Probabilistic machine learning · EURECOM

Simone Rossi

I am an Assistant Professor (Maître de Conférences) in the Data Science Department at EURECOM. We develop scalable methods at the intersection of Bayesian inference, probabilistic modeling, and deep learning.

We study how uncertainty, structure, and inductive bias can support reliable and data-efficient learning, with applications including autonomous driving.

Google Scholar GitHub ORCID LinkedIn Email

More about me

Research

Uncertainty quantification and Bayesian inference. Uncertainty quantification helps assess the reliability of predictions and make decisions from limited data. Bayesian inference offers one approach. Our recent work studies how predictive uncertainty scales with data, model capacity, and inference method. We also develop functional priors for encoding prior knowledge in neural networks.
In-context learning theory and applications. In-context learning raises questions about how a model adapts to a task from examples and when that adaptation generalizes. We study how the mean of the task prior and query initialization affect the correspondence between linear self-attention and a gradient-descent step, and how epistemic uncertainty relates to grokking. Recently, we have used in-context learning to build ALICE, which estimates mutual information from sample contexts.
Generative modeling and autonomous systems. Autonomous systems need models of their surroundings and of how other road users may move. We develop methods for joint trajectory prediction and controllable generation and for dynamic driving-scene reconstruction with DrivingVoxels.