Book of Abstracts

3rd Workshop of UMI Group

Department of Mathematics – University of Bari “Aldo Moro”

https://umi-math4aiml2025.uniba.it/
[email protected]
Contents
About Workshop 2

Plenary Speakers 3

Keynotes Speakers 8

Contributed Talks 17

Industry Talks 67

Posters 70
About MATH4AIML-2025
The MiδAs research group of the Department of Mathematics of the University of Bari Aldo Moro,
in collaboration with the Italian Mathematical Union (UMI) research group on "Mathematics
for Artificial Intelligence and Machine Learning”, organized at the Department of Mathematics,
the third edition of the Workshop Mathematics for Artificial Intelligence and Machine Learning
– MATH4AIML2025.
The main aim of the Mathematics for Artificial Intelligence and Machine Learning meetings
is to provide a platform for early-career researchers working on topics within the broad research
interests to present their work and network with peers and future collaborators.
The first two editions of the workshop, held respectively at the Polytechnic University of Turin
(in November 2022) and at the Bocconi University of Milan (in January 2024), were attended by
a large audience of researchers, also from outside the UMI group, and proved to be a platform
for sharing the work of many young researchers on emerging issues and mathematical aspects of
Artificial Intelligence, Machine Learning, and Optimization.
We have the honor of hosting the third edition of this workshop at UniBA, in the year of its
first centenary, as a corollary of scientific and cultural events organized in 2024. The three-day
event will be attended by more than 160 people, including young PhD students, researchers, and
senior members of the scientific community, and will feature three plenary lectures, eight keynote
speakers, two parallel sessions of contributed presentations and posters by young PhD students
and researchers.
In addition, this third edition of MATH4AIML will see the organization of a round table
bringing together the academic, corporate, and research worlds to discuss and explore the in-
teractions between MATH, ML and AI in addressing challenges and innovative applications for
industry and science.
Finally, there will also be opportunities to exchange ideas and opinions: all participants are
invited to take advantage of the social opportunities offered by the coffee breaks, lunch breaks,
and poster sessions.
We would like to thank the supporters of this edition, whose help was essential for the
organization of the workshop. In particular, we thank the industrial partners Pirelli S.p.A and
Planetek Italia S.r.l, for supporting this workshop, University of Bari Aldo Moro, ERC Seeds
Uniba project “Biomes Data Integration with Low-Rank Models” (CUP H93C23000720001) and
Piano Nazionale di Ripresa e Resilienza (PNRR), Missione 4 “Istruzione e Ricerca”-Componente
C2 Investimento 1.1, “Fondo per il Programma Nazionale di Ricerca e Progetti di Rilevante
Interesse Nazionale”, Progetto PRIN-2022 PNRR, P2022BLN38, Computational approaches for
the integration of multi-omics data. CUP: H53D23008870001.

The organizing committee

Nicoletta Del Buono
Flavia Esposito

Supported by

2
Plenary Speakers
• Claudia Angelini (page 4)
From single omics dataset to multi-omics and multi-datasets integration through a statistical
learning perspective and beyond.
• Tommaso Di Noia (page 6)
Current and future Trends in Recommender Systems
• Yurii Nesterov (page 7)
Optimization, the philosophical background of artificial intelligence

3
From single omics dataset to multi-omics and multi-datasets
integration through a statistical learning perspective and
beyond
Claudia Angelini
Istituto per le Applicazioni del Calcolo "M. Picone", Napoli, Italy
[email protected]

The widespread availability of high-throughput instruments for collecting omics data has
opened new avenues in personalized medicines, disease etiology understanding, and bio-marker
discoveries. However, analyzing this data presents several challenges, including high dimensional-
ity, distribution heterogeneity, and elevated noise levels. Various statistical and machine-learning
methods have been proposed to address these issues in contexts such as classification, cluster-
ing, survival analysis, and network inference. Recently, data collection efforts have evolved from
focusing on a single omics dataset (e.g., gene expression) to gathering multiple datasets from
different individuals on specific omics or datasets encompassing multiple omics from the same
individuals (e.g., gene expression, methylation, and gene structural variants). The availability
of such a large amount of data, including single-cell resolution data, can enhance the accuracy
of predictions when combined with appropriate computational approaches, to cite only a few
examples.
This work first provides an overview of our recent methods for analyzing single omics, such
as gene expression, within the context of survival analysis [1, 2]. Subsequently, we discuss how
such statistical methods can be generalized to accommodate scenarios with multiple datasets
or multiple omics. Therefore, we will present our recent extension of the cooperative learning
approach [3] to survival analysis and our latest methods for network inference, such as [4, 5].
Finally, we provide insights into how artificial intelligence methods can further move steps ahead
in extracting valuable knowledge and improving performance.
Acknowledgments This work is part of an extended collaboration with several colleagues and is
partially supported by the PRIN 2022 PNRR P2022BLN38 project, “Computational approaches
for the integration of multi-omics data” funded by European Union - Next Generation EU, CUP
B53D23027810001.

References
[1] C. Angelini, D. De Canditiis, I. De Feis, A. Iuliano A Network-Constrain Weibull AFT Model for
Biomarkers Discovery, Biometrical Journal 2024, 66 (7), e202300272
[2] A. Iuliano, A. Occhipinti, C. Angelini, I. De Feis, P. Liò COSMONET: An R Package for Survival
Analysis Using Screening-Network Methods , Mathematics 2021, 9, 3262
[3] D.Y. Ding, S. Li, B. Narasimhan, R. Tibshirani Cooperative learning for multiview analysis, Pro-
ceedings of the National Academy of Sciences 119 (38), e2202113119
[4] C. Angelini, D. De Canditiis, A. Plaksienko Jewel 2.0: An Improved Joint Estimation Method for
Multiple Gaussian Graphical Models , Mathematics 2022, 10, 21, 3983
[5] V. Policastro, M. Magnani, C. Angelini, A. Carissimo INet for network integration, Computational
Statistics, 2024, 1-23
Short Bio Dr. Claudia Angelini graduated in Mathematics in 1994 at the University of
Naples "Federico II"; where she also obtained her Ph.D. in Applied Mathematics and Computer
Science in 2002. Since 2001, she has worked as a permanent Researcher at the Institute for
Applied Calculus (IAC-CNR). She became a Senior researcher in 2019, and since January 2020,
she has held the position of Director of Research. Moreover, since July 2024, she has been acting
as head of the Naples branch of the Institute for Applied Calculus. Her main research activ-
ity is devoted to developing new statistical and machine learning methods to analyze complex

4
data, focusing on the analysis and integration of omics data. She has been the scientific coor-
dinator of the IAC-CNR research unit in several scientific and industrial projects at national
and international levels. She has co-authored more than 100 full articles in ISI peer-reviewed
international journals and numerous other international publications in conference proceedings
and book chapters. Over the years, she has supervised the research activities of several Ph.D.
students, Master students, and research fellows. She also gave courses and seminars in Statistics
and Computational Biology at several universities for Master’s and Ph.D. students and was a
member of the evaluation committee for several projects, including European projects.

5
 Current and Future Trends in Recommender Systems
Tommaso Di Noia
Politecnico di Bari, Italy.

[email protected]

Recommender systems have become an integral component of modern digital ecosystems,

shaping user experiences across various domains such as e-commerce, social media, and streaming
services. This talk will explore the current landscape of recommender systems and address
emerging trends and future directions in the field. We will give an overview of current trends in
recommender systems research and discuss potential evolutions of the recommendation problem.

Short Bio Tommaso Di Noia is a Professor of Computer Science at Politecnico di Bari

(Italy). His research activities, mainly focused on Artificial Intelligence and Data Management,
were initially devoted to theoretical and practical issues in knowledge representation and auto-
mated reasoning. In these fields, he proposed innovative solutions to knowledge-aware resource
retrieval and matching by exploiting non-monotonic automated reasoning techniques. Then, he
moved to study how to apply knowledge representation techniques and tools both to automated
negotiations among rational agents with preferences and to mobile and ubiquitous computing
scenarios and protocols. Following these ideas, he started to study applications of knowledge
graphs and Linked Open Data datasets to user modeling and recommender systems. He has
recently been publishing many works covering theoretical, algorithmic, and experimental aspects
on the subject of recommender systems. During the last years, he has also focused on secu-
rity and privacy issues related to recommender systems with a specific emphasis on adversarial
and federated machine learning. Tommaso Di Noia has published many papers in international
journals, conferences, and book chapters related to his research interests. Some of them have
been awarded the Best Paper Award in different conferences. He is a recipient of IBM Ph.D.
Fellowship in 2015 and HP Labs Innovation Research Program Award in 2011 and 2012.

6
 Optimization, the philosophical background of artificial
intelligence
Yurii Nesterov
UCLouvain, Belgium.

[email protected]

We discuss new challenges in the modern Science, created by Artificial Intelligence (AI).
Indeed, AI requires a system of new sciences, mainly based on computational models. Its devel-
opment has already started by the progress in Computational Mathematics. In this new reality,
Optimization plays an important role, helping the other fields with finding tractable models and
efficient methods, and significantly increasing their predictive power. We support our conclusions
by several examples of efficient optimization schemes related to human activity.

Short Bio Yuri Nesterov is a renowned mathematician and one of the leading experts in
optimization theory. He is a professor at the Université catholique de Louvain (UCLouvain) in
Belgium, where he has made groundbreaking contributions to the field of convex optimization,
particularly in the development of fast gradient methods. Prof. Nesterov is best known for intro-
ducing Nesterov’s Accelerated Gradient (NAG) method, a cornerstone of modern optimization
algorithms widely used in machine learning and artificial intelligence. His work spans convex and
non-convex optimization, large-scale optimization, and polynomial optimization, with profound
impacts on both theoretical and applied aspects of the field. He has authored several influen-
tial books, including Introductory Lectures on Convex Optimization: A Basic Course, and has
received numerous prestigious awards, such as the John von Neumann Theory Prize, for his
significant contributions to optimization and mathematical sciences.

7
Keynotes Speakers
• Andersen Ang (page 9)
MGProx: A nonsmooth multigrid proximal gradient method with adaptive restriction for
strongly convex optimization
• Stefano Coniglio (page10)
Graph and Hypergraph Learning via Complex- and Quaternion-Valued Spectral Convolu-
tional Operators
• Giacomo De Palma (page 11)
Trained quantum neural networks are Gaussian processes
• Stefania Fresca (page 12)
Latent Dynamics Models
• Alessandro Gianola (page 13)
Formal Analysis of Data-Aware Processes via Symbolic AI
• Cesare Molinari (page 14)
Stochastic (but structured) zeroth order optimization
• Katerina Papagiannouli (page 15)
Bures-Wasserstein gradient-based learning of covariance operators in Gaussian processes
• Monica Pragliola (page 16)
Whiteness-based learning of parameters in inverse imaging problems

8
MGProx: A nonsmooth multigrid proximal gradient method
with adaptive restriction for strongly convex optimization
Andersen Ang
School of Electronics and Computer Science University of Southampton, UK

[email protected]

We study the combination of proximal gradient descent with multigrid for solving a class of
possibly nonsmooth strongly convex optimization problems. We propose a multigrid proximal
gradient method called MGProx, which accelerates the proximal gradient method by multigrid,
based on utilizing hierarchical information of the optimization problem. MGProx applies a newly
introduced adaptive restriction operator to simplify the Minkowski sum of subdifferentials of the
nondifferentiable objective function across different levels. We provide a theoretical character-
ization of MGProx. First we show that variables at all levels exhibit a fixed-point property at
convergence. Next, we show that the coarse correction is a descent direction for the fine variable
in the general nonsmooth case. Lastly, under some mild assumptions we provide the convergence
rate for the algorithm, such as the classical sub-linear rate and also the linear rate. By treat-
ing the multigrid proximal gradient iteration as a black-box, we also proposed a fast MGProx
with Nesterov’s acceleration, together with the classical rate. In the numerical experiments, we
show that MGProx has a significantly faster convergence speed than proximal gradient descent
and proximal gradient descent with Nesterov’s acceleration on nonsmooth convex optimization
problems such as the Elastic Obstacle Problem, which the restriction operator is well known.

9
 Graph and Hypergraph Learning via Complex- and
Quaternion-Valued Spectral Convolutional Operators
Stefano Coniglio
Department of Economics, University of Bergamo [email protected]

In many learning problems, graphs and hypergraphs are powerful abstractions that can used
to model various types of interactions among the elements of a given dataset. Over the past
years, these structures have been attracting a growing interest in the deep-learning literature
thanks to many successful applications in several fields, including key ones in chemistry and
biology. Hypergraphs, in particular, are crucial for their capability of representing real-world
phenomena involving polyadic (many-to-many) relations between the elements, generalizing the
simpler diadic (pairwise) relationships that are classically captured by a graph. While the possi-
bility of capturing asymmetric relationships (either diadic or polyadic) within a dataset is crucial
in many applications, (hyper)edge directions are often ignored in many state-of-the-art works
that rely on a convolutional operator of spectral type, i.e., one grounded in graph-signal theory.
In this presentation, we survey recent results in directed (hyper)graph learning based on the
construction of complex- or quaternion-valued graph Laplacian matrices which are suitably de-
signed to capture the (hyper)edge directions while being amenable for the construction of spectral
convolutional operator. In particular, we present the Sign-Magnetic Laplacian and SigMaNet,
a generalized Graph Convolutional Network (GCN) capable of handling both undirected and
directed graphs with weights not restricted in sign nor magnitude; a quaternion-valued extension
of the Sign-Magnetic Laplacian which is suitable for graphs involving digons (antiparallel edges)
of asymmetric weights and its associated GCN QuaterGCN; the Generalized Directed Lapla-
cian and GeDi-HNN, a Hypergraph Neural Network (HNN) suitable for hypergraph-learning
tasks involving hyperedge directions; and the Directed Line Graph Laplacian and its associ-
ated HNN DLGNet, which are designed to tackle chemical-reaction classification problems by a
suitably-designed transformation of the input directed hypergraph to a directed line graph with
complex-valued edge weights.

References
[1] Stefano Fiorini, Stefano Coniglio, Michele Ciavotta, Enza Messina SigMaNet: One Laplacian to
Rule Them All, Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 11, pp.
7568-7576, 2023.
[2] Stefano Fiorini, Stefano Coniglio, Michele Ciavotta, Enza Messina Graph Learning in 4D: A
Quaternion-Valued Laplacian to Enhance Spectral GCNs, Proceedings of the AAAI Conference
on Artificial Intelligence, vol. 38, no. 11, pp. 12006-12015, 2024.
[3] Stefano Fiorini, Stefano Coniglio, Michele Ciavotta, Alessio Del Bue Let There be Direction in
Hypergraph Neural Networks, Transactions on Machine Learning Research, 2024.
[4] Stefano Fiorini, Giulia M. Bovolenta, Stefano Coniglio, Michele Ciavotta, Pietro Morerio, Michele
Parrinello, Alessio Del Bue DLGNet: Hyperedge Classification through Directed Line Graphs for
Chemical Reactions, arXiv preprint arXiv:2410.06969, October 2024.

10
 Trained quantum neural networks are Gaussian processes
Giacomo De Palma
University of Bologna, Department of Mathematics, Piazza di Porta San Donato 5, 40126 Bologna BO,

Italy [email protected]

Filippo Girardi
Scuola Normale Superiore, Piazza dei Cavalieri 7, 56126 Pisa PI, Italy [email protected]

Quantum neural networks represent the quantum analog of deep neural networks, leveraging
the unique properties of quantum mechanics to potentially enhance machine-learning algorithms.
Despite their promise, quantum neural networks currently lack a solid mathematical foundation.
This work aims to establish such foundation.
We investigate quantum neural networks for supervised learning, constructed with parametric
one-qubit gates and fixed two-qubit gates, where the output function is the expectation value of
the sum of single-qubit observables across all qubits.
First, we demonstrate that the probability distribution of the function generated by untrained
quantum neural networks with randomly initialized parameters converges in distribution to a
Gaussian process in the limit of infinite width, provided that each measured qubit is correlated
with only a few other qubits.
Then, we analytically characterize the gradient-descent training dynamics of the network in
the limit of infinite width. We prove that the loss function decays exponentially in the training
time, and therefore that the trained network can perfectly fit the training set. Moreover, we
prove that during the whole training, the probability distribution of the generated function still
converges in distribution to a Gaussian process. The proof of such a result relies on proving that
training occurs in the lazy regime, i.e., that the maximum variation of each parameter vanishes
in the limit of infinite width.
Finally, we address the statistical noise in measurements at the output of the network, proving
that number of measurements growing polynomially with the number of qubits is sufficient to
ensure the convergence to a Gaussian process, and therefore that the network can be trained in
polynomial time.

References
[1] Filippo Girardi, Giacomo De Palma Trained quantum neural networks are Gaussian processes,
arXiv:2402.08726

11
 Latent Dynamics Models
Stefania Fresca
Address [email protected]

Nicola Farenga, Simone Brivio, Andrea Manzoni

Address {nicola.farenga, simone.brivio, andrea1.manzoni}@polimi.it

Solving differential problems using full order models (FOMs), such as the finite element
method, usually results in prohibitive computational costs, particularly in real-time simulations
and multi-query routines. Reduced order modeling aims to replace FOMs with reduced order
models (ROMs) characterized by much lower complexity but still able to express the physical
features of the system under investigation. Within this context, deep learning-based reduced
order models (DL-ROMs) have emerged as a novel and comprehensive approach, offering efficient
and accurate surrogates for solving parametrized time-dependent nonlinear PDEs. By leveraging
the mathematical properties of the system, the accuracy and generalization capabilities of DL-
based ROMs can be further enhanced.
In this respect, latent dynamics models (LDMs) represent a novel mathematical framework
in which the latent state is constrained to evolve according to an (unknown) ODE. A time-
continuous setting is employed to derive error and stability estimates for the LDM approximation
of the FOM solution. The impact of using an explicit Runge-Kutta scheme in a time-discrete
setting is then analyzed, resulting in the ∆LDM formulation. Additionally, the learnable setting,
∆LDMθ , is explored, where deep neural networks approximate the discrete LDM components,
ensuring a bounded approximation error with respect to the high-fidelity solution. Moreover,
the framework demonstrates the capability to achieve a time-continuous approximation of the
FOM solution in a multi-query context, thus being able to compute the LDM approximation at
any given time instance while retaining a prescribed level of accuracy.

References
[1] N. Farenga, S. Fresca, S. Brivio, A. Manzoni On latent dynamics learning in nonlinear reduced
order modeling, arXiv preprint arXiv:2408.15183, 2024.
[2] S. Fresca, A. Manzoni POD-DL-ROM: enhancing deep learning-based reduced order models for
nonlinear parametrized PDEs by proper orthogonal decomposition, Computer Methods in Applied
Mechanics and Engineering, 388, 114181, 2022.

