Gabriel Victorino Cardoso

Biography

Gabriel Victorino Cardoso is a researcher specializing in computational methods for statistical inference and generative modeling. His work focuses on developing advanced algorithms to solve inverse problems and Bayesian inference tasks, particularly in contexts where the underlying models rely on complex simulators. He explores innovative approaches such as simulation-based inference (SBI), score-based diffusion models, and sequential Monte Carlo methods, with applications in computational neuroscience and medical imaging, particularly for the analysis of electrocardiographic (ECG) signals. His contributions include algorithmic improvements to the efficiency and stability of estimators, as well as theoretical frameworks for reducing bias and controlling variance in stochastic approximations.

Publication(s)

2026
Diffusion posterior sampling for simulation-based inference in tall data settings
2025
Reconstructing ECG from indirect signals: a denoising diffusion approach DOI : 10.1098/rsta.2024.0330
2024
MONTE CARLO GUIDED DENOISING DIFFUSION MODELS FOR BAYESIAN LINEAR INVERSE PROBLEMS
2024
PARTICLE-BASED, RAPID INCREMENTAL SMOOTHER MEETS PARTICLE GIBBS DOI : 10.5705/ss.202020.0215
2024
Leveraging an ECG Beat Diffusion Model for Morphological Reconstruction from Indirect Signals
2023
Particle-based, Rapid Incremental Smoother Meets Particle Gibbs DOI : 10.5705/ss.202022.0215
2023
State and parameter learning with PARIS particle Gibbs
2022
A Patient-Specific Single Equivalent Dipole Model DOI : 10.22489/CinC.2022.260
2022
BR-SNIS: Bias Reduced Self-Normalized Importance Sampling

Teaching

Generative Models

Course Director

The goal of this course is to introduce generative models used for generating modalities other than language (particularly images, but not limited to them). The course will begin with an introduction to the various models used (Normalizing Flows, VAEs, GANs) and will quickly move on to diffusion models, which will be the focus of this course. Throughout the course, we will examine both the statistical properties of these models and their practical implementation and effectiveness. To introduce diffusion models, we will start by learning about the score and how we can use it to generate samples from certain distributions (via the Langevin sampler). We will then move on to the current approach to using the score to generate data, first in a discrete manner (Markov chains), which is subsequently extended to the continuous-time formalism (stochastic differential equations). This will allow us to investigate, in particular, the convergence theorems for these models. Subsequently, the course will address how diffusion models can be used for conditional generation (images, text, and others), covering the following topics in particular: Classifier-guided, Classifier-free, and Training-free approaches. Throughout the course, we will work in parallel with a toy data set (due to computing power limitations) where participants can implement the methods covered in class. The last hour of each class will be dedicated to notebooks, worksheets, and questions. The exam will cover diffusion models, the various mathematical formalisms, and their statistical properties.

Large-Scale Machine Learning and Data Mining

Course Director

The week is organized around three types of activities: Lectures (mornings), hands-on sessions (afternoons), and conferences and roundtable discussions (evenings)