Machine Learning for Scientific Data

Scientific observations often provide only indirect evidence of the physical or functional properties we want to study. My work uses machine learning to make those latent properties accessible while keeping the method tied to the structure of each scientific problem.

The representation and evaluation therefore vary by domain: graph models and explainability for functional connectomes, perceptual and saliency-based evaluation for visual brain decoding, and deep neural models for supernova light curves. Applications include identifying discriminative brain sub-systems, evaluating fMRI-based visual reconstructions, and characterising the progenitors of astrophysical transients from simulated and observed data.