12
 Formal Analysis of Data-Aware Processes via Symbolic AI
Alessandro Gianola
INESC-ID/Instituto Superior Técnico, Universidade de Lisboa, Lisbon, Portugal

[email protected]

Contemporary organizations are complex organisms involving multiple actors that use mul-
tiple resources to perform activities, interact with data objects and take decisions based on
this interaction. This inherent complexity highlights the growing need for advanced modeling
and analysis of business processes using modern and efficient techniques from various areas of
computer science, and for the automatic regulation of their internal work by exploiting Arti-
ficial Intelligence (AI) methods. Business process management (BPM [5]) has emerged as a
well-established research field and industry-oriented discipline at the crossroads of operations
management, computer science, data science, and software and systems engineering. Its pri-
mary goal is to support managers, analysts, ICT professionals, and domain experts in designing,
deploying, enacting, and continuously improving processes to meet organizational objectives.
Addressing the intricate nature of modern business processes requires safe and trustworthy sys-
tems that stakeholders and practitioners can depend on, posing significant challenges for both
modeling and analysis.
The complexity intensifies when business processes are analyzed not only through their con-
trol flow but also by examining their interaction with data [4]. The data dimension can take
various forms, such as case variables that encapsulate data objects or more intricate persistent
storage systems like relational databases. Recently, considerable research across different fields
has focused on integrating data and processes [3] to gain a deeper understanding of their dynamic
interaction. This integration necessitates exploring how data influences process behavior and,
conversely, how the control flow of the process affects the data it accesses and modifies. We call
such complex systems data-aware processes.
Overall, the integration of BPM and AI is transforming the development of intelligent and
reliable information systems, in particular when these systems integrate processes and data. On
the one hand, BPM raises novel and challenging questions about processes and the event data
they generate during execution. On the other hand, AI provides a set of robust techniques that
require continuous adaptation and refinement to address these questions successfully. In this
talk, I will explore how the integration of BPM and AI paves the way for innovative systems
capable of managing organizational complexities while meeting operational goals. A particular
emphasis will be placed on advanced techniques for the analysis of data-aware processes [6],
considering tasks such as formal verification [2] and conformance checking [1]. Specifically, I
will argue how symbolic AI and formal methods provide rigorous foundational approaches and
powerful tools to precisely specify and analyze generic relational dynamic systems for capturing
data-aware processes, ensuring both reliability and robustness.

References
[1] W. M. P. van der Aalst. Process Mining - Data Science in Action, Second Edition, Springer, 2016
[2] C. Baier, J.-P. Katoen. Principle of Model Checking, MIT Press, 2008
[3] D. Calvanese, G. De Giacomo, M. Montali. Foundations of data-aware process analysis: a database
theory perspective, In Proceedings of PODS, 2013
[4] M. Dumas. On the convergence of data and process engineering, In Proceedings of ADBIS 2011,
volume 6909 of LNCS. Springer, 2011
[5] M. Dumas, M. La Rosa, J. Mendling, H. A. Reijers. Fundamentals of Business Process Management,
Second Edition, Springer, 2018
[6] A. Gianola. Verification of Data-Aware Processes via Satisfiability Modulo Theories, Lecture Notes
in Business Information Processing 470, Springer, 2023

13
 Stochastic (but structured ) zeroth order optimization
Cesare Molinari
Università di Genova, MaLGa Center; [email protected]

M. Rando, C. Traoré, D. Kozak, L. Rosasco, S. Villa

Finite-difference methods are a class of algorithms designed to solve black-box optimization

problems by approximating a gradient of the target function on a set of directions. In black-box
optimization, the non-smooth setting is particularly relevant since, in practice, differentiability
and smoothness assumptions cannot be verified. To cope with non-smoothness, several authors
use a smooth approximation of the target function and show that finite difference methods ap-
proximate its gradient. Recently, it has been proved that imposing a structure in the directions
allows improving performance. However, only the smooth setting was considered. To close this
gap, we introduce and analyze O-ZD, the first structured finite-difference algorithm for non-
smooth black-box optimization. Our method exploits a smooth approximation of the target
function and we prove that it approximates its gradient on a subset of random orthogonal direc-
tions. We analyze the convergence of O-ZD under different assumptions. For non-smooth convex
functions, we obtain the optimal complexity. In the non-smooth non-convex setting, we charac-
terize the number of iterations needed to bound the expected norm of the smoothed gradient.
For smooth functions, our analysis recovers existing results for structured zeroth-order methods
for the convex case and extends them to the non-convex setting. We conclude with numerical
simulations where the assumptions are satisfied, observing that our algorithm has very good
practical performance.

References
[1] D. Kozak, C. M., S. Villa, L. Rosasco, L. Tenorio: Zeroth-order optimization with orthogonal random
directions, Mathematical Programming, 199 (1-2), 1179-1219 (2023)
[2] M. Rando, C. M., L. Rosasco, S. Villa: An Optimal Structured Zeroth-order Algorithm for Non-
smooth Optimization, Advances in Neural Information Processing Systems 36 (2023)
[3] M. Rando, C. M., S. Villa, L. Rosasco: Stochastic Zeroth order Descent with Structured Directions
Computational Optimization and Applications, 1-37 (2024)
[4] M. Rando, C. Traoré, C. M., L. Rosasco, S. Villa: A Structured Proximal Stochastic Variance
Reduced Zeroth-order Algorithm, in preparation

14
 Bures-Wasserstein gradient-based learning of covariance
operators in Gaussian processes
Katerina Papagiannouli
University of Pisa & Max Planck Institute Mathematics in the Sciences;

[email protected]

We study gradient-based learning of covariance operators in Gaussian processes, emphasizing

low-rank approximations. Based on the Bures-Wasserstein gradient flow framework, we propose
methods to learn covariance eigenvalues and eigenfunctions feature-by-feature. Utilizing the
dynamics of the Kullback-Leibler (KL) divergence within the framework of Gaussian distribu-
tions equipped with Bures-Wasserstein geometry, we show convergence guarantees for eigenvalue
learning in various settings, including neural network architectures. Our approach extends to
neural network-based parametrizations, enabling scalable and efficient learning for complex data
distributions.

References
[1] K. P., P. Bréchet, j. An, G. Montufar Critical Points and Convergence Analysis of Generative Deep
Linear Networks Trained with Bures-Wasserstein Loss ICML (2023)
[2] K. P, P. Brechét, A. Agazzi: Learning covariance operators feature by feature Gradient-based low
rank approximation of Gaussian processes, in preparation

15
 Whiteness-based learning of parameters in inverse imaging
problems
Monica Pragliola
Department of Mathematics and Applications, University of Naples Federico II

[email protected]

Variational methods for ill-posed imaging inverse problems aim to minimize a functional
which is sum of a fidelity term and of a regularization term, the two terms being balanced by the
so-called regularization parameter. It is well-established that flexible models are characterized
by highly-parametrized regularizers, and it is thus crucial to design robust methods for the
selection of the possibly high number of parameters arising in the models of interest. In this
talk, we take a journey through the different instances of the Residual Whiteness Principle,
an unsupervised approach that has been originally introduced for the estimation of the single
regularization parameter in variational models [3]. In its seminal version, the RWP is applied to
white-noise corrupted data and it amounts to maximize the whiteness of the residual image, i.e.
to minimize the autocorrelation of its entries. We will discuss how the RWP can be extended
so as to be applied to non-white yet whitenable noise statistics, such as, e.g. Poisson noise and
mixed Poisson-Gaussian noise [2, 1]. Moreover, we will show how the bilevel optimization task
expressing the RWP can be tackled so as to reduce the computational costs and to make it
possible to employ the whiteness-based unsupervised principle for the estimation of a general
large number of unknown parameters [4, 1].
This talk summarizes the results achieved with several co-authors: Francesca Bevilacqua, Alessan-
dro Lanza, Fiorella Sgallari, Luca Calatroni, Marco Donatelli, Carlo Santambrogio.

References
[1] Bevilacqua F., Lanza A., Pragliola M., Sgallari F. A general framework for whiteness-based param-
eters selection in variational models , Computational Optimization and Applications (2024)
[2] Bevilacqua F., Lanza A., Pragliola M., Sgallari F. Whiteness-based parameter selection for Poisson
data in variational image processing , Applied Mathematical Modelling, 117 (2023)
[3] Lanza A., Pragliola M., Sgallari F. Residual whiteness principle for parameter-free image restoration,
Electronic Transactions on Numerical Analysis, 53 (2020)
[4] Santambrogio C., Pragliola M., Lanza A., Donatelli M., Calatroni L. Whiteness-based bilevel learning
of regularization parameters in imaging, European Signal Processing Conference (2024)

16
Contributed Talks
• Linda Albanese (page 20)
Boolean SK model
• Andrea Alessandrelli (page 21)
Networks of neural networks: disentanglement of overlapping inputs

• Antonioreneè Barletta (page 22)

Exploring Deep Learning in Seismology for Early Warning systems
• Vittorio Bauduin (page 23)
Simulations of Water Distribution Systems via Radial Basis Function Neural Networks

• Cristian Belfiore (page 24)

An efficient matheuristic for nurse rostering problems
• Alessia Benevento (page 25)
Semi-Supervised Learning for Time Series Clustering Using Copulas

• Giovanni Bocchi (page 26)

Graph distinction through GENEOs and Permutants
• Simone Brivio (page 27)
Mitigating the adverse effects of data scarcity through pre-trained physics-informed DL-
ROMs

• Filippo Camellini (page 28)

Majorization-Minimization for multiclass classification in a big data scenario
• Davide Carrara (page 30)
Implicit Neural Field Reconstruction on Complex Shapes from Scattered Data

• Edoardo Centofanti (page 31)

Operator Learning Techniques in Computational Cardiology
• Francesco Conti (page 32)
A unified framework for equivariant neural network
• Ivan Cucchi (page 33)
Integrating Molecular Dynamics and Machine Learning Algorithms to Predict the Func-
tional Profile of Kinase Ligands
• Ben William Gerriety Cullen (page 34)
GANs through the Lens of Topological Data Analysis

• Arturo De Marinis (page 35)

Approximation properties of neural ODEs
• Francesco Della Santa (page 36)
Learning Variably Scaled Kernels and Scaling Functions via Discontinuous Neural Networks
• Simmaco Di Lillo (page 37)
Spectral Complexity of Deep Neural Networks
• Nunzio Dimola (page 38)
A Neural Preconditioner for the Numerical Solutions of Parametrised PDEs
• Davide Duma (page 39)
Optimizing patient admission in the emergency department with machine learning-based
survival models

17
• Flavia Esposito (page 40)
Low-rank approximation methods for real data analysis and integration

• Alberto Fachechi (page 41)

A random matrix approach to Hopfield-like neural networks: addressing generalization and
overfitting
• Nicola Farenga (page 42)
On latent dynamics learning in nonlinear reduced order modeling

• Stefania Ferrisi (page 43)

Mathematical Transformations and Deep Learning Methodologies to enhance Tool Wear
Monitoring using Audio Data
• Nicola Rares Franco (page 44)
Deep orthogonal decomposition: an adaptive basis approach to dimensionality reduction
• Bharath Krishnan Girishkumar (page 45)
Penalized Maximum Likelihood and Loss Minimization for Classification
• Marc Hirschvogel (page 46)
Learning Passive Left Ventricular Mechanics via Shape Encoding Neural Networks

• Sofia Imperatore (page 47)

Data-driven parameterization for adaptive spline model reconstruction
• Samira Iscaro (page 48)
A new mathematical model to analyze the spread of misinformation on Social Media

• Giacomo Lancia (page 49)

Constructing Interpretable Prediction Models with Semi-Orthogonal 1D DNNs: An Exam-
ple in Irregular ECG Classification
• Giulia Lombardi (page 50)
Hexagonal Grid-Based Reinforcement Learning Environments for Marine Biodiversity Mon-
itoring
• Andrea Manzoni (page 51)
Multi-fidelity reduced-order surrogate modelling
• Anderson Melchor Hernandez (page 52)
Convergence of quantum neural networks at infinite width
• Marta Menci (page 53)
An all-around perspective on hybrid coupled models and parameter calibration for collective
cell dynamics
• Giovanni Pagano (page 54)
Step-by-Step Time-Discrete Physics Informed Neural Networks for PDEs models
• Davide Pastorello (page 55)
Training a quantum GAN with classical data
• Danilo Pezzi (page 56)
Linesearch-Enhanced Forward-Backward Methods for Inexact Nonconvex Scenarios
• Moreno Pintore (page 57)
The Neural Approximated Virtual Element Method on general polygons
• Domenico Pomarico (page 58)
Grokking as an entanglement transition in tensor network machine learning

18
• Maria Grazia Quarta (page 59)
A CNN-LSTM approach for parameter estimation for lithium metal battery cycling model

• Luca San Mauro (page 60)

On the complexity of infinite argumentation
• Alessandro Scagliotti (page 61)
Trade-off Invariance Principle for regularized functionals

• Vincenzo Schiano di Cola (page 62)

Quantum Optimization in Environmental Resource Management: A Focus on Irrigation
Scheduling
• Dhruv Singhvi (page 63)
A Framework Combining Machine Learning and Statistical Modeling for Detecting Extreme
Events in High-Dimensional Data
• Cristiano Tamborrino (page 64)
A Deep-QLP Decomposition Algorithm and Applications
• Ilaria Trombini (page 65)
Variable metric proximal stochastic gradient methods with additional sampling

19
 Boolean SK model
Linda Albanese
University of Salento, [email protected]

Andrea Alessandrelli
University of Pisa, [email protected]

In recent years, the rapid development of Artificial Intelligence (AI) solutions has profoundly
influenced contemporary scientific research. Its impact is reshaping the scope of applied disci-
plines [1, 2] while simultaneously inspiring theoretical interest in automated systems across fields
such as neuroscience, statistics, complex systems physics, engineering, and information theory.
The statistical mechanics of spin glasses has traditionally served as a paradigm for modelling
and interpreting diverse phenomena, spanning from quantitative biology to computer science.
Despite the substantial body of research in this field, there remains a notable gap concerning
the substitution of Ising spins with Boolean spins; given the role of Boolean variables as binary
units in Machine Learning, addressing this gap is now essential.
In this presentation, we will discuss an approach to filling this lacuna for the mean-field model
with Boolean variables and disordered couplings governed by a Gaussian distribution. Given the
similarities with the Sherrington-Kirkpatrick (SK) model [3, 4] – a foundational framework for
mean-field spin glasses – this model is naturally referred to as the Boolean SK model. Due
to time constraints, our focus will be on the application of Guerra’s interpolation method [5] to
derive the thermodynamic expression of the quenched statistical pressure under both the Replica
Symmetric and first-step Replica Symmetry Breaking assumptions.
However, despite the structural similarities, the Boolean SK model exhibits distinct character-
istics compared to the original SK model. Specifically, due to the breaking of spin-flip symmetry,
it exhibits an inherent magnetisation, and the overlap (an analogue for the SK model) lacks the
conventional phase transition. Instead, the system transitions continuously from a random state
to a disordered phase. All theoretical results are substantiated by numerical analyses.
This work may serve as a foundation for a series of studies aimed at understanding other
network models where Ising spins are replaced by Boolean spins.
This research is inspired by joint work with Andrea Alessandrelli (University of Pisa) [6].

References
[1] J. Leskovec, A. Rajaraman, J. D. Ullman Mining of massive datasets , Cambridge University Press,
2014.
[2] K. K. Jain Personalized medicine , Current Opinion in Molecular Therapeutics, 4(6):548-558, 2002.
[3] Sherrington, D., Kirkpatrick, S. Solvable model of a spin-glass , Physical review letters 35.26 (1975):
1792.
[4] Mézard, M., Parisi, G., Virasoro, M. A. Spin glass theory and beyond: An Introduction to the
Replica Method and Its Applications , Vol. 9. World Scientific Publishing Company, 1987.
[5] Guerra, Francesco Broken replica symmetry bounds in the mean field spin glass model , Communi-
cations in mathematical physics 233 (2003): 1-12
[6] Albanese, L., Alessandrelli, A. Boolean mean field spin glass model: rigorous results , arXiv preprint
arXiv:2409.08693 (2024).

20
 Networks of neural networks: disentanglement of
overlapping inputs
Andrea Alessandrelli
Università di Pisa [email protected]

Elena Agliari
Sapienza Università di Roma [email protected]
Adriano Barra
Sapienza Università di Roma [email protected]

Martino S. Centonze
Università di Bologna [email protected]
Federico Ricci-Tersenghi
Sapienza Università di Roma [email protected]

This work investigates the intersection of Artificial Intelligence and Statistical Mechanics,
focusing on the hetero-associative extension of the classic Hopfield network [1]. Indeed, we present
an extended version of the Bidirectional Associative Memory (BAM) [3] that can concurrently
process three or more patterns [2].
Our analysis shows that an ensemble of BAM models exhibits emergent capabilities absent
in a single network. Specifically, we design a layered associative Hebbian network that not only
performs standard pattern recognition but also achieves pattern disentanglement. For instance,
when we present a composite input – such as a musical chord – the network can extract the
individual elements constituting it, i.e. the distinct notes. In our investigation, we restrict to
notes represented as Rademacher vectors and chords constructed as their mixtures, analogous
to the spurious states in a Hopfield model. Through a statistical-mechanical analysis (both
analytical and computational), we derive the conditions on the model parameters that enable
successful pattern disentanglement.
Leveraging statistical mechanics, interpolation techniques, and phase diagrams, we character-
ize critical computational features and optimize network configurations. Numerical experiments
on hierarchical synthetic datasets confirm the model’s capability for input disentanglement, with
theoretical predictions aligning closely with the empirical results. This statistical-mechanical
framework not only enables optimized network parameterization but also provides a pathway for
a priori optimization of deep learning architectures, aligning network structure with the intrinsic
organization of the data under analysis.

References
[1] J. J. Hopfield, Neural networks and physical systems with emergent collective computational abilities,
Proceedings of the National Academy of Sciences of the United States of America, 79:2554–2558
(1982).
[2] E. Agliari, A. Alessandrelli, A. Barra, M.S. Centonze, F. Ricci-Tersenghi, Generalized hetero-
associative neural networks, arXiv preprint arXiv:2409.08151 (2024).
[3] B. Kosko, Bidirectional associative memories, IEEE Transactions on Systems, man, and Cybernet-
ics,18(1):49–60 (1988).

21
 Exploring Deep Learning in Seismology
for Early Warning systems
Antonioreneè Barletta
Università degli studi di Napoli Federico II [email protected]

S. Cuomo, G. Milano
Università degli studi di Napoli Federico II [email protected],
Spici srl [email protected]

One of the major challenges in seismology is the development of fast, precise, and robust
solutions for early warning (EW) systems. EW involves methodologies for detecting and rapidly
analyzing an earthquake’s initial, non-damaging primary (P) waves. These approaches aim to
estimate critical parameters such as the earthquake’s epicenter, magnitude, and potential impact,
allowing alerts to be issued before the arrival of the slower, destructive secondary (S) waves. In
this domain, seismic data are collected from sensors (typically seismographs) and recorded as time
series. These data capture essential characteristics of seismic waves, including their amplitude,
frequency, and timing, providing crucial information for accurate analysis and interpretation.
The research literature demonstrates the effectiveness of various machine learning approaches
for EW applications, for example, models such as random forests, gradient boosting algorithms,
and Support Vector Machines (SVMs) have been widely explored due to their robustness and
reliability. More recently, the emergence of deep learning, driven by advancements in high-
performance hardware like GPUs and TPUs, has revolutionized this research field. Techniques
involving Recurrent Neural Networks (RNNs), Long Short-Term Memory networks (LSTMs), and
Convolutional Neural Networks (CNNs) have shown excellent performance in seismology-related
tasks. In this study, we explore the application of a Temporal Convolutional Network (TCN) for
analyzing earthquake seismograms in the context of an EW system. Our investigation focuses
on leveraging the unique capabilities of TCNs to enhance the speed and accuracy of seismic data
analysis, oriented to the development of applications and services designed for EW scenarios.

References
[1] C. Satriano, Y. Wu, A. Zollo, H. Kanamori Earthquake early warning: Concepts, methods and
physical grounds, Soil Dynamics and Earthquake Engineering 31 (2), 106-118.
[2] W. Zhu, G. C. Beroza PhaseNet: a deep-neural-network-based seismic arrival-time picking method,
Geophysical Journal International, Volume 216, Issue 1, January 2019, Pages 261–273.
[3] R. Rea, S. Colombelli, L. Elia, A. Zollo Retrospective performance analysis of a ground shaking early
warning system for the 2023 Turkey–Syria earthquake, Communications Earth & Environment -
Nature, 5 (1), 332.
[4] X. Liu, T, Ren, H. Chen, G. M. Dimirovski, F. Meng, P. Wang Earthquake magnitude estimation
using a two-step convolutional neural network, Journal of Seismology - Springer, 2024
[5] F. Piccialli, S. Cuomo, F. Giampaolo, E. Prezioso Prediction method and related system, US Patent
App. 17/815,737.

