Program

The event  will start in the afternoon of Monday 20th January 2025 around 2pm and end on Friday 24 January 2025 around 1pm. 

The schedule is reported below. 

Invited Talks

• Stefania Bellavia, University of Florence

Title: Adaptive Optimization Methods based on  Random Models

Abstract: We discuss trust region methods using random models and stochastic function estimates. The primary motivation for developing these methods is the need to solve optimization problems that arise in machine learning, which because of the enormous amounts of data involved in each computation of the function and gradient, usually require a stochastic optimization approach. The main issues that we address in this talk are the  adaptive choice of step-size and of the accuracy conditions that the models and  function values have to satisfy. In particular we assume that our models  satisfy some good quality conditions with some fixed probability, but can be arbitrarily bad otherwise. Both theoretical and computational results will be presented.

• Sandra Pieraccini, Polytechnic University of Turin

Title: Graph-Instructed Neural Networks: features and applications

      Abstract: In many applications involving complex problems, the availability of reliable surrogate models is a crucial issue. This is particularly true for many-query tasks, such as uncertainty quantification or optimization problems. To reduce the computational effort of such processes, recent literature suggests that neural networks (NN) can be a valuable tool for regression tasks. In various applications (such as transportation systems, epidemic modeling, and social interactions), network analysis plays a significant role, with graphs being central to these frameworks. Recently, the NN community has made key advancements by extending deep learning approaches to graphs through the introduction of graph neural networks and the more recent graph convolutional networks (GCNs). While GCNs have shown promising results in many applications, several challenges remain. In this talk, we introduce a novel layer designed for regression tasks on graphs, a framework where GCNs are not ideally suited and multi-layer perceptrons (MLPs) are often preferred. This new layer leverages the graph structure to enhance NN training (compared to MLPs) and enables the development of deep NNs with an approach that is scalable for large graphs. We will demonstrate the potential of these graph-instructed layers and present examples of successful applications.

• Silvia Villa, MaLGa & University of Genoa

Title: Implicit regularization for machine learning and inverse problems

Abstract: In many applications, the goal is to recover an unknown quantity of interest from noisy/random measurements. To achieve a satisfactory reconstruction, a priori information about the problem is crucial. The most common way to inject this information into the reconstruction method is by regularization, either explicit or implicit.  In this talk, I will focus on implicit regularization, that is based on the bias intrinsically induced by the method used to optimize the parameters involved. I will present different ways to implicitly induce a bias in the optimization technique and some theoretical recoveryresults. 

Poster session:

• Gianluca Audone, Polytechnic University of Turin

Title: Beyond linearity:  exploring the potential machine learning techniques for space weather

• Katsiaryna Bahamazava, Polytechnic University of Turin 

Title: Physics-based multi-class prediction of geo-effective events through machine learning

• Ivan Bioli, University of Pisa

Title: Preconditioned low-rank Riemannian optimization for symmetric positive definite linear matrix equations

• Davide Carrara, Polytechnic University of Milan

Title: Implicit neural field reconstruction on electro-anatomical maps from noisy scattered data

• Massimiliano Ghiotto, University of Pavia 

Title: HyperNOs: automated and parallel Hyperparameter Optimization library for Neural Operators

• Laura Girometti, University of Bologna 

Title: Unfolding strategy for automatic image decomposition

• Luca Pellegrini, University of Pavia

Title: Physics-Informed Neural Networks for ionic models in electrophysiology

• Andrea Sebastiani, University of Modena and Reggio Emilia 

Title: Accelerated Plug-and-Play methods for medical imaging

• Ilaria Trombini, University of Ferrara

Title: Variable metric proximal stochastic gradient methods with additional sampling