Giosuè Migliorini

I am a fourth year Ph.D. Student in Statistics at UC Irvine, advised by Prof. Padhraic Smyth.

I had the privilege to spend the summer of 2025 working at Flagship Pioneering as an AI research intern, working on mechanistic interpretability for protein co-folding models. Previously, I spent the summer of 2023 working at Los Alamos National Laboratory, where I applied Bayesian deep learning to scientific problems.

Before moving to the U.S., I lived and studied in Italy, where I completed my Master's in Data Science at Bocconi University. You can check out my Master's thesis on computational methods for Bayesian variable selection at this link, completed under the supervision of Prof. Giacomo Zanella at Bocconi's Bayes Lab.

Email  /  GitHub  /  Google Scholar  /  LinkedIn  /  CV

Research

My research centers on probabilistic inference for latent variable models, from modern deep generative methods (flows, diffusion) to latent feature extraction for interpretability in large foundation models. My current emphasis is on spatiotemporal latent variable models for continuous-time dynamics and on making them scalable to high-dimensional state spaces.

Efficient Inference for Coupled Hidden Markov Models in Continuous Time and Discrete Space

We introduce latent interacting particle systems, a class of hidden Markov models whose dynamics is described by interacting continuous-time Markov chains. Our inference method involves estimating look-ahead functions (twist potentials) that anticipate future information, and incorporating them into a twisted sequential Monte Carlo scheme.

PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Folding

Variational Inference for Interacting Particle Systems with Discrete Latent States

We explore variational inference methods for irregularly observed continuous-time Markov chains in high-dimensional settings.

A Generative Diffusion Model for Probabilistic Ensembles of Precipitation Maps Conditioned on Multisensor Satellite Observations

Residual diffusion model for probabilistic predictions of precipitation maps from infrared and microwave radiometric measurements.

Dynamic Conditional Optimal Transport through Simulation-Free Flows

We introduce a dynamic formulation of conditional optimal transport, and incorporate it into the flow matching framework. We showcase the method on challenging Bayesian inverse problems.

Functional Flow Matching

Theory and methodology for flow matching in infinite-dimensional spaces, for training and sampling at arbitrary resolutions.

Teaching

• Statistical Methods II (STATS 211P - Teaching Assistant): weekly discussion sessions, office hours, and homework grading for students of the Masters of Data Science at UCI. Topics include generalized linear models, PCA, regularization, model selection, splines, mixed-effects models.
• Introduction to Probability and Statistics (STATS 120B - Reader): undergraduate-level class focusing on maximum likelihood and hypothesis testing.
• Introduction to Probability and Statistics for Computer Science (STATS 67 - Reader): undergraduate-level class on basic statistics and probability theory.

Education

• Ph.D. in Statistics - University of California, Irvine (2022 - present)
• M.Sc. in Data Science (summa cum laude) - Bocconi University (2020 - 2022)
• B.Sc. in Economics (summa cum laude) - University of Padova (2017 - 2020)

Design and source code from Jon Barron's website