22
Simulations of Water Distribution Systems via Radial Basis
Function Neural Networks
Vittorio Bauduin
University of Campania L. Vanvitelli, email: [email protected]

Salvatore Cuomo
University of Naples Federico II, email: [email protected]

Water Distribution Networks (WDNs) are critical infrastructures consisting of thousands of

interconnected elements, characterized by a meshed and irregular topology shaped by urban
layouts. Their structural robustness is inherently linked to the operational efficiency of both the
networks themselves and associated infrastructures. Analyzing such complex systems necessitates
a balanced methodology that integrates holistic and reductionist perspectives, supported by
tools to examine network topology, behavior, and dynamic evolution using advanced analytical
frameworks. This study adopts a predictive approach by integrating Radial Basis Function Neural
Networks (RBF-NNs) with real-time sensor data to enhance understanding and management of
WDNs. Numerous methods for approximating multivariate functions have their own benefits but
struggle with high dimensionality, common in water distribution networks (WDN) applications.
This “curse of dimensionality" limits traditional methods’ effectiveness due to computational
challenges. As a result, more specialized approaches like reduced-rank techniques, sparse grid
approximations, or neural networks are preferred for their computational efficiency and accuracy
in high-dimensional contexts. RBF-NNs emerge as a robust solution to address these challenges,
as capturing the local-to-global temporal inference inherent to the dynamic behavior of these
networks. Their efficacy has been validated through numerical testing in this study, with the
primary objective of developing an interpolative model capable of accurately predicting pressures
and flows across the entire WDN using available sensor data. The proposed model demonstrated
excellent performance metrics, achieving a Loss Function value and Mean Squared Error (MSE)
of approximately 10−12 , alongside a Mean Absolute Error (MAE) of around 10−7 . These results
underline the high accuracy and reliability of the predictive approach, validating its potential for
effective WDN analysis and management.
The work was carried out in collaboration with Prof. Giovanni Francesco Santonastaso from the Department
of Engineering at the University of Campania "L. Vanvitelli", who served as the tutor for the research project
PRIN 2022 “SMART RENEW” - Rehabilitation of Water Distribution Networks through a Data-Driven Approach
(CUP: B53D23006080006).

References
[1] A. Di Nardo, C. Giudicianni, R. Greco, M. Herrera, G. F. Santonastaso, Applications of Graph
Spectral Techniques to Water Distribution Network Management, Water, vol. 10, no. 1, p. 45, Jan.
2018, doi: 10.3390/w10010045.
[2] M. D. Buhmann, Radial basis functions, Acta Numerica, vol. 9, pp. 1–38, Jan. 2000, doi:
10.1017/s0962492900000015.
[3] V. Bauduin, S. Cuomo, V. Schiano Di Cola, Constraint Satisfaction approach for Neuron Configu-
rations in Neural Networks,(submitted for pubblication)

23
 An efficient matheuristic for nurse rostering problems
Cristian Belfiore
de-Health Lab - Laboratory of Decision Engineering for Health Care Services Dep. of Mechanical,

Energy and Management Engineering, University of Calabria, Ponte Pietro Bucci Cubo 41C, 87036

Arcavacata di Rende (CS), Italy [email protected]

Rosita Guido, Domenico Conforti

de-Health Lab - Laboratory of Decision Engineering for Health Care Services Dep. of Mechanical,

Energy and Management Engineering, University of Calabria, Ponte Pietro Bucci Cubo 41C, 87036

Arcavacata di Rende (CS), Italy [email protected] [email protected]

Matheuristics represent a novel approach to problem-solving, combining elements of both

exact and heuristic methodologies. FiNeMath is a matheuristic that employs a combination of a
Large Neighbourhood Search approach and a Destroy-and-Reconstruct strategy. In this study,
we adapt FiNeMath to address the well-known nurse rostering problem. The goal of this problem
is to assign a set of nurses to shifts within a predefined period. Nurses are classified according
to one or more skills, are engaged under a contract that governs their work, and may express
preferences regarding days-off, shifts-off, or shifts-on. For each shift in the planning period,
a preferred level of staffing is defined. This value represents the number of nurses considered
optimal by the hospital. The assignments must comply with a set of constraints derived from
the hospital’s internal rules and legal requirements, as well as the preferences expressed by the
nurses. The objective is to obtain shifts coverage while ensuring staff satisfaction and workload
balance.
In our approach, we start by constructing an initial feasible solution to the problem by using a
Fix-and-Optimize strategy. Then, we attempt to iteratively improve this solution by partially
destroying it and then reconstruct it through a solver. The ruined portion is identified by a
destruction operator that varies iteration to iteration. The experimental campaign conducted
on benchmark instances available in the literature demonstrated promising outcomes for our
solution approach.

References
[1] Cristian Belfiore. An effective matheuristic approach to solve Nurse Rostering Problem, 7AYW-
8AYW· Operations Research Beyond Frontier – Proceedings of the 7th and the 8th AIROYoung
Workshops. AIRO Springer Series. (2024) [in press]
[2] Guido, R., Groccia, M. C., Conforti, D. An efficient matheuristic for offline patient-to-bed assign-
ment problems , European Journal of Operational Research, 268, 2, 486-503, (2018).
[3] Curtois, T., Qu, R. Computational results on new staff scheduling benchmark instances , Technical
Report, ASAP Research Group, School of Computer Science, University of Nottingham, NG8 1BB,
Nottingham, UK. (2014).

24
Semi-Supervised Learning for Time Series Clustering Using
Copulas
Alessia Benevento
Dipartimento di Matematica e Fisica “Ennio De Giorgi”, Università del Salento, Lecce, Italy

[email protected]

Fabrizio Durante
Dipartimento di Matematica e Fisica “Ennio De Giorgi”, Università del Salento, Lecce, Italy

[email protected]

Roberta Pappadà
Dipartimento di Scienze Economiche, Aziendali, Matematiche e Statistiche “B. de Finetti”, Università
degli Studi di Trieste, Trieste, Italy [email protected]

Time-series data containing one or multiple variables that vary with time is extensively
recorded and analyzed in various fields, such as science, engineering, medicine, economics, and
finance. Clustering is a powerful data mining technique for classifying these temporal data into
related groups in the absence of sufficient prior knowledge of the groups. Clustering methods
for time series are typically performed in unsupervised learning settings, where the aim is to
uncover hidden structures in the data. However, if the data comes with additional background
information, such as pairwise positive/negative relationships with associated degrees among the
time series, this can impose constraints on the clustering process. In such cases, the approach
is more accurately described as semi-supervised learning. The first goal of this presentation is
to review certain aspects of dissimilarity-based clustering methods that have been introduced
within a copula framework.
Additionally, in many applications, the identification of clusters among time series is compli-
cated by the presence of spatial constraints and the need to capture complex dependence struc-
tures, including tail dependencies. This talk presents a novel semi-supervised learning framework
for clustering time series based on copula models, inspired by the methodologies introduced in
[1]. We leverage copula-based measures to model temporal dependence structures and tail be-
haviors. The semi-supervised approach lead to the clustering of the time-series while taking
into account spatial proximities. We demonstrate the method’s efficacy through simulated and
real-world datasets, highlighting its applicability in fields such as environmental monitoring.

References
[1] Benevento, A., Durante, F., and Pappadà, R. Tail-dependence clustering of time series with spatial
constraints , Environmental and Ecological Statistics (2024): 1-17

25
 Graph distinction through GENEOs and Permutants
Giovanni Bocchi
Address [email protected]

Massimo Ferri, Patrizio Frosini

Address [email protected], [email protected]

The theory of Group Equivariant Non-Expansive Operators (GENEOs) was initially devel-
oped in Topological Data Analysis for the geometric approximation of data observers, including
their invariances and symmetries. In this work we depart from that line of research and ex-
plore the use of GENEOs for distinguishing graphs up to isomorphisms. In doing so, we aim to
test the capabilities and flexibility of the operators obtained exploiting Generalized Permutants
specifically designed to search for interesting subgraph structures in the graphs to be tested.
Our experiments show that the isomorphism test we obtained using a minimal number of GE-
NEOs learned from data offers the best compromise between efficiency and computational costs
when tested on the comparison r-regular graphs. In addition, the actions on data of the learned
operators are easily interpretable. This helps to support the idea that GENEOs could be a
general-purpose approach to discriminative problems in Machine Learning when some structural
information about data and observers is explicitly given.

References
[1] Giovanni Bocchi, Patrizio Frosini, Alessandra Micheletti, et al. A geometric XAI approach to protein
pocket detection, xAI-2024 Late-breaking Work, Demos and Doctoral Consortium Joint Proceedings
- The 2nd World Conference on eXplainable Artificial Intelligence. CEUR https://ceur-ws.org/Vol-
3793/ (2024).
[2] Giovanni Bocchi, Stefano Botteghi, Martina Brasini, et al On the finite representation of linear group
equivariant operators via permutant measures, Annals of Mathematics and Artificial Intelligence
91.4 (2023), pp. 465-487. ISSN: 1012-2443. DOI: 10.1007/s10472-022-09830-1.
[3] Faraz Ahmad, Massimo Ferri, and Patrizio Frosini Generalized Permutants and Graph GENEOs,
Machine Learning and Knowledge Extraction 5.4 (2023), pp. 1905-1920. DOI: 10.3390/make5040092
[4] Mattia G. Bergomi, Patrizio Frosini, Daniela Giorgi, et al. Towards a topological–geometrical the-
ory of group equivariant non-expansive operators for data analysis and machine learning, Nature
Machine Intelligence 1.9 (2019), pp. 423-433. ISSN: 2522-5839. DOI: 10.1038/s42256-019-0087-3
[5] Ryoma Sato A Survey on The Expressive Power of Graph Neural Networks, Preprint at arXiv
(2020). DOI:10.48550/arXiv.2003.04078.

26
 Mitigating the adverse effects of data scarcity through
pre-trained physics-informed DL-ROMs
Simone Brivio† , Stefania Fresca, Andrea Manzoni
MOX, Dept. of Mathematics, Politecnico di Milano, P.zza Leonardo da Vinci 32, Milano, I-20133, Italy

[email protected], [email protected], [email protected]

Deep learning-based reduced order models (DL-ROMs) provide a comprehensive paradigm for
nonlinear model order reduction enabling the construction of fast and efficient surrogate models
for the simulation of nonlinear parametrized PDEs [3]. Experimental evidence and theoretical
results have recently demonstrated that the prediction accuracy of data-driven DL-ROMs is of-
ten unsatisfactory when only an insufficient amount of labeled data is available at the training
stage [1]. Unfortunately, data scarcity is common in Scientific Machine Learning (SciML) ap-
plications. Indeed, data are usually generated through synthetic solvers, which provide highly
accurate and reliable simulations, but generally demand excessive computational resources. For
this reason, we are normally only able to generate only a handful of labeled data, which are often
not representative of the entire parametric space.

To compensate for the accuracy shortfall brought about by data scarcity, we build on the fact
that the governing equations convey the same information as the data synthetically generated
through numerical solvers. Consequently, it is sound to minimize the residual of the governing
equation in the regions of the parametric space that are not properly covered by labeled training
data. The resulting physics-informed approach is unsupervised by nature and does not need
additional input-output pairs.
However, especially as the problem complexity increases, such physics-informed architecture
requires a significant amount of computational resources to be suitably trained, and its optimiza-
tion phase is prone to convergence failure. To avoid these side effects, by further intertwining
data and physics, we devise a novel two-step training strategy, consisting of (i) a fast and efficient
pre-training stage that enables the optimizer to quickly and stably approach the minimum in
the loss landscape, and (ii) a fine-tuning phase that further enhances the prediction accuracy.
Ultimately, we showcase the potential of the resulting paradigm, termed Pre-Trained Physics-
Informed DL-ROM (PTPI-DL-ROM), by assessing its performance in terms of prediction accu-
racy and training efficiency [2]. To this end, we consider a series of numerical experiments
involving parametrized PDEs stemming from computational fluid dynamics and mathematical
biology.

References
[1] Brivio, S., Fresca, S., Franco, N. & Manzoni, A. Error estimates for POD-DL-ROMs: a deep
learning framework for reduced order modeling of nonlinear parametrized PDEs enhanced by proper
orthogonal decomposition. Adv. Comput. Math.. 50 (2024)
[2] Brivio, S., Fresca, S. & Manzoni, A. PTPI-DL-ROMs: Pre-trained physics-informed deep learning-
based reduced order models for nonlinear parametrized PDEs. Computer Methods In Applied Me-
chanics And Engineering. 432 pp. 117404 (2024)
[3] Fresca, S., Dede’, L. & Manzoni, A. A comprehensive deep learning-based approach to reduced
order modeling of nonlinear time-dependent parametrized PDEs. Journal Of Scientific Computing.
87 pp. 1-36 (2021)

27
 Majorization-Minimization for multiclass classification in a
big data scenario
Filippo Camellini
Department of Physics, Informatics and Mathematics, Via Campi 213/B, 41125, Modena, Italy
[email protected]

Emilie Chouzenoux, Jean–Christophe Pesquet

CVN, Inria, CentraleSupélec, University Paris Saclay, 9 rue Joliot Curie, 91190, Gif-sur-Yvette, France

[email protected], [email protected]
Giorgia Franchini, Federica Porta
Department of Physics, Informatics and Mathematics, Via Campi 213/B, 41125, Modena, Italy

[email protected], [email protected]

Majorization-minimization (MM) algorithms are numerical optimization methods that sim-

plify the original problem by iteratively replacing the objective function with more tractable
approximations. Through an iterative process, at each step the minimization problem associated
with a surrogate function is solved. This surrogate function, also called the tangent majorant,
approximates the objective function and exhibits good properties, typically convexity.
The MM algorithms just described are widely used in the field of supervised learning. For
instance, in [1], an MM approach with a quadratic surrogate function is exploited to train a
binary SVM-based linear model with a squared hinge loss as the data fidelity term and a smooth
regularization term that induces sparsity.
The main objective and original contribution of our work is to propose a highly scalable
MM algorithm for training a linear multiclass classification model that leverages a data-fidelity
function and a regularizer term which are L-smooth. Specifically, when extending the MM
approach presented in [1] to the multiclass classification case, two challenges arise. The first
involves formulating a differentiable objective function, which leads to the use of the Weston-
Watkins formulation [2]. The second challenge relates to the case where the size of the training
set makes gradient computation unfeasible. In this big data context, it becomes necessary to
introduce techniques that use gradient approximation while still allowing the exploitation of MM
methods.
The Incremental MM algorithm we propose is inspired by a classical incremental gradient
scheme [3] in which the descent direction is rescaled using a symmetric positive definite matrix.
The use of a scaling matrix derived from the MM Quadratic approach presented in [1] allows
us to leverage a second-order approximation of the function, rather than relying solely on the
gradient. Unlike a classic second-order method that involves high computational costs at each
iteration, the structure of our proposed matrix enables more efficient computation. A square
matrix of order equal to the gradient dimension is calculated in a warm-up phase that precedes
the optimization process and can be easily parallelized. Subsequently, for each epoch, this matrix
remains constant and is updated by computing only a diagonal matrix.
The experiments presented in our work highlight the advantage of using the second-order
information introduced by the scaling matrix. In particular, MM Incremental exhibits better
performance compared to other algorithms commonly used in big data contexts, such as Incre-
mental or Stochastic Gradient Descent.

References
[1] Alessandro Benfenati, Emilie Chouzenoux, Giorgia Franchini, Salla Latva-Äijö, Dominik Narnhofer,
Jean–Christophe Pesquet, Sebastian J. Scott, Mahsa Yousefi Majoration-Minimization for Sparse

28
 SVMs, Advanced Techniques in Optimization for Machine Learning and Imaging, Springer Nature
Singapore, 2024
[2] Yutong Wang, Clayton Scott Weston-Watkins Hinge Loss and Ordered Partitions, Advances in
Neural Information Processing Systems, 2020
[3] Dimitri P. Bertsekas, John N. Tsitsiklis Gradient Convergence in Gradient methods with Errors,
SIAM Journal on Optimization, 2000

29
 Implicit Neural Field Reconstruction on Complex Shapes
from Scattered Data
Davide Carraraa , Marc Hirschvogela , Francesco Regazzonia , Simone
Pezzutob , Stefano Pagania
a
MOX, Dipartimento di Matematica, Politecnico di Milano, Milan, Italy
a
{davide.carrara, marc.hirschvogel, francesco.regazzoni, stefano.pagani}@polimi.it
b
Dipartimento di Matematica, Università di Trento, Trento, Italy
b
[email protected]

In many engineering and medical applications, reconstructing physical fields and domain
geometries from noisy, scattered data collected by local sensors is a critical task. Both the
statistical reconstruction of distributed quantities and the simulation of physical processes (typ-
ically modeled by means of partial differential equations) depend heavily on accurate geometry
reconstruction.
Meshless approaches, such as using Multi-Layer Perceptrons to represent Signed or Unsigned
Distance Function (S/U-DF) from target geometries, have been effective in tackling this chal-
lenge, but often require intense preprocessing of the data and are not suited for sparse datasets.
We propose two novel approaches for geometry reconstruction, tailored to scenarios of low and
high data numerosity, that require only point cloud representations and do not need mesh or
point correspondences for input. We present applications of each method for reconstructing
cardiac geometries.
For cases with high-quality data, we propose a supervised reconstruction pipeline using the
DeepSDF architecture [1]. This method combines an embedding model and a regression network
to learn and reconstruct the shapes of multiple objects using a shared network. Each geometry is
associated with a latent code that encodes shape information, enabling the generation of realistic
new synthetic shapes by sampling the latent space. We demonstrate the application of this
method for solving nonlinear PDEs on reconstructed geometries, where the latent code is used
for network conditioning [2]. For scenarios with limited or noisy data where SDF computation
is not feasible, we introduce a novel method [3] that reconstructs the geometry from surface-
level point measurements. Our approach employs a tailored loss function combining fit and
regularization terms, including a differential term based on the eikonal equation to enhance model
generalization. The reconstructed shape model is then used to predict distributed quantities on
the surface, taking into account its geometry. High accuracy and geometric fidelity are ensured
through supervised training and validation against derived surface properties such as gradients,
which are computed using automatic differentiation. We validate this method on both synthetic
datasets and an atrial cardiac geometry.
This project has been partially funded from the project PRIN2022, MUR, Italy, 2023–2025,
P2022N5ZNP “SIDDMs: shape-informed data-driven models for parametrized PDEs, with ap-
plication to computational cardiology”.

References
[1] Park, J. J., Florence, P., Straub, J., Newcombe, R., and Lovegrove, S. DeepSDF: Learning Con-
tinuous Signed Distance Functions for Shape Representation. , 2019 IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), 165–174.
[2] Regazzoni F., Pagani S., Quarteroni A. Universal Solution Manifold Net works (USM-Nets): Non-
Intrusive Mesh-Free Surrogate Models for Problems in Variable Domains , Journal of Biomechanical
Engineering, 144(12), 2022
[3] Carrara D., Regazzoni F., Pagani S. Implicit neural field reconstruction on complex shapes from
scattered and noisy data , Mox Report 40/2024

30
Operator Learning Techniques in Computational Cardiology
Edoardo Centofanti
Università di Pavia, Via Ferrata 5, 27100, Pavia, Italy [email protected]

Operator Learning methods are gaining significant attention in biomathematics and compu-
tational cardiology due to their ability to efficiently approximate complex dynamical systems.
These methods offer new opportunities for reducing computational costs while maintaining accu-
racy, making them particularly suited for addressing challenges in cardiac modeling. In this talk,
we will explore the application of Operator Learning techniques to tackle two key challenges in
cardiac electrophysiology. First, we examine their capability of learning ionic models [1], which
play a critical role in describing cellular excitability and action potential generation, but present
a challenging nonlinear and stiff dynamics. Second, we focus on applying the Fourier Neural
Operator (FNO) to learn activation and repolarization times [2], as evaluated through the mon-
odomain cardiac model. By comparing these approaches to traditional numerical solvers, we
will highlight their potential for accurately reconstructing electrophysiological dynamics while
significantly improving computational efficiency.

