Two-days Follow on ANMMLDL

Program

The event  will start in the afternoon of Wednesday 12 March 2025 around 2pm and end on Thursday 13 March 2025 around 3pm. 

The event will be held in room B1 at Palazzo Manfredini, Via Monsignore Bovelli 59, Ferrara.

The schedule is reported below.

Invited Long Talks

• Alessandro Benfenati, University of Milan

Title: Deep Learning Techniques for Imaging Problems - Deep Image Prior and Autoencoders for blind deconvolution, segmentation and tomography reconstruction

Abstract: Deep learning techniques are increasingly playing a significant role in addressing imaging problems across several scientific  fields, including Medicine, Astronomy, and Microscopy. We present two distinct approaches designed to tackle three different tasks: blind deconvolution, semantic segmentation, and tomography reconstruction. The first task involves reconstructing both the image and the unknown blur kernel; semantic segmentation focuses on classifying regions of an image based on color and/or shape, pixel by pixel, via Mumford-Shah functional. The deep learning technique used to solve these two problems is the Deep Image Prior (DIP), a recent approach that reparametrizes an image through the weights of a neural network. Tomography reconstruction is addressed using two distinct neural autoencoders, whose latent spaces are connected via a Bridge network. This approach resembles an SVD solver, thus offering a layer of interpretability of the entire process within an XAI framework.

• Tatiana Bubba, University of Ferrara

Title:  Deeply learned regularization for limited-angle tomography

Abstract: In recent years, limited angle CT has become a challenging testing ground for several theoretical and numerical studies, where both variational regularization and data-driven techniques have been investigated extensively. In this talk, I will present hybrid reconstruction frameworks that combine model-based regularization with data-driven deep learning by relaying on the interplay between sparse regularization theory, harmonic analysis and microlocal analysis. The underlying idea is to only learn the part that provably can not be handled by model-based methods, while applying the theoretically controllable sparse regularization technique to the remaining parts. The numerical results show that these approaches significantly surpass both pure model- and more data-based reconstruction methods.

Invited Short Talks

• Anna De Magistris, University of Campania "Luigi Vanvitelli"

Title: A line-search based SGD algorithm with Adaptive Importance Sampling.

• Ilaria Trombini, University of Ferrara

Title: Stochastic gradient methods with Additional Sampling.

• Mahsa Yousefi, University of Florence

Title: Fully stochastic trust-region methods with Barzilai-Borwein steplengths.

• Greta Malaspina, University of Florence

Title: A Variable-Dimension Sketching Strategy for Nonlinear Least-Squares.

• Gianluca Audone, Polytechnic of Turin

Title: Variably Scaled Kernels: Improving Interpolation Accuracy with Learned Scaling Functions.

• Andrea Sebastiani, University of Modena and Reggio-Emilia

Title: Self-Supervised Deep Equilibrium models: learning from limited data.

• Luca Pellegrini, University of Pavia

Title: Learning Ionic Model Dynamics Using Fourier Neural Operators.

• Massimiliano Ghiotto, University of Pavia

Title: HyperNOs: Automated and Parallel Library for Neural Operators Research.

• Davide Carrara, Polytechnic of Milan

Title: Implicit Neural Field Reconstruction on Complex Shapes from Scattered And Noisy Data.

• Caterina Millevoi, University of Padua

Title: PINN-Based Inverse Modelling for Hydro-Poromechanics.

• Katsiaryna Bahamazava, Polytechnic of Turin

Title: Optimizing Multi-Class Classification of Geo-Effective Events in Machine Learning Using a Beta Distribution-Based Loss Function.