↓Skip to main content

Home

Eliezer de Souza da Silva #

Assistant Professor, Department of Informatics Engineering, University of Coimbra.

I develop methods in probabilistic AI and machine learning, with emphasis on Bayesian modeling, amortized inference, and principled uncertainty quantification. My current work combines theory and practice across GFlowNets, prior predictive modeling, and trustworthy decision-making systems.

Publications ยท Contact

Research overview #

Amortized Inference and GFlowNets #

I study training objectives, stability, and expressiveness in GFlowNets and related amortized inference models, including divergence-based learning and assessment metrics.

Prior Predictive and Bayesian Modeling #

I work on prior specification and identifiability in matrix/tensor factorization and broader Bayesian models, using prior predictive criteria to make modeling choices explicit and robust.

Sequential Decision-Making, Causality, and Trustworthy AI #

I investigate uncertainty-aware methods for sequential decision-making and causal discovery, including human-in-the-loop settings and reliability-aware evaluation. I am interested also in unlearning as a task for trustworthy AI.

Applications #

I collaborate on applications in recommender systems, Math reasoning and evaluation, and scientific ML tasks where uncertainty and calibration matter in deployment.

Selected publications #

  • On the Identifiability of Tensor Ranks via Prior Predictive Matching (AISTATS 2026). Extends prior predictive ideas to tensor rank identifiability with closed-form characterizations. Paper
  • MATH-PT: A Math Reasoning Benchmark for European and Brazilian Portuguese (PROPOR 2026). A benchmark of 1,729 native Portuguese math problems for evaluating mathematical reasoning in LLMs. Paper
  • When do GFlowNets Learn the Right Distribution? (ICLR 2025, Spotlight). Clarifies when balance violations harm sampling correctness and introduces a better assessment perspective. Paper
  • On Divergence Measures for Training GFlowNets (NeurIPS 2024). Unifies GFlowNet training via divergence minimization and develops lower-variance gradient estimators. Paper
  • Prior Specification for Bayesian Matrix Factorization via Prior Predictive Matching (JMLR 2023). Provides a practical framework for selecting priors and latent rank using prior predictive objectives. Paper
  • Human-in-the-Loop Causal Discovery under Latent Confounding using Ancestral GFlowNets (Preprint 2023). Combines uncertainty-aware causal graph sampling with expert feedback loops. Paper

Ongoing directions #

  • Exploration-exploitation and quality-diversity in generative models and decision-making (towards deliberative reasoning).
  • Continual adaptation and unlearning in modern foundation-model pipelines.
  • Prior predictive identifiable aspects of tensor factorization models, with connections to larger family of models.

Collaboration and contact #

I am open to research collaborations, student supervision (BSc/MSc/PhD), and invited talks on probabilistic AI, Bayesian ML, and generative inference.

Please reach out via email at eliezer.silva@uc.pt