References
[1] E. Centofanti, M. Ghiotto, L.F. Pavarino Learning the Hodgkin-Huxley Model with Operator Learn-
ing Techniques, Computer Methods in Applied Mechanics and Engineering 432, Part A (2024):
117381.
[2] Joint work with G. Ziarelli, S. Scacchi, in preparation

31
 A unified framework for equivariant neural network
Francesco Conti
Address [email protected]

Equivariant neural networks are proving effective in many real-world scenrios [1]. For ex-
ample, Convolutional Neural Networks are the state-of-the-art in computer vision tasks and
Topological Data Analysis [2] (TDA) is achieving great accomplishment with noisy datasets.
In this talk, we are going to present a unified mathematical framework for equivariant neural
networks and show that both CNNs and TDA can be expressed using this framework that we
call Group Equivariant Non-Expansive Operators (GENEOs) [3].

References
[1] Gerken, Jan E and Aronsson, Jimmy and Carlsson, Oscar and Linander, Hampus and Ohlsson,
Fredrik and Petersson, Christoffer and Persson, Daniel Geometric deep learning and equivariant
neural network , Springer Artificial Intelligence Review
[2] Wasserman, Larry Topological data analysis , Annual Review of Statistics and Its Application
[3] Bergomi, Mattia G and Frosini, Patrizio and Giorgi, Daniela and Quercioli, Nicola Towards a
topological–geometrical theory of group equivariant non-expansive operators for data analysis and
machine learning , Nature Machine Intelligence

32
 Integrating Molecular Dynamics and Machine Learning
Algorithms to Predict the Functional Profile of Kinase
Ligands
Ivan Cucchi
University of Pavia - Dept. of Mathematics [email protected]

Elena Frasnetti
University of Pavia - Dept. of Chemistry

The modulation of protein function via designed small molecules is providing new opportuni-
ties in chemical biology and medicinal chemistry. While drugs have traditionally been developed
to block enzymatic activities through active site occupation, a growing number of strategies now
aim to control protein functions in an allosteric fashion, allowing for the tuning of a target’s ac-
tivation or deactivation via the modulation of the populations of conformational ensembles that
underlie its function. In the context of the discovery of new active leads, it would be very useful
to generate hypotheses for the functional impact of new ligands. Since the discovery and design
of allosteric modulators (inhibitors/activators) is still a challenging and often serendipitous tar-
get, the development of a rapid and robust approach to predict the functional profile of a new
ligand would significantly speed up candidate selection. Herein, we present different machine
learning (ML) classifiers to distinguish between potential orthosteric and allosteric binders. Our
approach integrates information on the chemical fingerprints of the ligands with descriptors that
recapitulate ligand effects on protein functional motions. The latter are derived from molecu-
lar dynamics (MD) simulations of the target protein in complex with orthosteric or allosteric
ligands. In this framework, we train and test different ML architectures, which are initially
probed on the classification of orthosteric versus allosteric ligands for cyclin-dependent kinases
(CDKs). The results demonstrate that different ML methods can successfully partition allosteric
versus orthosteric effectors (although to different degrees). Next, we further test the models with
FDA-approved CDK drugs, not included in the original dataset, as well as ligands that target
other kinases, to test the range of applicability of these models outside of the domain on which
they were developed. Overall, the results show that enriching the training dataset with chemical
physics-based information on the protein–ligand dynamic cross-talk can significantly expand the
reach and applicability of approaches for the prediction and classification of the mode of action
of small molecules.

References
[1] E. Frasnetti, I. Cucchi, S. Pavoni, F. Frigerio, F. Cinquini, S. A. Serapian, L. F. Pavarino, G.
Colombo Integrating Molecular Dynamics and Machine Learning Algorithms to Predict the Func-
tional Profile of Kinase Ligands , Journal of Chemical Theory and Computation, Vol. 20, Issue
20

33
 GANs through the Lens of Topological Data Analysis
Ben Cullen
University of Pisa [email protected]

B. T. Corradini , C. Gallegatib , S. Marzialib , G. A. D’Invernoc , M.

a

Bianchinib , F. Scarsellib
a
University of Florence, b University of Siena, c SISSA, Trieste

Generative Adversarial Networks (GANs) [1] aim to produce realistic samples by mapping a low-
dimensional latent space to a high-dimensional data space by exploiting an adversarial training
mechanism. Despite achieving state-of-the-art results, GAN training faces significant challenges
such as mode collapse, vanishing gradients, and inefficiencies in hyperparameter tuning, relying
on computationally expensive trial-and-error methods. In addition, GANs lack a clear early
stopping criterion, often leading to resource-intensive training processes.
This work investigates GANs using Topological Data Analysis (TDA) tools [3] to gain deeper
insights into their training dynamics and generative capabilities. By employing persistent ho-
mology, we examine the evolution of topological features during training, focusing on the conver-
gence of the generated manifold to that of real data. Through various experiments on MNIST
and CIFAR-10 datasets with different GAN models, we analyze the interplay between model ar-
chitecture, training stability, and performance, as well as characterise common issues in GANs.
In particular, we show that the Wasserstein distance between persistence diagrams, which sum-
marise the topological features of manifolds, is a robust tool for quantifying similarities between
generated and real data, offering a novel perspective on evaluating samples beyond conventional
metrics like the Frechet Inception Distance (FID) [2]. Indeed, the FID score is shown to be
insufficient in assessing the quality of generated images, neither alone nor in combination with
the Intrinsic Dimension estimation [4]. Our results suggest that homological features provide a
suitable characterisation of the generative process that can be valuable for uncovering insights
about the structural transformations occurring during the training of a GAN. This study lays
the foundation for integrating topology-based approaches into the optimization and assessment
of generative models, potentially enabling the formulation of an early stopping criterion.

References
[1] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A.
& Bengio, Y. Generative adversarial nets. Advances In Neural Information Processing Systems. 27
(2014)
[2] Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B. & Hochreiter, S. Gans trained by a two time-
scale update rule converge to a local nash equilibrium. Advances In Neural Information Processing
Systems. 30 (2017)
[3] Chazal, F. & Michel, B. An introduction to topological data analysis: fundamental and practical
aspects for data scientists. Frontiers In Artificial Intelligence. 4 pp. 667963 (2021)
[4] Pope, P., Zhu, C., Abdelkader, A., Goldblum, M. & Goldstein, T. The intrinsic dimension of images
and its impact on learning. ArXiv Preprint ArXiv:2104.08894. (2021)

34
 Approximation properties of neural ODEs
Arturo De Marinis
GSSI - Gran Sasso Science Institute, L’Aquila, Italy arturo.demarinis[at]gssi.it

Davide Murari1 , Elena Celledoni2 , Nicola Guglielmi3 , Brynjulf Owren4 ,

Francesco Tudisco5
1
University of Cambridge, Cambridge, United Kingdom dm2011[at]cam.ac.uk
2
NTNU - Norges teknisk-naturvitenskapelige universitet, Trondheim, Norway elena.celledoni[at]ntnu.no
3
GSSI - Gran Sasso Science Institute, L’Aquila, Italy nicola.guglielmi[at]gssi.it
4
NTNU - Norges teknisk-naturvitenskapelige universitet, Trondheim, Norway brynjulf.owren[at]ntnu.no
5
University of Edinburgh, Edinburgh, United Kingdom f.tudisco[at]ed.ac.uk

We study the universal approximation property (UAP) of shallow neural networks whose
activation function is defined as the flow of a neural ODE. We prove the UAP for the space of
such shallow neural networks in the space of continuous functions. In particular, we also prove
the UAP with the weight matrices constrained to have unit norm.
Furthermore, in [1] we are able to bound from above the Lipschitz constant of the flow of the
neural ODE, that tells us how much a perturbation in input is amplified or shrunk in output. If
the upper bound is large, then so it may be the Lipschitz constant, leading to the undesirable
situation where certain small perturbations in input cause large changes in output. Therefore,
in [2] we compute a perturbation to the weight matrix of the neural ODE such that the flow of
the perturbed neural ODE has Lipschitz constant bounded from above as we desire. This leads
to a stable flow and so to a stable shallow neural network.
However, the stabilized shallow neural network with unit norm weight matrices does not
satisfy the universal approximation property anymore. Nevertheless, we are able to prove ap-
proximation bounds that tell us how poorly and how accurately a continuous target function can
be approximated by the stabilized shallow neural network.
The results presented during this talk are being collected in [3].

References
[1] N. Guglielmi, A. De Marinis, A. Savostianov, and F. Tudisco. Contractivity of neural ODEs: an
eigenvalue optimization problem. arXiv preprint arXiv:2402.13092, 2024.
[2] A. De Marinis, N. Guglielmi, S. Sicilia, and F. Tudisco. Stability of neural ODEs by a control over
the expansivity of their flows. Work in progress.
[3] A. De Marinis, D. Murari, E. Celledoni, N. Guglielmi, B. Owren and F. Tudisco, Approximation
properties of neural ODEs. Work in progress.

35
Learning Variably Scaled Kernels and Scaling Functions via
Discontinuous Neural Networks
Francesco Della Santa
Dipartimento di Scienze Matematiche, Politecnico di Torino

[email protected]

Gianluca Audone, Emma Perracchione, Sandra Pieraccini

Dipartimento di Scienze Matematiche, Politecnico di Torino
[email protected], [email protected], [email protected]

This presentation describes a novel methodology to improve the accuracy of interpolation

techniques based on Variably Scaled Kernels (VSKs) by learning the scaling function directly
from data. The importance of selecting an appropriate scaling function in VSK methods is
well-documented, with studies suggesting that the function should mimic the key features of the
target, such as its discontinuities [1]. However, theoretical explanations for these observations
and practical methods for constructing such scaling functions are almost missing. This work
addresses both challenges, offering a theoretical framework alongside a practical, automated
solution for learning scaling functions using Discontinuous Neural Networks (δNNs) [2].
The theoretical results illustrated in this presentation justify the intuition that having scaling
functions mimicking the behavior of the target function can significantly improve approximation
accuracy. These theoretical findings also provide a robust foundation for understanding the role
of scaling functions in interpolation. To bridge the gap between theory and application, we
propose a novel approach for automatically learning scaling functions through δNNs, which are
designed to effectively learn both continuous and discontinuous features of a target function [2].
By leveraging the properties of δNNs, our method constructs scaling functions with characteristics
that resemble the target function’s ones; this is observed both for continuous and discontinuous
target functions.
The presentation includes numerical experiments that validate the theoretical claims, demon-
strating the practical efficacy of our approach. These examples involve classical interpolation
problems with both continuous and discontinuous target functions. Notably, the results highlight
that δNN-based scaling functions enable VSK methods to achieve greater accuracy, particularly
in challenging scenarios with discontinuities, outperforming conventional kernel interpolation
techniques.

References
[1] S. De Marchi, W. Erb, F. Marchetti, E. Perracchione, M. Rossini Shape-driven interpolation with
discontinuous kernels: Error analysis, edge extraction, and applications in magnetic particle imag-
ing, SIAM J. Sci. Comput. 42 (2) (2020) B472–B491
[2] F. Della Santa, S. Pieraccini Discontinuous neural networks and discontinuity learning, J. Comput.
Appl. Math. 419 (2023)

36
 Spectral Complexity of Deep Neural Networks
Simmaco Di Lillo, Domenico Marinucci, Michele Salvi, Stefano
Vigogna
RoMaDS - Department of Mathematics, University of Rome Tor Vergata, Rome, Italy

dilillo[at]mat.uniroma2.it

Understanding the spectral properties of neural networks is critical for unveiling their theoret-
ical foundations and practical performance. Fully connected networks with random initialization
are known to converge to isotropic Gaussian processes in the infinite-width limit. In this work,
we propose a novel approach to characterize network complexity by leveraging the angular power
spectrum of these limiting Gaussian fields. Specifically, we define sequences of random variables
associated with the angular power spectrum and provide a comprehensive asymptotic character-
ization of their distribution as network depth grows.
This framework enables a new classification of neural networks into three categories: low-
disorder, sparse, and high-disorder. Our analysis reveals distinct behaviors of common activa-
tion functions, with particular attention to the sparsity properties of ReLU networks. These
theoretical insights are supported by extensive numerical simulations.

37
 A Neural Preconditioner for the Numerical Solutions of
Parametrised PDEs
Nunzio Dimola∗
MOX, Department of Mathematics, Politecnico di Milano, Italy. [email protected]

Nicola Rares Franco

MOX, Department of Mathematics, Politecnico di Milano, Italy. [email protected]
Paolo Zunino
MOX, Department of Mathematics, Politecnico di Milano, Italy. [email protected]

Numerical solution of PDEs is widely recognized as fundamental in scientific and engineering

applications; nonetheless, the algebraic structure of the resulting discretized system presents
challenges in constructing efficient solution algorithms in many relevant applications. In addition,
the computational effort increases as the problem needs to be solved multiple times, to address
various instances of the parameter.
In this work, we propose a novel, matrix-free preconditioning strategy that leverages operator
learning to efficiently address a class of parametrized 3D-1D mixed-dimensional PDEs [1]. The
proposed preconditioner generalizes across varying shapes of the 1D manifold without requiring
retraining procedure, making it robust to changes in graph topology. Theoretical ground for pre-
conditioner learning approach is established, developing a fully unsupervised training procedure,
removing the need for prior problem solution data.
A key contribution is the introduction of a problem-specific data augmentation set, tailored
to the spectral properties of the 3D-1D coupling operator’s kernel. This enhancement enables the
preconditioner to smooth high-frequency error components associated with the coupling term,
thereby removing the need for explicit pre- or post-smoothing stages, often required by other
approaches (e.g. [1, 2]).
Numerical experiments demonstrate the competitiveness of the proposed approach against
established preconditioners, particularly in accelerating convergence in iterative solvers. The
preconditioner maintains robust performance across the parameter space without requiring a
setup stage, showing scalability for large-scale problems.
This study establishes a foundation for extending machine learning-based preconditioning
techniques to broader classes of coupled multi-physics systems, providing a powerful tool for
overcoming complex computational challenges in scientific computing.

References
[1] Federica Laurino and Paolo Zunino Derivation and analysis of coupled PDEs on manifolds with
high dimensionality gap arising from topological model reduction., ESAIM: Mathematical Modelling
and Numerical Analysis, 53(6), 2047-2080.
[2] Yael Azulay and Eran Treister Multigrid-augmented deep learning preconditioners for the helmholtz
equation., SIAM Journal on Scientific Computing, 45(3):S127–S151, 2022
[3] Alena Kopanivcáková and George Em Karniadakis Deeponet based preconditioning strategies for
solving parametric linear systems of equations., arXiv preprint arXiv:2401.02016, 2024.805

38
Optimizing patient admission in the emergency department
with machine learning-based survival models
Davide Duma, Vittorio Meini
Dipartimento di Matematica, Università di Pavia [email protected],

[email protected]

Roberto Aringhieri
Dipartimento di Informatica, Università degli Studi di Torino [email protected]

In Emergency Department (ED) management, optimizing key performance indicators such as

Door-To-Doctor Time (DTDT) and Emergency Department Length of Stay (EDLOS) is crucial
for improving efficiency and care quality[1]. DTDT measures the time from patient arrival to
doctor consultation, reflecting the speed of initial care. EDLOS tracks the total time a patient
spends in the ED, from admission to discharge. Reducing both indicators helps minimize waiting
times, prevent overcrowding, and ensure timely care delivery.
In addition, the rate of patients who Leave Without Being Seen (LWBS) is a critical per-
formance metric for EDs. LWBS refers to patients who leave the ED before being seen by a
doctor, often due to long wait times or overcrowding. These episodes pose a significant risk to
patients, as delays or lack of medical attention can lead to a worsening of their health conditions.
In contrast, prior optimization approaches have primarily focused on metrics such as DTDT and
EDLOS, overlooking the direct minimization of the LWBS rate.
Despite the widespread use of descriptive analytics to study this issue in the medical literature
[2], there have been limited efforts to explore predictive and prescriptive analytics. Previous
Machine Learning (ML) studies have proposed classification models to identify categories of
patients at risk of abandon (e.g., see [3]), but these models are not suitable to capture their
behavior over time. However, to model LWBS as an optimization objective, it is essential to
estimate the risk of abandonment as a function of waiting time.
This work introduces a dynamic physician-patient assignment framework based on an Integer
Linear Programming (ILP) model informed by predictions about the LWBS risk and solved in a
reactive way when the waiting list is updated. Using ML for survival analysis on censored data [4],
patient behavior is modeled based on their information collected at the triage. This approach
not only supports better decision-making in patient admission processes but also enables the
simulation of scenarios with extended DTDT compared to the one observed in the electronic
health records.
We consider a real case study of a medium-size Italian ED. A computational analysis evaluates
the proposed framework’s performance, analyzing the trade-off between DTDT, EDLOS, and
LWBS rates, by considering both the aleatory and the epistemic uncertainty. Results demonstrate
the potential of the proposed interplay beetween ML, ILP, and simulation for improving ED
admissions.

References
[1] D Duma, R Aringhieri. Real-time resource allocation in the emergency department: A case study,
Omega 117, 102844, 2023.
[2] M Johnson, S Myers, J Wineholt, M Pollack, AL Kusmiesz. Patients Who Leave the Emergency
Department Without Being Seen, Journal of Emergency Nursing 35(2), 105-108, 2009.
[3] M Sheraton, C Gooch, R Kashyap. Patients leaving without being seen from the emergency de-
partment: A prediction model using machine learning on a nationwide database, Journal of the
American College of Emergency Physicians Open 1(6), 1684-1690, 2020.
[4] P Wang, Y Li, CK Reddy. Machine Learning for Survival Analysis: A Survey, ACM Computing
Surveys 15(6): 1-36, 2019.

39
Low-rank approximation methods for real data analysis and
integration
Flavia Esposito
Dipartimento di Matematica, Università degli Studi di Bari Aldo Moro [email protected]

Over the years, low-rank approximation models have gained significant attention due to their
effectiveness in analyzing real data.
The key idea is that real data has a structured form (such as vectors, matrices, or tensors)
and admits a low-rank representation. A data matrix X ∈ Rn×m , with n samples and m features,
can be represented as a product of two factors W ∈ Rn×r and H ∈ Rr×m , with r < min(m, n),
such that X ≈ W H.
The problem of finding such a pair (W, H) can be mathematically formulated as a penalized
optimization task:

min Div(X, W H) + µ1 J1 (W ) + µ2 J2 (H) + µ3 J3 (W, H) (1)

W,H∈C

where Div(·, ·) is a divergence function that evaluates the quality of the approximation, C is a
feasible set that encodes structural or physical information about the data, Ji (i = 1, 2, 3)are
the penalty functions that enforce additional properties on W and H, and µi are the penalty
hyperparameters, balancing the bias-variance trade off in approximating X and satisfying factor
properties.
In this talk, we review some theoretical and computational issues related to specific low-rank
approximation models and numerical methods defined on the set C of nonnegative matrices.
We address several mathematical challenges, including the selection of an appropriate diver-
gence function tailored to the specific data domain, and the proper definition of Ji to integrate
domain-specific prior knowledge. We also emphasize real-world applications, particularly in the
biomedical and environmental fields. Morover, we also investigate how additional constraints
encoded by the peculiar form of Ji can be advantageously handled using manifold optimization
techniques.

