Research

Research Interests

My main occupation is the development of Machine Learning methodologies aimed at solving particular problems other disciplines might present. In particular, I have extensively worked with both simulated and observed astronomical data sets. In the first case, I have contributed significantly to development of the public toolbox 1-DREAM. We present here 5 methodologies that gradually discover one-dimensional structures (filaments, streams, etc.). For each detected filament we recover an explicit parametric formulation, as a mapping from a latent abstract graph to its embedding in the data space. Thanks to this smooth mapping, we are able to walk along the tangent space of the streams and study their physical properties. At the same time, taking advantage of the probabilistic modelling over their skeleton, we can produce maps on the co-moving orthonormal coordinate frames along the manifold. This allows for detailed studies of the dynamics and hcemical properties of the filaments. The same methodology has been applied to both simulations of Jellyfish galaxies and the Large-Scale-Structure of the Universe (Cosmic Web).

Synthetic data sets for testing of 1-DREAM toolbox. Seven manifolds of different dimensionalities, each one being noisily sampled, are embedded in a noisy environment. The goal is to recover their true intrinsic dimensionality and model each manifold separately.
End result of the application of the full methodology to the synthetic data set. Each manifold is correctly recovered as a skeleton (black), and an isosurface for a given isovalue of the log-likelihood estimated from their individual models. Even non-orientable manifolds, such as the Mobius band, are recovered, thanks to the smooth mapping between latent abstract and embedded graphs.

During my years as a Research Fellow at UoB I have also worked on biomedical problems. In particular, multiple works have been performed on the study of metabolomic pathways in the early diagnosis of Cushing disease. Further studies on feature relevance in the distinction between Non-fat Adrenocortical Tumors (NFAT) and Mild-Autonomous-Cortisol-Secretion (MACS) are on the way and soon to be submitted.

I also had the chance to colaborate with epydemiologists at the Institute of Applied Health Research (@UoB) within the OPTIMAL project. The focus is here to describe multi-morbidity and develop a toolbox for clinicians and researches capable of guiding epidemiological studies on different sub-populations, distinct by demography. An additional project focuses on poli-pharmacy and how to prevent over-exposure to medications to patients with multiple long term cronic diseases.

Publications