References
[1] Gillis, N. Nonnegative matrix factorization. (SIAM, 2020)
[2] Boumal, N. An introduction to optimization on smooth manifolds. (Cambridge University
Press, 2023)

40
 A random matrix approach to Hopfield-like neural
networks: addressing generalization and overfitting
Alberto Fachechi
Department of Mathematics, Sapienza University of Rome [email protected]

Theoretical investigations in modern Artificial Intelligence focus on developing robust math-

ematical frameworks for understanding information processing in learning neural networks. A
key challenge lies in characterizing the ability of these systems to extract hidden features from
empirical data and utilize them for effective generalization. In this context, spin glass models
with structural disorder (to use the words by Marc Mézard [1]) serve as essential toy models for
exploring the functional regimes of artificial neural networks. In this talk, I will present recent
findings on example-based Hopfield-like models, which provide an ideal theoretical framework for
this purpose and naturally emerge from a statistical inference perspective. Specifically, we exploit
[2, 3] the properties of random Hebbian-like coupling matrices in these models to explore the
attractor landscape, thereby providing insights into the mechanisms underlying generalization
and overfitting. Furthermore, we extend the Marchenko-Pastur theorem to this class of random
matrices, using the resulting distribution to estimate crucial model characteristics, including the
attractive capacity of hidden ground truths.

References
[1] M. Mézard, Spin glass theory and its new challenge: structured disorder. Indian J Phys 98,
3757–3768 (2024).
[2] E. Agliari et al., "Regularization, early-stopping and dreaming: a Hopfield-like setup to
address generalization and overfitting." Neural Networks 177 (2024): 106389.

[3] E. Agliari, A. Fachechi, D. Luongo. "A spectral approach to Hebbian-like neural networks."
Applied Mathematics and Computation 474 (2024): 128689.

41
 On latent dynamics learning in nonlinear reduced order
modeling
Nicola Farenga
MOX, Department of Mathematics, Politecnico di Milano, Milan, Italy

[email protected]

Stefania Fresca, Simone Brivio, Andrea Manzoni

MOX, Department of Mathematics, Politecnico di Milano, Milan, Italy

{stefania.fresca, simone.brivio, andrea1.manzoni} @polimi.it

In this work, we present the novel mathematical framework of latent dynamics models (LDMs)
for reduced order modeling of parameterized nonlinear time-dependent PDEs. Our framework
casts this latter task as a nonlinear dimensionality reduction problem, while constraining the
latent state to evolve accordingly to an (unknown) dynamical system, namely a latent vector
ordinary differential equation (ODE). A time-continuous setting is employed to derive error and
stability estimates for the LDM approximation of the full order model (FOM) solution. We an-
alyze the impact of using an explicit Runge-Kutta scheme in the time-discrete setting, resulting
in the ∆LDM formulation, and further explore the learnable setting, ∆LDMθ , where deep neural
networks approximate the discrete LDM components, while providing a bounded approximation
error with respect to the FOM. Moreover, we extend the concept of parameterized Neural ODE
– recently proposed as a possible way to build data-driven dynamical systems with varying in-
put parameters – to be a convolutional architecture, where the input parameters information
is injected by means of an affine modulation mechanism, while designing a convolutional au-
toencoder neural network able to retain spatial-coherence, thus enhancing interpretability at
the latent level. Numerical experiments, including the Burgers’ and the advection-reaction-
diffusion equations, demonstrate the framework’s ability to obtain, in a multi-query context, a
time-continuous approximation of the FOM solution, thus being able to query the LDM approx-
imation at any given time instance while retaining a prescribed level of accuracy. Our findings
highlight the remarkable potential of the proposed LDMs, representing a mathematically rigorous
framework to enhance the accuracy and approximation capabilities of reduced order modeling
for time-dependent parameterized PDEs.

References
[1] N. Farenga, S. Fresca, S. Brivio, A. Manzoni On latent dynamics learning in nonlinear reduced
order modeling, https://arxiv.org/abs/2408.15183

42
 Mathematical Transformations and Deep Learning
Methodologies to enhance Tool Wear Monitoring using
Audio Data
Stefania Ferrisi*, Rosita Guido, Giuseppina Ambrogio
Department of Mechanical, Energetic and Management Engineering, University of Calabria, Ponte P.

Bucci, Rende, 87036, CS, Italy

*[email protected]

The integration of deep learning methodologies with Internet of Things (IoT) sensor systems
offers significant potential for real-time monitoring of tool conditions in milling processes. Tool
condition monitoring systems provide critical insights into tool wear, allowing for timely replace-
ment decisions, minimizing machine downtime, and preserving the quality of machined surfaces.
These advancements contribute to the sustainability of manufacturing operations by reducing
waste and optimizing resource utilization. Among various IoT sensors, microphones that capture
audio signals during machining have emerged as a cost-effective and non-invasive approach.
This study investigates the use of mathematical transformations for audio signals to enhance the
predictive accuracy of tool wear monitoring. This study examines two primary methods for pro-
cessing audio data: numerical feature extraction and audio conversion into spectrograms using
the Fast Fourier Transform (FFT). By decomposing complex audio waveforms into their fre-
quency components, the FFT retains essential information that characterizes the progression of
tool wear. The generated spectrograms, represented as high-resolution images, provide a detailed
depiction of frequency and amplitude variations over time. When analyzed using convolutional
neural networks, these spectrograms enable accurate classification of tool wear stages and estima-
tion of the remaining useful life of cutting tools. This methodology highlights the effectiveness
of combining rigorous mathematical signal processing techniques with artificial intelligence to
address challenges in predictive maintenance.
The findings emphasize the potential of this approach to develop robust and scalable systems
for real-time tool monitoring, aligning with the principles of modern manufacturing to improve
efficiency, reduce operational costs, and support sustainable practices.

References
[1] Ferrisi Stefania, Ambrogio Giuseppina, Guido Rosita and Umbrello Domenico. Artificial Intelli-
gence techniques and Internet of things sensors for tool condition monitoring in milling: A review
, Materials Research Proceedings, Vol. 41, pp 2000-2010, 2024
[2] Stefania Ferrisi, Gabriele Zangara, David Rodríguez Izquierdo, Danilo Lofaro, Rosita Guido,
Domenico Conforti and Giuseppina Ambrogio. Tool Condition Monitoring for milling process using
Convolutional Neural Networks , Procedia Computer Science, Vol. 232, pp 1607-1616, 2024

43
Deep orthogonal decomposition: an adaptive basis approach
to dimensionality reduction
Nicola Rares Franco1
Andrea Manzoni1 , Paolo Zunino1 , Jan S. Hesthaven2
1
MOX, Department of Mathematics, Politecnico di Milano, 20133, Milan, Italy
2
CMSS, École Polytechnique Fédérale de Lausanne, Station 8, 1015, Lausanne, Switzerland

[email protected]

Linear dimensionality reduction methods like Principal Component Analysis (PCA) and Sin-
gular Value Decomposition (SVD) are ubiquitous in statistics, machine learning, and numerical
analysis. Recently, several researchers have developed adaptive variants of these methods to
address the challenge of integrating external sources of information —such as, e.g., contextual
information or parameter dependency— within the dimensionality reduction process. We refer
to these methods as to «algorithms for parameter-dependent low-rank approximation». Such
approaches enable enhanced interpretability in statistical applications, such as extracting key
patterns in data (e.g., ECG signals, images, or audio) conditioned on covariates like age or time,
and improved performance in numerical applications, such as reduced-order modeling of PDEs
with slowly decaying Kolmogorov n-widths.
Starting from here, we present a unified theoretical framework for parametric low-rank ap-
proximations and propose Deep Orthogonal Decomposition (DOD) as a novel approach for di-
mensionality reduction in the context of reduced-order modeling of parameterized PDEs. DOD
utilizes deep neural networks to construct an adaptive local bases that can capture the structure
of the solution manifold in a dynamical manner. By combining linear and nonlinear elements,
DOD overcomes the limitations of global methods, such as POD and deep autoencoders, pro-
viding both interpretability and precise error control. We validate the effectiveness of the DOD
through numerical experiments based on the Navier-Stokes and Eikonal equations, demonstrating
its capability to address challenging scenarios, including nonlinear PDEs, intricate geometries,
and large parameter spaces. In doing so, we also explore certain connections between the DOD
and the Grassmann manifold, thanks to which we are able to develop specific diagnostic tools
that can facilitate practical implementation and analysis.
Finally, we come back to the general framework, with the purpose of deepening our under-
standing through a more abstract mathematical analysis. Specifically, we shall present some novel
theoretical results that show how the efficacy of parametric low-rank approximation algorithms
—such as the DOD— relates to certain regularity properties, which, in turn, depend on how
the eigenvalues of the covariance operator change with the problem parameters. In particular,
branching phenomena (crossing of the eigenvalues) can significantly impact model performance
and needs to be accounted for when designing and implementing these approaches.

References
[1] Franco, N. R., Manzoni, A., Zunino, P., Hesthaven, J. S. (2024). Deep orthogonal decompo-
sition: a continuously adaptive data-driven approach to model order reduction, arXiv preprint
arXiv:2404.18841
[2] Franco, N. R. (2024). Measurability and continuity of parametric low-rank approximation in Hilbert
spaces: linear operators and random variables, arXiv preprint arXiv:2409.09102
[3] Gupta, A., Barbu, A. (2018). Parameterized principal component analysis, Pattern Recognition,
78, 215-227.
[4] Amsallem, D., Farhat, C. (2011). An online method for interpolating linear parametric reduced-order
models, SIAM Journal on Scientific Computing, 33(5), 2169-2198.

44
 Penalized Maximum Likelihood and Loss Minimization for
Classification
Bharath Krishnan Girishkumar
PhD student, MIDA group, Department of Mathematics, University of Genova

[email protected]

Federico Benvenuto
Associate Professor, MIDA group, Department of Mathematics, University of Genova

[email protected]

This talk explores the parallelism between empirical loss minimization and binary classifi-
cation as a maximum likelihood problem with data drawn from a Bernoulli distribution. We
demonstrate that empirical loss minimization corresponds to penalized maximum likelihood es-
timation, where the penalty depends on the specific loss function. Furthermore, we establish
a one-to-one correspondence between solutions of different loss functions via generalized linear
model link functions. Remarkably, the resulting binary classifiers remain identical across the con-
sidered loss functions. We also show that the classification problem can be solved numerically
using linear equations. However, due to potential ill conditioning in the case of square systems,
iterative algorithms are often more effective. Finally, we extend these concepts to multiclass
classification and present supporting numerical experiments.

45
 Learning Passive Left Ventricular Mechanics via Shape
Encoding Neural Networks
Marc Hirschvogela1 , Davide Carraraa , Stefano Pagania , Simone
Pezzutob , Francesco Regazzonia
a
MOX, Dipartimento di Matematica, Politecnico di Milano, Milan, Italy
b
Dipartimento di Matematica, Università di Trento, Trento, Italy

We present a novel scientific machine learning approach to predict the solution of partial
differential equations on unseen domains. The methodology consists of a two-step procedure:
First, the DeepSDF [1] neural network architecture is used to learn a signed-distance function
(SDF) that is representative of the object’s shape. Second, a fully connected neural network is
trained with PDE solutions on different geometries, leveraging a latent vector that encodes shape
information from the prior SDF training step [2]. The approach, in general, only requires a point
cloud representation of the geometry, hence neither meshes nor any type of point-to-point cor-
respondence between domains is needed. We test our approach for inferring anisotropic passive
mechanics on left ventricular patient-specific and synthetically generated geometries, investigat-
ing alternative shape encoding via principal component analysis or input feature enhancement by
universal ventricular coordinates. Our results highlight the potential of shape codes for surrogat-
ing nonlinear PDEs on a diverse cohort of ventricles and pave the way for real-time predictions
of multi-physics phenomena such as cardiac electromechanics on complex geometries.
The present research has been supported by the project PRIN2022, MUR, founded by the
European Union (grant P2022N5ZNP).

References
[1] Park, Jeong Joon and Florence, Peter and Straub, Julian and Newcombe, Richard and Lovegrove,
Steven DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation, 2019
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019
[2] Regazzoni, Francesco and Pagani, Stefano and Quarteroni, Alfio Universal Solution Manifold Net-
works (USM-Nets): Non-Intrusive Mesh-Free Surrogate Models for Problems in Variable Domains,
Journal of Biomechanical Engineering, 144(12), 2022

1 [email protected]

46
 Data-driven parameterization for adaptive spline model
reconstruction
Sofia Imperatore
IMATI-CNR “Enrico Magenes”, via Ferrata 5/A, 27100, Pavia, Italy [email protected]

Carlotta Giannelli, Angelos Mantzaflaris, Dominik Mokrivs, Felix

Scholz
Dipartimento di Matematica e Informatica “Ulisse Dini”, Università degli Studi di Firenze, Viale

Morgani 67/A, Firenze, 50137, Italy [email protected]

Inria Centre at Université Côte d’Azur, 2004 Route des Lucioles, 06902 Sophia Antipolis, France

[email protected]
MTU Aero Engines AG, Dachauer Strasse 665, Munich, 80995, Germany [email protected],

[email protected]

In this talk, we combine Computer Aided Geometric Design (CAGD) methods with Deep
Learning (DL) technologies. The final objective is the (re-)construction of highly accurate CAD
models for the design of complex data-driven free-form adaptive spline geometries. In particu-
lar, we present two novel geometric deep learning techniques for parameterizing scattered point
clouds in R3 on a planar parametric domain, by exploiting (graph) convolutional neural networks.
Firstly, we introduce a data-driven parameterization model that builds upon existing meshless
parameterization schemes and predicts the parametric values of the input point cloud from the
proximity information of its 3D items and its dual line graph [1]. Secondly, we present an al-
ternative learning model, that avoids line-graph computation, characterized by a new boundary
informed message-passing input layer, that takes in input boundary conditions and propagates
them into the new features of the interior points [2]. Finally, we show the effectiveness of these
learning models for surface fitting with adaptive spline constructions and moving parameteriza-
tion, thus merging CAGD methods with DL technologies [3].

References
[1] Giannelli, C., Imperatore, S., Mantzaflaris, A., and Scholz, F. Learning meshless parameterization
with graph convolutional neural networks, In International conference on WorldS4 (pp. 375–387).
Singapore: Springer Nature Singapore, 2023.
[2] Giannelli, C., Imperatore, S., Mantzaflaris, A., and Scholz, F. BIDGCN: boundary-informed dy-
namic graph convolutional network for adaptive spline fitting of scattered data, Neural Computing
and Applications, 1–24, 2024
[3] Giannelli, C., Imperatore, S., Mantzaflaris, A., and Mokrivs, D. Leveraging moving parameterization
and adaptive THB-splines for CAD surface reconstruction of aircraft engine components In Smart
Tools and Applications in Graphics-Eurographics Italian Chapter Conference. The Eurographics
Association, 2023

47
 A new mathematical model to analyze the spread of
misinformation on Social Media
Samira Iscaro, Dajana Conte, Giovanni Pagano and Beatrice
Paternoster
Dept. of Mathematics, University of Salerno, Fisciano (SA), Italy

[email protected], [email protected], [email protected], [email protected]

In recent years, research activity on the mathematical analysis of evolutionary problems

has been focused on a fundamental problem of our society: the spread out of fake information
on Social Media. Since the outcomes of many political or social events have been influenced
by fake news [5], researches have developed different approaches to describe and analyze the
aformentioned phenomenon, from the use of machine learning based detectors for fake information
to the use of mathematical epidemiological models to make an attempt of prediction of the
evolution of news spread through time [1, 4]. In particular, in [2], it has been shown that starting
from a dataset composed of real data extracted from X (Twitter) and using an epidemiological
model of SIR type, called the Ignorant-Spreader-Recovered model, it is possible to compute
optimized parameters to make predictions on the spread of a certain news in terms of both the
total number of individuals who share news and the moment of the peak of interest towards it.
The main aim of this talk is to focus on a new mathematical model to analyze the spread of
fake news on Social Media. More specifically, we will focus on a class of mathematical models
called the Ignorant-Spreader-Counter Spreader-Recovered (ISCR), initially proposed in [6], and
introduced to accurately describe the scenario in which there is a debate regarding a certain topic.
In general, this class of models is suited to describe the situation in which two different groups of
spreders are present: the one of the individuals that share the fake news (the Spreader class) and
the one of individuals who try to restore the truth (the Counter Spreader class). Mathematical
methods to numerically solve this kind of problem will also be introduced. In fact, we will show
that using an ISCR model it is possible to predict the evolution of certain fake news, spread on X
in recent years. However, in order to reduce the computational cost of the fitting phase as well as
to preserve the positivity of the model, starting from strategies described in [3], we will introduce
a Nonstandard Finite Difference (NSFD) scheme for the considered model, showing that it is
truly applicable to reality to make predictions, as confirmed by several numerical experiments.

References
[1] Cardone, A., Diaz de Alba, P., Paternoster, B. Analytical Properties and Numerical Preservation
of an Age-Group Susceptible-Infected-Recovered Model: Application to the Diffusion of Information.
J. of Comput.l and Nonl.r Dyn. 19.6. https://doi.org/10.1115/1.4065437 (2024).
[2] Castiello, M., Conte, D., Iscaro, S. Using Epidemiological Models to Predict the Spread of Informa-
tion on Twitter. Algorithms. 16, 391. https://doi.org/10.3390/a16080391 (2023).
[3] Conte, D., Guarino, N., Pagano, G., Paternoster, B. Positivity-preserving and elementary stable
nonstandard method for a COVID-19 SIR model. Dolomites Research Notes on Approximation,
15(DRNA Volume 15.5), 65-77. (2022).
[4] D’Ambrosio, R., Giordano, G., Mottola, S., Paternoster, B. Stiffness analysis to predict the spread
out of fake information. Future Internet, 13(9), 222. (2021).
[5] Maleki, M., Mead, E., Arani, M., Agarwal, N. Using an epidemiological model to study the spread of
misinformation during the Black Lives Matter Movement. arXiv. https://arxiv.org/abs/2103.
12191 (2021).
[6] Muhlmeyer, M., Agarwal, S. Information spread in a social media age. In Modelling and Control.
CRC Press: Boca Raton, FL, USA; Taylor and Francis Group: Boca Raton, FL, USA; London,
UK; New York, NY, USA. (2021).

48
 Constructing Interpretable Prediction Models with
Semi-Orthogonal 1D DNNs: An Example in Irregular ECG
Classification
Giacomo Lancia
Università di Roma "La Sapienza"; Dipartimento di Scienze di Base Applicate all’Ingegneria (SBAI);
Via Antonio Scarpa, 16; Roma [email protected]

Cristian Spitoni
Universiteit Utrecht; Mathematics Department; Budapestlaan, 6; Utrecht [email protected]

When considering deep learning (DL)–based applications in medicine, ensuring interpretabil-

ity is crucial for assessing the quality and safety of clinical predictions. At the same time, the
increasing availability of massive, complex, high-dimensional data has heightened the demand
for computational models capable of extracting salient features effectively and accurately.
Artificial Neural Network (ANN) based methods have emerged as powerful tools for achieving
highly accurate predictions across diverse medical domains , often outperforming traditional ap-
proaches. However, the inherent "black-box" nature of ANN models limits their interpretability,
making it challenging to establish causal relationships between input covariates and predictions.
As such, integrating ANN-based methods within interpretable frameworks is essential for ensur-
ing model transparency and clinical reliability.
To address this challenge, we propose a novel methodology that enforces a mathematical
structure within the neural network architecture. Specifically, we incorporate Semi-Orthogonal
constraints on convolutional kernels to ensure the invertibility of the learned features, thereby
enhancing the interpretability of the convolutional neural network’s (CNN) actions. This con-
straint enables a more transparent understanding of the transformations applied to the data,
bridging the gap between predictive accuracy and explainability.
Saliency Map (SM)-based methods and, more broadly, eXplainable Artificial Intelligence
(XAI) algorithms have been proposed to visualize and interpret the areas of input data where
an ANN focuses its attention during decision-making.These methods aim to uncover the con-
nections between outputs and inputs by analyzing the network’s propagation rules. While SM-
based approaches, such as Vanilla Gradient , CAM , and Layer-Wise Relevance Propagation ,
have demonstrated utility, they often fail to fully reveal causal relationships between inputs and
predictions. Instead, they primarily highlight regions of interest, which may lead to misleading
conclusions in clinical settings.
Building upon these foundations, our methodology leverages Semi-Orthogonal convolutional
weights to enhance the interpretability of 1-D Deep Neural Networks (1D DNN). Our approach
allows for the efficient reconstruction of 1D CNN inputs, providing deeper insights into how
features are extracted and contributing to more interpretable feature representations. These
extracted features are then incorporated into a Logistic Regression (LR) model, a simple yet
interpretable framework, to classify irregular ECG patterns.
The interpretability of this framework is further evaluated through feature extraction and
alignment with clinical knowledge. The proposed methodology ensures that the predictions not
only achieve high accuracy but also maintain a transparent decision-making process, crucial for
identifying conditions such as Atrial Fibrillation (AF), Myocardial Infarction (MI), and Sinus
Bradycardia (SBR).
Through this mathematically grounded approach, our research bridges the gap between ad-
vanced deep learning techniques and the need for interpretable, clinically reliable predictions
in medical applications. We demonstrate that Semi-Orthogonal constraints significantly im-
prove the explainability of 1D CNN-based predictions, setting a precedent for developing ro-
bust and interpretable diagnostic models. A pre-print of this work is available at https:
//arxiv.org/abs/2410.12059.

49
 Hexagonal Grid-Based Reinforcement Learning
Environments for Marine Biodiversity Monitoring
Giulia Lombardi
University of Trento, Via Sommarive 14, I-38123 Povo (Trento) [email protected]

Monitoring marine biodiversity presents profound scientific and logistical challenges, necessi-

tating the integration of advanced mathematical frameworks, computational methodologies, and

interdisciplinary expertise. These challenges are further compounded by the vast spatial scales
and ecological complexity of marine environments, necessitating a rigorous and multifaceted
approach to data acquisition and environmental monitoring [1].
Contemporary research leverages an array of technological modalities, including satellite
imagery, remote sensing, autonomous underwater vehicles (AUVs), remotely operated vehi-
cles (ROVs), environmental DNA (eDNA) sampling, and traditional vessel-based expeditions
[2, 3]. While these methodologies yield valuable insights, they are frequently constrained by
logistical limitations, financial burdens, and operational inflexibility. Moreover, the intensify-
ing impacts of climate change—manifested in dynamic ocean conditions and shifting species
distributions—exacerbate the demand for continuous, high-resolution monitoring systems.
To address these pressing challenges, we propose a novel framework for local data collection
and monitoring, specifically designed to optimize AUV navigation. Central to our approach is the
development of a reinforcement learning (RL) hexagonal grid-world environment which enables
the agent to navigate under conditions of partial observability. This model exploits the geomet-
ric properties of hexagonal tiling — notably uniform neighbor connectivity and minimized edge
effects — to enhance spatial coverage and enable realistic navigation in complex marine settings.
This design outperforms traditional Cartesian grid-based systems, which are susceptible to ineffi-
ciencies in irregular or dynamic environments. In fact, similar methodologies have demonstrated
success in terrestrial [5] and space-based [5] applications, yet remain underexplored in ecological
modeling and monitoring [6].
The proposed approach addresses diverse ecological and operational needs by allowing the
customization of multi-objective reward functions to suit specific monitoring tasks. This adapt-
ability allows researchers to leverage the same framework to pursue diverse objectives, such as
accurately identifying biodiversity hotspots, monitoring rare or endangered species, and detecting
plastic debris, ultimately offering a powerful AI-driven tool for advancing marine conservation
and environmental research.

References
[1] N. Bax et al. Seascape ecology: identifying research priorities for an emerging ocean sustainability
science, Marine Ecology Progress Series, 663, 1–29, 2019.
[2] A. Miller, J. I. Virmani. Advanced marine technologies for ocean research, Deep Sea Research Part
II: Topical Studies in Oceanography, Volume 212, 105340, 2023, ISSN 0967-0645.
[3] A. Miller, J. I. Virmani. Advanced marine technologies for ocean research, Deep Sea Research Part
II: Topical Studies in Oceanography, Volume 212, 105340, 2023, ISSN 0967-0645.
[4] Uber Technologies Inc. Hexagonal Grids in Urban Mobility Optimization: Applications to Uber,
Available at: https://www.uber.com/en-GB/blog/h3/.
[5] W. Wang, H. Zhou, S. Zheng, G. L"u, and L. Zhou. Ocean surface currents estimated from satellite
remote sensing data based on a global hexagonal grid, International Journal of Digital Earth, 16:1,
1073–1093, 2023.
[6] C.P.D. Birch, S.P. Oom, J.A. Beecham. Rectangular and hexagonal grids used for observation,
experiment and simulation in ecology, Ecological Modelling, 206(2007), 347–359.

50
 Multi-fidelity reduced-order surrogate modelling
Andrea Manzoni, Paolo Conti
MOX – Department of Mathematics, Politecnico di Milano, Italy
[email protected], [email protected]

Mengwu Guo
Centre for Mathematical Sciences, Lund University, Sweden
[email protected]

Attilio Frangi
Department of Civil and Environmental Engineering, Politecnico di Milano
[email protected]

Steven L. Brunton, J. Nathan Kutz

University of Washington, Seattle, United States
[email protected], [email protected]

High-fidelity numerical simulations of partial differential equations (PDEs) given a restricted

computational budget can significantly limit the number of parameter configurations considered
and/or time window evaluated. Multi-fidelity surrogate modeling aims to leverage less accurate,
lower-fidelity models that are computationally inexpensive in order to enhance predictive accu-
racy when high-fidelity data are scarce [4]. However, low-fidelity models, while often displaying
the qualitative solution behavior, fail to accurately capture fine spatio-temporal and dynamic
features of high-fidelity models.
To address this shortcoming, we present a data-driven strategy that combines dimensionality
reduction with multi- fidelity neural network surrogates [1]. The key idea is to generate a spatial
basis by applying proper orthogonal decomposition (POD) to high-fidelity solution snapshots,
and approximate the dynamics of the reduced states – time-parameter-dependent expansion
coefficients of the POD basis – using a multi-fidelity long short-term memory network [2, 3].
By mapping low-fidelity reduced states to their high-fidelity counterpart, the proposed reduced-
order surrogate model enables the efficient recovery of full solution fields over time and parameter
variations in a non-intrusive manner. A further extension to the case of multiple data sources,
with low-fidelity models of different type, is also considered, in the spirit of progressive learning
from multiple sources.
The generality of the proposed approach is demonstrated by a collection of PDE problems
where the low-fidelity model can be defined by coarser meshes and/or time stepping, as well as
by misspecified physical features.

References
[1] P. Conti, M. Guo, A. Manzoni, A. Frangi, S. L. Brunton, and J. Nathan Kutz. Multi-fidelity
reduced-order surrogate modelling. Proceedings of the Royal Society A, 480(2283):20230655, 2024.
[2] P. Conti, M. Guo, A. Manzoni, and J. S. Hesthaven. Multi-fidelity surrogate modeling using long
short-term memory networks. Computer methods in applied mechanics and engineering, 404:115811,
2023.
[3] M. Guo, A. Manzoni, M. Amendt, P. Conti, and J. S. Hesthaven. Multi-fidelity regression us-
ing artificial neural networks: Efficient approximation of parameter-dependent output quantities.
Computer methods in applied mechanics and engineering, 389:114378, 2022.
[4] M. Torzoni, A. Manzoni, and S. Mariani. A multi-fidelity surrogate model for structural health
monitoring exploiting model order reduction and artificial neural networks. Mechanical Systems
and Signal Processing, 197:110376, 2023.

51
 Convergence of quantum neural networks at infinite width
Anderson Melchor Hernandez
Piazza di Porta S. Donato, 5, Bologna, BO [email protected]

Filippo Girardi
Piazza dei Cavalieri, 7, Pisa, PI [email protected]

Giacomo De Palma
Piazza di Porta S. Donato, 5, Bologna, BO [email protected]

Davide Pastorello
Piazza di Porta S. Donato, 5, Bologna, BO [email protected]

Quantum neural networks constitute the quantum version of deep neural models. These new
models are based on quantum circuits and generate functions given by the expectation values of
a quantum observable measured on the output of a quantum circuit made by parametric one-
qubit and two-qubit gates [4]. The parameters of the circuit encode both the input data and the
parameters of the model itself. These parameters are typically optimized by gradient descent,
which involves iterative adjustment to minimize a cost function and improve the performance
of the quantum circuit in the processing and analysis of data [1]. Significant progress has been
made in addressing the question of whether training can perfectly fit the training examples
while simultaneously avoiding overfitting. A fundamental breakthrough has been the proof that,
in the limit of infinite width, the probability distribution of the function generated by a deep
neural network trained on a supervised learning problem converges to a Gaussian process [2].
This recent has inspired renewed interest in quantum machine learning, raising the question
of whether quantum neural networks exhibit analogous properties. In this presentation, I will
explore some of the recent advancements in this area, highlighting key insights and findings [3].

References
[1] F. Girardi, G. De Palma, Trained quantum neural networks are Gaussian processes,
arXiv:2402.08726 (2024).
[2] H. Boris, Which neural net architectures give rise to exploding and vanishing gradients? J. Adv
Neural Inf Process Syst 31 (2018).
[3] A. Melchor Hernandez, F. Girardi, G. De Palma, D. Pastorello, Quantitative conver-
gence of trained quantum neural networks to a Gaussian process, Preprint.
[4] M. Schuld, I. Sinayskiy, F. Petruccione, An introduction to quantum machine learning, J.
Contemporary Physics 56 (2015) no. 2.

52
 An all-around perspective on hybrid coupled models and
parameter calibration for collective cell dynamics
Marta Menci
Università Campus Bio-Medico di Roma, [email protected]

The study of collective dynamics has garnered significant interest across various scientific
domains due to its potential to model self-organization in complex systems and its wide range
of applications. In the biological and biomedical world, an increasing number of phenomena
benefits from the mathematical and numerical approach, aiming at in-silico models to inves-
tigate the phenomena of interest. In this field, collective cell dynamics play a critical role in
several biological processes characterizing the human body. The main feature of those kind of
collective behaviors, that need to be taken into account in the mathematical models, is that cells
not only interact mechanically, but are also driven by chemical signals which lead cells moving
towards higher concentrations of chemicals. In real applications, parameter estimation can be
exceptionally challenging due to the large number of parameters that need to be simultaneously
estimated and the costs of performing experiments to collect experimental data. To this end, ma-
chine learning algorithms are currently investigated, allowing for faster and robust optimization
procedures for solving inverse problems associated with parameter estimation.
The talk will explore a recent class of multiscale hybrid coupled models to simulate migrations
of cells in different scenarios [1, 2]. Originally conceived to model embriogenesis processes,
the particular structure combine discrete cellular dynamics with continuous chemical signaling,
offering a multiscale framework to describe the complex interactions between cells and their
environment.
Although hybrid models provide an accurate description of cell behaviors, they can be com-
putationally expensive, especially when dealing with large numbers of cells in higher-dimensional
settings. To address this challenge, a macroscopic pressureless Euler-type model with nonlocal
chemotaxis has been rigorously derived from the microscopic scale, describing cellular dynamics
in term of evolution of a cell density, hence on a macroscopic scale [3, 4].
Numerical simulations of the considered models at different scales will be presented, including
2D and 3D scenarios. In particular, the hybrid coupled model is validated against experimental
data (positions and velocities of cells acquired at different times during the experiments), whereas
the macroscopic-derived version makes use of synthetic data generated from original microscopic
real-data.
This work is based on ongoing collaborations with Roberto Natalini (Instituto per le Appli-
cazioni del Calcolo - CNR), Thierry Paul (LYSM - CNRS) and Tommaso Tenna (Université Côte
D’azur).

References
[1] E. Di Costanzo, M. Menci, E. Messina, R. Natalini and A. Vecchio A hybrid model of collective
motion of discrete particles under alignment and continuum chemotaxis , Discrete & Continuous
Dynamical Systems-B, 25(1), 2020.
[2] G. Bretti, E. Campanile, M. Menci, R. Natalini A scenario-based study on hybrid PDE-ODE model
for Cancer-on-chip experiment , In: Problems in Mathematical Biophysics: A Volume in Memory
of Alberto Gandolfi. Cham: Springer Nature Switzerland, 2024.
[3] R. Natalini and T. Paul. The mean-field limit for hybrid models of collective motions with chemo-
taxis, SIAM Journal on Mathematical Analysis, 55(2), 2023.
[4] M. Menci, R. Natalini, T. Paul. Microscopic, kinetic and hydrodynamic models of collective motions
with chemotaxis: a numerical study , Mathematics and Mechanics of Complex Systems, 12(1), 2024.

53
 Step-by-Step Time-Discrete Physics Informed Neural
Networks for PDEs models
Giovanni Pagano
Department of Mathematics, University of Salerno, Italy, [email protected]

1
C. Valentino, 2 D. Conte, 2 B. Paternoster, 1 F. Colace, 3 M. Casillo
1
Department of Industrial Engineering, University of Salerno, Italy, {cvalentino,fcolace}@unisa.it
2
Department of Mathematics, University of Salerno, Italy, {dajconte,beapat}@unisa.it
3
Department of Cultural Heritage Sciences, University of Salerno, Italy, [email protected]

Models based on Partial Differential Equations (PDEs) originate from different phenomena,
such as: life cycle of batteries [1], evolution of vegetation [3], corrosion of materials [5], production
of renewable energy [1]. For the related numerical solution, in addition to standard well-known
methods, several techniques based on Artificial Neural Networks (ANNs) have recently been
proposed, see e.g. [4]. In this context, the so-called Physics-Informed Neural Networks (PINNs)
are considered, i.e. ANNs generally constructed in such a way as to compute a time-continuous
and space-continuous approximation of the exact solution of the analyzed PDE.
This talk focuses on the derivation of a new approach based on PINNs, namely Time-Discrete
PINNs, for the solution of PDEs. They are called this way since provide a solution which is
continuous in space and discrete in time. Existing Time-Discrete PINNs from the literature are
based on the immersion of classical Runge-Kutta (RK) methods within ANNs. That is, given
every point of the spatial domain, the neural network is constructed in such a way as to furnish,
as output, approximations of the stages of the selected RK method at a fixed time step.
Here, we propose new Step-by-Step (SBS) Time-Discrete PINNs, based on the implicit Euler
and Crank-Nicolson methods [1]. We construct these PINNs in such a way as to obtain, as output,
an approximation of the solution by the above-mentioned methods at each time step (unlike RK-
based PINNs). Furthermore, we establish connections between the existing RK-based and the
new SBS PINNs, which allow to use the same workflow for both in implementation. Several
numerical experiments, conducted on PDEs models related to sustainability [1] and life cycle of
batteries [2], show the advantages of the new SBS PINN over the RK-based ones, and also over
classical continuous-time and continuous-space PINNs.
Acknowledgements: this work has been supported by the PRIN PNRR 2022 project
P20228C2PP “BAT-MEN”.

References
[1] M. Frittelli, B. Bozzini, I. Sgura. Turing patterns in a 3D morpho-chemical bulk-surface reaction-
diffusion system for battery modeling. MinE (Mathematics in Engineering) 6(2), 363-393 (2024).
[2] D. Conte, G. Pagano, B. Paternoster. Nonstandard finite differences numerical methods for a
vegetation reaction–diffusion model. J. Comput. Appl. Math. 419 (2023).
[3] G. Frasca-Caccia, C. Valentino, F. Colace, D. Conte. An overview of differential models for corrosion
of cultural heritage artefacts. Math. Model. Nat. Phenom. 18, 27 (2023).
[4] M. Raissi, P. Perdikaris, G. E. Karniadakis. Physics-informed neural networks: A deep learning
framework for solving forward and inverse problems involving nonlinear partial differential equations.
J. Comput. Phys. 378, 686–707 (2019).
[5] C. Valentino, G. Pagano, D. Conte, B. Paternoster, F. Colace, M. Casillo. Step-by-step time discrete
Physics Informed Neural Networks with application to a sustainability PDE model. Math. Comput.
Simul., doi.org/10.1016/j.matcom.2024.10.043 (2024).
[6] C. Valentino, G. Pagano, D. Conte, B. Paternoster, F. Colace. Physics Informed Neural Networks
for a Lithium-ion batteries model: a case of study. Submitted.

54
 Training a quantum GAN with classical data
Davide Pastorello*, Giacomo De Palma* and Tristan Klein†
*Dept. of Mathematics, University of Bologna, Piazza di Porta San Donato 5, 40126 Bologna, IT
†
ENS de Lyon, Département Informatique, 15 parvis René Descartes 69342 Lyon Cedex 07, France
[email protected], [email protected] , [email protected]

Quantum neural networks (QNNs) are defined by parametric quantum circuits which can
be trained by backpropagation in analogy to classical feedforward neural networks. Parametric
circuits can be applied to construct generators and discriminators within the quantum version of
generative adversarial networks (GANs). In quantum generative adversarial networks (QGANs),
the generator can be implemented using a series of quantum gates that manipulate the quantum
state of a set of qubits and it is designed to generate data resembling those from the training
dataset. The discriminator is also implemented as a quantum circuit, this circuit evaluates the
likelihood of the data generated by the generator, comparing it with the real data from the
training set. The loss function used to train a QGAN is often defined using quantum concepts,
such as quantum state overlap or quantum divergence, rather than traditional loss metrics like
cross-entropy. During the training, the parameters of the generator and discriminator quantum
circuits are optimized using variational algorithms within an adversarial framework. In the
quantum architecture, the training set is made by quantum states, which may encode classical
data, assumed to be stored in a quantum memory.
In [1], we considered the so-called shadow protocol that is a procedure to construct classical
estimates of quantum states, called classical shadows, by means of measurements and quan-
tum/classical processing. The classical shadow is computed classically and stored as classical
information and used to efficiently estimates expectation values of observables [2]. Moreover, for
any n-qubit quantum state ρ, the computation of a number of classical shadows that is loga-
rithmic in n provides an accurate estimate of ρ w.r.t. the local quantum Wasserstein distance of
order 1 that is a notion from the quantum optimal mass transport [1]. This distance is a measure
of distinguishability between quantum states of a n-qubit system and it can be used to evaluate
the convergence of the shadow protocol.
The accuracy in estimating a quantum state with classical shadows in this metric has a
remarkable consequence in the training of a QGAN [1]. Considering a QGAN where the dis-
criminator generates a classical estimate of the true state constructed as the empirical mean of
O(log n) classical shadows, as proved in [1], no more copies of the true state will be needed and
the information contained in its classical shadow will be sufficient. The generator and the dis-
criminator are trained against each other in the adversarial scenario, and the expectation value
of the discriminator observable on the true state is estimated via its classical estimate without
needing further copies of the true state. After enough iterations, the generated state will be close
to the classical shadow of the true state in the local quantum Wasserstein distance of order 1.
As a consequence, a QGAN can be equivalently trained over classical shadows in place of true
quantum states, if no prior information about the state is available.
In this talk we introduce the notion of the local quantum Wasserstein distance of order 1 as
a tool in quantum optimal mass transport, its role in quantifying the convergence of the shadow
protocol and how a QGAN can be trained by classical data estimating the quantum states of the
training set in terms of classical shadows.

References
[1] De Palma, G., Klein, T., Pastorello, D. Classical shadows meet quantum optimal mass transport.
Journal of Mathematical Physics. 65, 092201 (2024)
[2] Huang, H. Y., Kueng, R.,Preskill, J. Predicting many properties of a quantum system from very few
measurements. Nature Physics 16, 1050-1057 (2020).

55
 Linesearch-Enhanced Forward-Backward Methods for
Inexact Nonconvex Scenarios
Danilo Pezzi
[email protected]

Silvia Bonettini
[email protected]
Giorgia Franchini
[email protected]

Marco Prato
[email protected]
Via Campi 213/B, Modena

In recent times, optimization techniques have been widely applied to imaging problems, lead-
ing to increasingly sophisticated variational models in current research. Significant advancement
from previous state-of-the-art methods have been achieved by considering nonconvex settings
and combining machine learning strategies with the classical variational techniques. In this talk
we introduce a forward-backward framework aimed at the minimization of an objective function
composed of a differentiable term and a convex, non differentiable one. The scheme is able to
handle two different challenges that can be presented by the objective function. On one hand,
even if the differentiable part of the function may be non-convex, the method is is able to achieve
convergence to a stationary point. On the other hand, only partial knowledge of the function
is required. Indeed, all the key steps of the method can be performed inexactly. As this is a
general scheme, it can incoporate a variety of algorithms for different problems. Here we present
an application in the realm of bilevel optimization for imaging problems, where the scope is to
combine classical variational techniques with machine learning approaches to improve the quality
of the reconstructed images. The numerical experience shows that the method is competitive
with other existing approaches [1][2].

References
[1] Pedregosa, Fabian Hyperparameter Optimization with Approximate Gradient, International Con-
ference on Machine Learning, 2016
[2] Suonperä, Ensio and Valkonen, Tuomo Linearly convergent bilevel optimization with single-step
inner methods, Computational Optimization and Applications, 2023

56
 The Neural Approximated Virtual Element Method on
general polygons
Moreno Pintore
Laboratoire Jacques-Louis Lions, Sorbonne Université, INRIA, 4 place Jussieu, 75005 Paris, France

[email protected]

Stefano Berrone, Gioana Teora

Dipartimento di Scienze Matematiche "G. L. Lagrange", Politecnico di Torino, Corso Duca degli

Abruzzi 24, 10129 Turin, Italy [email protected], [email protected]

In the Scientific Machine Learning framework, numerous new methods to solve engineering
problems have been proposed in the last few years. Such methods combine the accuracy and
stability of classical numerical methods with the efficiency and adaptability of machine learning
techniques. The Neural Approximated Virtual Element Method (NAVEM) perfectly fits in this
context, since it is a method inspired by the Virtual Element Method (VEM) [1], with which
shares some features, and that heavily relies on the nonlinear approximation properties of deep
neural networks.
The VEM is a numerical method used to solve partial differential equations using meshes
comprising very general elements and basis functions that are not known in closed form. The
idea of the NAVEM is to use the same meshes and to explicitly approximate the VEM basis
functions through one or more neural networks. This approximation leads to a completely
different method, that does not include projection or stabilization operators, but that relies on
an offline-online splitting.
The NAVEM has been firstly introduced in [2] and then extended in [3] to more general two-
dimensional meshes. In this presentation we focus on this second formulation, characterized by an
approximation of the VEM basis functions through a novel set of harmonic functions. This choice
is crucial in order to accurately approximate the VEM basis functions while reducing spurious
oscillations that may characterize the output of a standard neural network. We also present the
architecture of the involved neural networks and we theoretically discuss their approximation
properties. We propose several numerical results to illustrate the performances of the method
on different meshes and on different problems.

References
[1] L. Beirão da Veiga, F. Brezzi, A. Cangiani, G. Manzini, and A. Russo Basic principles of Virtual
Element Methods, Mathematical Models and Methods in Applied Sciences, vol. 23, no. 01, pp.
199–214, 2013.
[2] S. Berrone, D. Oberto, M. Pintore, and G. Teora The lowest-order neural approximated virtual ele-
ment method, ENUMATH 2023, Accepted. [Online]. Available: https://arxiv.org/abs/2311.18534.
[3] S. Berrone, M. Pintore, and G. Teora The lowest-order neural approximated virtual element method,
ArXiv preprint arXiv:2409.15917, 2024.

57
 Grokking as an entanglement transition in tensor network
machine learning
Domenico Pomarico
National Institute for Nuclear Physics, Bari [email protected]

Alfonso Monaco, Giuseppe Magnifico, Antonio Lacalamita,

Ester Pantaleo, Loredana Bellantuono, Sabina Tangaro,
Tommaso Maggipinto, Marianna La Rocca, Nicola Amoroso,
Sebastiano Stramaglia, Roberto Bellotti
National Institute for Nuclear Physics, Bari [email protected]

Generalizability is a fundamental property for machine learning algorithms, detected by a

grokking transition during training dynamics [1]. In the quantum-inspired machine learning
framework we numerically prove that a quantum many-body system shows an entanglement
transition corresponding to a performances improvement in binary classification of unseen data.
Two datasets are considered as use case scenarios, namely fashion MNIST and genes expression
communities of hepatocellular carcinoma. The measurement of qubits magnetization and corre-
lations is included in the matrix product state (MPS) simulation [2], in order to define meaningful
genes subcommunities, verified by means of enrichment procedures.

References
[1] Kenzo Clauw, Sebastiano Stramaglia and Daniele Marinazzo, Information-Theoretic Progress Mea-
sures reveal Grokking is an Emergent Phase Transition, arXiv:2408.08944 (2024)
[2] E. Miles Stoudenmire and David J. Schwab, Supervised Learning with Quantum-Inspired Tensor
Networks, arXiv:1605.05775 (2017)

58
 A CNN-LSTM approach for parameter estimation for
lithium metal battery cycling model
Maria Grazia Quarta, Ivonne Sgura
Department of Mathematics and Physics “Ennio De Giorgi”, University of Salento, via per Arnesano,

73100 Lecce (LE), Italy

[email protected], [email protected]
Benedetto Bozzini
Department of Energy, Politecnico di Milano, Via Lambruschi 4, 20156 Milano, Italy

[email protected]

Raquel Barreira
Instituto Politécnico de Setúbal, Escola Superior de Tecnologia de Setúbal Campus do IPS Estefanilha,

2914-508 Setúbal, Portugal [email protected]

Symmetric coin cell cycling is an important tool for the analysis of battery materials, en-
abling the study of electrode/electrolyte systems under realistic operating conditions. Moreover,
understanding the behavior of metal anodes in batteries and accurately predicting their perfor-
mance is a challenge due to the methodological gap between theoretical models and experimental
observations. In order to address this challenge, a PDE model describing the voltage profiles
behavior of symmetrical coin cells testing the Galvanostatic Discharge-Charge (GDC) protocol
has been developed [1, 2]
In this talk, based on [3], we propose a hybrid architecture of Convolutional Neural Network
and Long-Short Term Memory layers (CNN-LSTM) to estimate some relevant physico-chemical
parameters in the PDE system that describe GDC cycling of Li/Li symmetric cells. Our results
show the neural network ability to capture characteristics of voltage profiles, such as peak and
valley, saddle points, and concavity variations [1], that other traditional methods, such as Least
Squares (LS) fitting, may overlook. Moreover, our Deep Learning algorithm can successfully
estimate parameters also for experimental discharge-charge time series data. These results high-
light the robustness of our approach, which allows us to bridge the gap between theory and
experiments.

References
[1] F. Rossi, L. Mancini, I. Sgura, M. Boniardi, A. Casaroli, A.P. Kao, B. Bozzini, Insight into the
Cycling Behaviour of Metal Anodes, Enabled by X-ray Tomography and Mathematical Modelling,
ChemElectroChem 9, 2022.
[2] B. Bozzini, E. Emanuele, J. Strada, I. Sgura, Mathematical modelling and parameter classification
enable understanding of dynamic shape-change issues adversely affecting high energy-density battery
metal anodes, Applications in Engineering Science 13, 100125, 2023.
[3] M.G. Quarta, I. Sgura, E. Emanuele, J. Strada, R. Barreira, B. Bozzini, A deep-learning approach
to parameter fitting for a lithium metal battery cycling model, submitted.

59
 On the complexity of infinite argumentation
Luca San Mauro
University of Bari [email protected]

Uri Andrews
University of Wisconsin [email protected]

The theory of abstract argumentation frameworks (AFs), introduced in Dung’s seminal work
[3], has become a foundational topic in knowledge representation. AFs provide a versatile and
powerful tool for modeling diverse reasoning problems, especially in scenarios requiring the reso-
lution of conflicting arguments. To accommodate varying argumentative contexts, a wide range
of semantics has been developed to determine which arguments or extensions (i.e., sets of argu-
ments) are considered acceptable (for an in-depth overview, see the handbook [2]).
While research has extensively explored finite AFs, the study of infinite AFs remains under-
developed, creating theoretical, conceptual, and practical gaps. Our work [1] addresses these
gaps by systematically analyzing the algorithmic complexity of problems associated with infinite
AFs. Leveraging concepts from computability theory, we define computable AFs as those where
a Turing machine can determine, for any pair of arguments, whether one attacks the other. Our
results reveal that, for several established semantics, determining whether an argument is (cred-
ulously or skeptically) accepted reaches maximal complexity, properly belonging to the so-called
Σ11 and Π11 classes.
Moreover, we demonstrate that a single, carefully constructed infinite AF suffices to witness
our hardness results, highlighting that argument acceptability remains highly undecidable for an
individual, specific framework. Finally, we propose a way of using Turing degrees to calibrate,
for a given infinite AF, the exact difficulty of computing an extension in a given semantics. This
approach uncovers a rich and intricate landscape of complexities, significantly advancing our
understanding of infinite AFs and their computational properties.

References
[1] U. Andrews and L. San Mauro, On computational problems for infinite argumentation frameworks:
The complexity of finding acceptable extensions, in Proceedings of the 22nd International Workshop
on Nonmonotonic Reasoning (NMR 2024), CEUR Workshop Proceedings, 3835: 3–13, 2024
[2] Pietro Baroni, Dov Gabbay, Massimilino Giacomin, and Leendert van der Torre (eds), Handbook of
Formal Argumentation, College Publications, London, 2018
[3] P. M. Dung, On the acceptability of arguments and its fundamental role in nonmonotonic reasoning,
logic programming and n-person games, Artificial intelligence, 77: 321–357, 1995

60
 Trade-off Invariance Principle for regularized functionals
Alessandro Scagliotti
Technical University of Munich & Munich Center for Machine Learning (MCML) [email protected]

Massimo Fornasier, Jona Klemenc

Technical University of Munich & Munich Center for Machine Learning (MCML)

[email protected], [email protected]

In this talk, we consider functionals Hα : U → R ∪ {+∞} of the form Hα (u) = F (u) + αG(u)
with α ∈ [0, +∞), and where U ̸= ∅ is a set without further structure. Assuming that

Hα⋆ := arg min Hα

U

is non-empty for every α ∈ [a, b] ⊂ [0, +∞) (with 0 ≤ a < b), we first show that —excluding at
most countably many exceptional values of α ∈ [a, b]— we have the following:

inf⋆ F = sup F, inf⋆ G = sup G,

Hα ⋆
Hα Hα ⋆
Hα

i.e., for every u⋆1 , u⋆2 ∈ Hα⋆ the identities F (u⋆1 ) = F (u⋆2 ) and G(u⋆1 ) = G(u⋆2 ) hold true.
We further prove a stronger result, which asserts that for all but countable many α ∈ [0, +∞), if
inf u∈U Hα (u) > −∞, then there exists a value Gα ∈ [−∞, +∞] such that G(ui ) → Gα for every
sequence (ui )i∈N such that Hα (ui ) → inf u∈U Hα (u) as i → ∞.
This fact in turn implies an unexpected consequence for functionals regularized with uniformly
convex norms: excluding again at most countably many values of α, it turns out that for a
minimizing sequence, convergence to a minimizer in the weak or strong sense is equivalent.

References
[1] M. Fornasier, J. Klemenc, A. Scagliotti Trade-off Invariance Principle for minimizers of regularized
functional, arXiv:2411.11639 (preprint).

61
 Quantum Optimization in Environmental Resource
Management: A Focus on Irrigation Scheduling
Vincenzo Schiano Di Cola
Istituto di Ricerca sulle Acque, Consiglio Nazionale delle Ricerche; Quantum2Pi s.r.l.

[email protected]

Dimitri Jordan Kenne, Gabriele Intoccia

Dipartimento di Matematica e Fisica, Università degli Studi della Campania “Luigi Vanvitelli”

Salvatore Cuomo
Dipartimento di Matematica e Applicazioni, Universit‘a degli Studi di Napoli Federico II

[email protected]

Effective resource management in agriculture is essential for sustainability, given the in-
creasing demand on water resources. Traditional optimization methods for irrigation scheduling
frequently encounter difficulties in reconciling complex constraints, such as temporal dependen-
cies, resource availability, and environmental considerations. Date et al. [1] recently proposed
the use of quantum computers to accelerate machine learning model training. They formu-
lated three machine learning problems (linear regression, support vector machine, and balanced
k-means clustering) as Quadratic Unconstrained Binary Optimization (QUBO) problems and
proposed solving them using adiabatic quantum computing. In this context, Quantum Approxi-
mate Optimization Algorithm (QAOA) serves as an alternative approach for obtaining effective
approximate solutions to these problems. This could lead to the use of QAOA in deep learning
for neural network training and to boost novel research opportunities including non-Gaussian
gates, exploring quantum advantages with decoherence, developing specialized Quantum Neu-
ral Networks (QNNs), and a more profound examination of fundamental concepts in quantum
physics as they relate to QNNs [2].
This research examines the application of quantum algorithms, specifically the QAOA, to
enhance resource management in agriculture. This research presents irrigation scheduling as a
QUBO problem and explores various ansatz in the setting of a Variational Quantum Eigensolver
(VQE) [3, 4]. This study emphasizes the potential of quantum optimization in addressing critical
challenges in agricultural water management, offering a method for improved sustainability via
enhanced resource allocation. The suggested approach illustrates the broader applicability of
quantum approximation in solving complex optimization problems across diverse environmental
and industrial contexts, extending beyond irrigation.

References
[1] Date, P., Arthur, D., & Pusey-Nazzaro, L. (2021). QUBO formulations for training machine learning
models. , Scientific reports, 11(1), 10029.
[2] Blekos, K., Brand, D., Ceschini, A., Chou, C. H., Li, R. H., Pandya, K., & Summer, A. (2024). A
review on quantum approximate optimization algorithm and its variants. , Physics Reports, 1068,
1-66.
[3] Muhamediyeva, Dilnoz & Niyozmatova, Nilufar & Yusupova, Dilfuza & Samijonov, Boymirzo.
(2024). Quantum optimization methods in water flow control., E3S Web of Conferences. 590.
02003. 10.1051/e3sconf/202459002003.
[4] Scherer, Wolfgang. (2019). Mathematics of Quantum Computing: An Introduction., J10.1007/978-
3-030-12358-1

62
A Framework Combining Machine Learning and Statistical
Modeling for Detecting Extreme Events in
High-Dimensional Data
Dhruv Singhvi
Bond Street, Norwich, UK [email protected]

The detection and identification of extreme events in complex, high-dimensional datasets

present significant challenges across various fields, including finance, environmental science, and
engineering. We introduce a framework that combines machine learning-driven dimensionality
reduction with advanced statistical transformations to address this challenge. This approach sim-
plifies high-dimensional data while preserving its inherent probabilistic and relational properties,
enabling effective and interpretable linear analysis.
The framework begins with dimensionality reduction using algorithms such as Uniform
Manifold Approximation and Projection (UMAP) [McInnes et al., 2018], which project
high-dimensional data into a two-dimensional space. UMAP facilitates the creation of a vi-
sually interpretable representation, ensuring the retention of essential topological features for
subsequent analysis.
Following dimensionality reduction, the data is transformed into a linear progression through a
combination of cumulative distribution alignment and monotonic mapping [Wasserman,
2006]. This process involves calculating the empirical cumulative distribution functions (ECDFs)
of the reduced dimensions to map the data to a uniform distribution. The uniform coordinates
are then aligned to a linear configuration, such as y = mx+c, using transformations that preserve
the original dataset’s probabilistic structure and relative relationships.
Once aligned, linear regression models are employed to identify deviations from expected
patterns, with residual analysis pinpointing outliers and anomalies. These outliers are interpreted
as potential extreme events, characterized by their deviation from the linear progression. This
linear framework simplifies the detection of rare or unexpected phenomena, allowing traditional
statistical techniques such as hypothesis testing and confidence interval estimation to be applied
effectively [Wasserman, 2006].
This methodology bridges the gap between machine learning and statistical modeling, pro-
viding a scalable, interpretable, and versatile solution for extreme event detection. In the work-
ing paper, we demonstrate the effectiveness of the framework through applications to finan-
cial datasets, highlighting its capability to identify anomalies. Future research will explore its
adaptability to domains such as environmental systems, biological data, and urban modeling,
expanding its applicability to a wide range of critical challenges.

References
[1] McInnes, L., Healy, J., & Melville, J. (2018) UMAP: Uniform Manifold Approximation and Pro-
jection for Dimension Reduction, Journal of Open Source Software
[2] Wasserman, L. (2006) All of Statistics: A Concise Course in Statistical Inference

63
 A Deep-QLP Decomposition Algorithm and Applications
Cristiano Tamborrino
[email protected]
In collaboration with: Antonella Falini, Francesca Mazzia
Dipartimento di Informatica, Università degli Studi di Bari Aldo Moro, Italy

Abstract
Singular value decomposition (SVD) is a fundamental tool in data analysis and
machine learning. Starting from the Stewart’s QLP decomposition [1], we propose
an innovative Deep-QLP decomposition algorithm for efficiently computing an
approximate Singular Value Decomposition (SVD) based on the preliminary work
in [2]. Given a specified tolerance τ , the algorithm automatically computes a
positive integer f and a factorization Uf LD T D
f Vf , with Lf diagonal matrix, Uf , Vf
matrices of rank f with orthonormal columns such that
∥A − Uf LD T
f Vf ∥2 ≤ 3τ ∥A∥2 .

The Deep-QLP algorithm stands out for its ability to return an approximation
of the largest singular values, based on a fixed tolerance, to achieve significant
dimensionality reduction while simultaneously preserving essential information in
the data. In addition, it can also be used to return an approximation of the small-
est singular values that can be used in some applications.
The algorithm has been successfully integrated with the randomized SVD [3], mak-
ing the Deep-QLP algorithm particularly effective for sparse matrices, which are
prevalent in numerous applications such as text mining.
Several numerical experiments have been conducted, demonstrating the effective-
ness of the proposed method.

Acknowledgements. Cristiano Tamborrino, Francesca Mazzia and Antonella Falini acknowledge

the support of the PNRR project FAIR - Future AI Research (PE00000013), Spoke 6 - Symbiotic AI (CUP
H97G22000210007) under the NRRP MUR program funded by the NextGenerationEU. The authors
thank the GNCS for its valuable support under the INdAM-GNCS project CUP E55F22000270001.

References
[1] Gilbert W. Stewart. The QLP Approximation to the Singular Value Decomposition. SIAM J. Sci.
Comput., 20:1336–1348, 1999. https://api.semanticscholar.org/CorpusID:15701097.
[2] Antonella Falini and Francesca Mazzia. Approximated Iterative QLP for Change Detection in Hy-
perspectral Images. AIP Conference Proceedings, 3094(1):370003, 2024. https://doi.org/10.1063/
5.0210496.
[3] Nathan Halko, Per-Gunnar Martinsson, and Joel A. Tropp. Finding Structure with Random-
ness: Probabilistic Algorithms for Constructing Approximate Matrix Decompositions. SIAM Review,
53(2):217–288, 2011. SIAM.

64
 Variable metric proximal stochastic gradient methods with

additional sampling
Ilaria Trombini, Valeria Ruggiero
Dept. of Mathematics and Computer Science, University of Ferrara, Ferrara, 44121, Italy

[email protected], [email protected]

Natavsa Krklec Jerinkić

Faculty of Sciences, Department of Mathematics and Informatics, University of Novi Sad, Trg Dositeja
Obradovića 4, Novi Sad, 21000, Serbia [email protected]

Federica Porta
Dept. of Physics, Informatics and Mathematics, University of Modena and Reggio Emilia, Modena,

41125, Italy [email protected]

Regularized empirical risk minimization problems are prevalent across various

domains, such as machine learning, signal processing, and image processing. Many
optimization challenges within machine learning can be formulated as the mini-
mization of a composite function: the first component typically represents the
expected risk, which is often substituted by empirical risk in practice, while the
second imposes prior information on the solution. Generally, the first term is
differentiable and the second term is convex, making proximal gradient methods
particularly well-suited for tackling these optimization problems.
However, in the context of large-scale machine learning, calculating the full gra-
dient of the differentiable term can be computationally prohibitive, rendering stan-
dard proximal gradient algorithms impractical. Consequently, proximal stochastic
gradient methods have been extensively explored in optimization research over
recent decades [1, 2, 3].
This talk introduces a class of proximal stochastic gradient methods built on
three foundational elements: a variable metric underlying the iterative process, a
stochastic line search mechanism to control the descent properties, and an incre-
mental mini-batch size technique based on additional sampling. Convergence re-
sults for the proposed algorithms are established under varying assumptions on the
objective function. Notably, no assumption regarding the Lipschitz continuity of
the gradient of the differentiable part of the objective function is required. Possible
strategies for the automatic selection of parameters in the proposed framework are
also discussed. Numerical experiments on binary classification problems demon-
strate the effectiveness of this approach in comparison to other leading proximal
stochastic gradient methods.

References
[1] Xiao, L and Zhang, T A proximal Stochastic Gradient Method with Progressive Variance Reduction,
SIAM J. Optim. 24, 4, (2014), 2057–2075.
[2] Phamy. N H, and Nguyen, L M and Phan, D T and Tran–Dinh, Q ProxSARAH: an efficient
algorithmic framework for stochastic composite nonconvex optimization, J. Mach. Learn. Res. 21,
1, Article 110 (2020), 1–48.

65
[3] Wang, Zhe and Ji, Kaiyi and Zhou, Yi and Liang, Yingbin and Tarokh, Vahid SpiderBoost and
Momentum: Faster Stochastic Variance Reduction Algorithms, Proceedings of the 33rd International
Conference on Neural Information Processing Systems, Curran Associates Inc. 216, (2019), 2406–
2416.

66
Industry Talks
• Pirelli & C. S.p.A. (Mattia Beretta - Generative AI tech lead) (page 68)
Pirelli pratical development of LLM application for risk prevention on the workplace
• Planetek Italia S.r.l (Nicolò Taggio - GeoAI team coordinator) (page 69)
Data, Math, and Machine Learning: Revolutionizing Earth Observation Technologies

67
 Pirelli pratical development of LLM application for risk
prevention on the workplace
Mattia Beretta
Generative AI tech lead, Pirelli & C. S.p.A.

[email protected]

Pirelli, a leader in tire manufacturing, strengthens its commitment to ensuring workplace safety. The
"Health, Safety and Environment" department, with the support of GenAI, can now not only analyze
thousands of textual reports from global facilities more efficiently each year but also implement preventive
actions to mitigate risk situations. By leveraging the capabilities of natural language LLM, Pirelli is able
to automate and optimize the risk assessment process by summarizing reports and highlighting critical
points.

68
Data, Math, and Machine Learning: Revolutionizing Earth
Observation Technologies
Nicolò Taggio
GeoAI team coordinator, Planetek Italia S.r.l, [email protected]

The advent of advanced machine learning (ML) algorithms and imaging technologies (such as hy-
perspectral and multispectral sensors), has significantly transformed Earth Observation (EO). The con-
nection between data, mathematics, and ML will be explored to understand how they are driving this
transformation, revolutionizing interpretation and use of EO data for various applications.
A foundational overview of hyperspectral and multispectral imaging, highlighting their key differences
and advantages, will be presented. By diving into the feature space, mathematical operations and
machine learning techniques can be applied to combine spectral bands, creating indexes that enhance
the detection and classification of surface features. Furthermore, to illustrate the practical application
of these concepts, will be highlighted a Burned Area Detection Using Non-Negative Matrix
Factorization (NMF). This unsupervised approach leverages spectral signatures to identify and map
burned areas accurately, showcasing the power of data-driven feature extraction.
Finally, a service called Rheticus Network alert will be presented, which integrates ML algorithms
with data and mathematical models to provide actionable insights for pipeline monitoring. This service
emphasizes user interaction, showcasing the importance of tailoring EO solutions to meet end-user needs.
Looking ahead, the focus will be on the potential of cognitive cloud computing to optimize complex
satellite networks through cooperative swarming. This approach leverages multi-objective functions
inspired by game theory, enabling autonomous self-organization of satellite assets to achieve tasks even
in scenarios with incomplete information. This future-oriented perspective highlights how advances in
distributed intelligence and autonomous decision making are reshaping the next generation of space-
based technologies.

69
Posters
• Carlo Abate (page 71)
MaxCutPool: Differentiable Feature-Aware MAXCUT for Pooling in Graph Neural Networks
• Sara Cambiaghi (page 72)
Distributional forecast approaches to stochastic optimization in healthcare appointment scheduling
• Anna Livia Croella (page 73)
Anticlustering for Large Scale Clustering
• Serena Grazia De Benedictis (page 74)
ROI Image Identification via Topological Data Analysis: A Case Study of Brain Tumor MRI
• Roberta De Fazio (page 75)
Inferring Failure Processes via Causality Analysis: from Event Logs to Predictive Fault Trees
• Anna De Magistris (page 76)
A line-search based SGD algorithm with Adaptive Importance Sampling
• Bernardo Forni (page 78)
Adapting SAM2 for Few-Shot Multi-Class Semantic Segmentation
• Caterina Gallegati (page 79)
GANs through the Lens of Topological Data Analysis
• Daniela Gallo (page 80)
CAP: Copyright Audit via Prompt generation
• Grazia Gargano (page 81)
A Low-Rank Multi-Factor Approach to Identify Differentially Expressed Genes in Transcriptome
Data
• Letizia Lorusso (page 82)
Analysis of Decision-Making Styles and Personality Traits in Women Undergoing Voluntary Ter-
mination of Pregnancy: A Bayesian Network Approach Using bnstruct
• Maura Mecchi (page 83)
COSMONET 2.0: An R Package for Survival Analysis Using Screening-Network Methods
• Giuseppina Monteverde (page 84)
Efficiency-driven 3D CNN architectures for hyperspectral classification
• Laura Selicato (page 85)
Bi-level algorithm for optimizing hyperparameters in penalized NMF
• Alessandra Serianni (page 86)
Hybrid knowledge and data-driven approaches for Diffuse Optical Tomography reconstruction
• Gaetano Settembre (page 87)
Spatial Informed Hierarchical Clustering for Hyperspectral Imagery via Total Variation
• Paolo Sorino (page 88)
Empowering Clinicians with Explainable AI: Predicting Mortality Risk in MAFLD with Counter-
factual Analysis

70
 MaxCutPool: Differentiable Feature-Aware MAXCUT for
Pooling in Graph Neural Networks
Carlo Abate
[email protected]

Filippo Maria Bianchi

[email protected]

We propose a novel approach to compute the MAXCUT in attributed graphs, i.e. graphs with features
associated with nodes and edges, by exploiting heterophilic message passing to assign connected nodes to
different partitions. The approach is fully differentiable, making it possible to find solutions that jointly
optimize the MAXCUT along with other objectives. Based on the obtained MAXCUT partition, we implement
MaxCutPool, a hierarchical graph pooling layer for graph neural networks. The layer is sparse, differ-
entiable, and particularly suitable for downstream tasks on heterophilic graphs. Our key contributions
include: (1) a novel MAXCUT computation method for attributed graphs, (2) a new hierarchical pooling
layer especially effective for heterophilic graphs, (3) a general scheme for node-to-supernode assignment,
and (4) the introduction of the first heterophilic dataset for graph classification. Experimental results
demonstrate that MaxCutPool achieves state-of-the-art performance across various graph classification
and node classification tasks, highlighted by perfect accuracy on expressiveness tests and significant
improvements on heterophilic graph classification.

References
[1] C. Abate and F. M. Bianchi. Maxcutpool: differentiable feature-aware maxcut for pooling in graph
neural networks, 2024, https://arxiv.org/abs/2409.05100

71
 Distributional forecast approaches to stochastic
optimization in healthcare appointment scheduling
Sara Cambiaghi
Department of Mathematics “F. Casorati”, University of Pavia

[email protected]

Davide Duma
Department of Mathematics “F. Casorati”, University of Pavia

[email protected]

Appointment scheduling in healthcare involves optimizing the allocation of appointment slots to

balance patient access and resource utilization, with one of the primary sources of uncertainty being the
service time. In this talk we propose a stochastic optimization approach that leverages distributional
forecasts to identify patient groups with similar service times.
We explore two alternative strategies to achieve the desired distributional forecast. The first method
utilizes random forest leaf embeddings, where the encoded data are clustered using the K-means algo-
rithm. The second method employs a Decision Tree to generate an initial partition of patients and their
corresponding distributions, which are subsequently clustered using the Wasserstein K-Means algorithm.
Both strategies enable the derivation of probability distributions for each resulting cluster.
The appointment scheduling problem is formulated through a stochastic programming model that
minimizes waiting times for outpatients, completion times for emergency patients and inpatients, and
overtime. A genetic algorithm is used to solve the optimization problem by estimating the Pareto front
within a reasonable timeframe.
As a case study, we examine the CT scan service at Policlinico San Matteo in Pavia, Italy, where
three categories of patients–outpatients, inpatients, and emergency cases–use the same resources but have
different priorities and needs. The considered multi-objective decision problem concerns the outpatient
CT scan scheduling over a weekly planning horizon, avoiding conflicts with the real-time scheduling of
inpatients and emergencies and ensuring doctors have enough time to complete both the scans and their
reports.
A computational analysis is performed to evaluate the effectiveness of the proposed optimization and
machine learning approach for improving the efficiency of the CT scan service. We also compare different
point predictions with distributional forecasts to represent the uncertainty of service time durations and
understand which method provides better results.

References
[1] J. Marcak, Y.L. Huang, Radiology procedure time slot redesign to improve scheduling efficiency,
Proceedings of the 62nd IIE Annual Conference and Expo.
[2] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective genetic algorithm:
NSGA-II, IEEE Transactions on Evolutionary Computation.
[3] P. Bhattacharjee, P.K. Ray, Scheduling appointments for multiple classes of patients in presence
of unscheduled arrivals: Case study of a CT department, Proceedings of the IISE Transactions on
Healthcare Systems Engineering.
[4] T. Gneiting, A.E. Raftery, Strictly Prover Scoring Rules, Prediction, and Estimation, Journal of
the American Statistical Association.
[5] M. Bicego, K–Random Forest: a K–Means style algorithm for Random Forest clustering, Interna-
tional Joint Conference on Neural Networks (IJCNN).
[6] Y. Zhuang, X. Chen, Y. Yang, Wasserstein K–means for clustering probability distributions, Ad-
vances in Neural Information Processing Systems.

72
 Anticlustering for Large Scale Clustering
Anna Livia Croella
Sapienza University of Rome, Rome, Italy [email protected]

Veronica Piccialli, Antonio Maria Sudoso

Sapienza University of Rome, Rome, Italy [email protected],

[email protected]

This research develops innovative methodologies for integrating clustering and anticlustering tech-
niques into large-scale data analysis within AI frameworks. The objective is to establish a mechanism
that generates tighter lower bounds for the clustering problem, starting from a heuristic solution that
minimizes the Within-group Sum of Squares (WSS). A key insight is that the minimum WSS of the union
of disjoint subsets is always greater than or equal to the sum of the minimum WSS of the individual
subsets [1]. This indicates that summing the minimum WSS values of disjoint subsets provides a valid
lower bound for the optimal WSS of the entire dataset. To enhance this lower bound, we maximize
the minimum WSS of each subset by creating groups of points with high dissimilarity, a process known
as anticlustering [2]. Through anticlustering, we developed a certification process to validate clustering
solutions obtained using the k-means algorithm. We tested this mechanism on large-scale datasets con-
taining 2,000 to 10,000 data points and between 2 and 500 features. Our procedure consistently achieved
gaps between the clustering solution and the lower bound ranging from 0.1% to 5%. Future work will
focus on iterative improvements to the clustering solutions through feedback loops, as well as integrating
the generation of lower bounds into a Branch & Bound algorithm [3].

References
[1] Diehr, G. Evaluation of a Branch and Bound Algorithm for Clustering, SIAM J. Sci. Stat. Comput.
6, 2, 268–284 (1985)
[2] Papenberg, M. K-Plus anticlustering: An improved k-means criterion for maximizing between-group
similarity, British Journal of Mathematical and Statistical Psychology, 77(1), 80–102 (2023)
[3] Piccialli, V., Russo, A.R., and Sudoso, A.M. An Exact Algorithm for Semi-supervised Minimum
Sum-of-Squares Clustering, Computers & Operations Res., 147, 105958 (2021)

73
 ROI Image Identification via Topological Data Analysis: A
Case Study of Brain Tumor MRI
Serena Grazia De Benedictis
University of Bari Aldo Moro, [email protected]

Grazia Gargano, Gaetano Settembre

University of Bari Aldo Moro, [email protected], [email protected]

In the medical context, modern imaging methods such as magnetic resonance imaging (MRI) have
completely changed how diseases are diagnosed and tracked. Advanced image processing algorithms are
increasingly employed to automate the interpretation of medical images, facilitating faster and more
accurate diagnosis. This work presents a novel ensemble of methods using MRI data for the detection
and classification of common brain cancers. The proposed approach combines dimensionality reduction
technique with machine learning (ML) algorithms, and then integrates ML prediction with topological
data analysis (TDA)-based results [2]. A low-rank Tucker decomposition [3] is used to reduce data
dimensionality while maintaining the key structures and properties of preprocessed MRI scans. Robust
tumor classification models can be developed with supervised machine learning classifiers that are trained
on the low-dimensional representations of the data. The MRI scans are also parallelly processed using
persistent homology (PH) [4], an algebraic method for measuring topological features of data to explore
the spatial relationships and patterns present in the pixel distribution and the geometry of the images.
Indeed, by extracting the most persistent connected component of the MRI scan, we can precisely
identify region of interest (ROI) that can suggest the existence or features of a possible tumor and
require further investigation. The promising results obtained by applying the proposed framework to a
brain tumor image dataset demonstrate the effectiveness of integrating low-rank approximation, ML and
TDA techniques for tumor detection and classification. This comprehensive approach provides a robust
strategy for future research and clinical application, potentially extendable to other solid tumors.

References
[1] S.G. De Benedicitis, G. Gargano, and G. Settembre Enhanced MRI brain tumor detection and clas-
sification via topological data analysis and low-rank tensor decomposition, Journal of Computational
Mathematics and Data Science (2024), 13, 100103. doi:10.1016/j.jcmds.2024.100103
[2] Dey, Tamal Krishna and Wang, Yusu Computational Topology for Data Analysis, Cambridge
University Press, 2022. doi:10.1017/9781009099950.
[3] Kolda, Tamara G. and Bader Tensor Decompositions and Applications, SIAM Review 51 (3) (2009)
455–500. doi:10.1137/07070111x.
[4] Schenck, Hal Algebraic Foundations for Applied Topology and Data Analysis , Springer International
Publishing, 2022. doi:10.1007/978-3-031-06664-1.

74
 Inferring Failure Processes via Causality Analysis:
from Event Logs to Predictive Fault Trees
Roberta De Fazio
Dipartimento di Matematica e Fisica, Università degli Studi della Campania Luigi Vanvitelli, Italy

[email protected]

Benoît Depaire
Faculty of Business Informatics, Hasselt University, Belgium

[email protected]

Stefano Marrone, Laura Verde

Dipartimento di Matematica e Fisica, Università degli Studi della Campania Luigi Vanvitelli, Italy
stefano.marrone,[email protected]

In the current Artificial Intelligence era, the integration of the Industry 4.0 paradigm in real-world
settings requires robust and scientific methods and tools. Two concrete aims are the exploitation of
large datasets [1] and the guarantee of a proper level of explainability, demanded by critical systems
and applications [2]. Focusing on the predictive maintenance problem, this work leverages causality
analysis to elicit knowledge about system failure processes. The result is a model expressed according
to a newly introduced formalism: the Predictive Fault Trees [3]. This model is enriched by causal
relationships inferred from dependability-related event logs. The proposed approach considers both
fault-error-failure chains between system components and the impact of environmental variables (e.g.,
temperature, pressure) on the health status of the components. A proof of concept shows the effectiveness
of the methodology, leveraging an event-based simulator [4].

References
[1] R. De Fazio, A. Balzanella, S. Marrone, F. Marulli, L. Verde, V. Reccia, P. Valletta CaseID Detection
for Process Mining: A Heuristic-Based Methodology, Process Mining Workshops, Springer Nature
Switzerland
[2] S. Ramezani, L. Cummins, B. Killen, R. Carley, A. Amirlatifi, S. Rahimi, M. Seale, L. Bian Scalabil-
ity, Explainability and Performance of Data-Driven Algorithms in Predicting the Remaining Useful
Life: A Comprehensive Review, IEEE Access,Institute of Electrical and Electronics Engineers
(IEEE)
[3] R. De Fazio, S. Marrone, L. Verde, V. Reccia, P. Valletta Towards an extension of Fault Trees
in the Predictive Maintenance Scenario, 19th European Dependable Computing Conference, arXiv
pre-print
[4] C. Abate, L. Campanile, S. Marrone A flexible simulation-based framework for model-based/data-
driven dependability evaluation, Proceedings - 2020 IEEE 31st International Symposium on Software
Reliability Engineering Workshops, ISSREW 2020

75
 A line-search based SGD algorithm with Adaptive
Importance Sampling
Anna De Magistris
Dipartimento di Matematica e Fisica Luigi Vanvitelli, [email protected]

Filippo Camellini, Serena Crisci, Giorgia Franchini

Università degli Studi di Modena e Reggio Emilia, [email protected]

Dipartimento di Matematica e Fisica Luigi Vanvitelli, [email protected]

Università degli Studi di Modena e Reggio Emilia, [email protected]

Stochastic Gradient Methods are essential for solving large-scale optimization problems, particularly
when the objective function F is expressed as the sum of n functions fi , each with an Li -Lipschitz
continuous gradient [1]. Stochastic Gradient Descent (SGD), which computes an approximate gradient
by sampling a function fik from a probability distribution pk , is highly efficient and scalable. However, its
asymptotic performance is limited; with a constant step size, it converges only to a neighborhood of the
optimum even under strong convexity assumptions [2]. To address this, variance-reduction techniques
like SVRG [3] and SAGA [4] combine stochastic gradients with partial updates of the full gradient.
Another approach involves dynamic sampling to increase the batch size progressively, as in algorithms
like LISA [5, 6]. Importance sampling is also explored, optimizing the sampling distribution pk to reduce
variance based on Lipschitz constants L [7]. Yet, estimating L remains challenging, especially in deep
learning contexts. A notable advancement is the SGD-AIS algorithm, which approximates an optimal
sampling distribution without relying on L and demonstrates superior performance compared to SGD
with uniform sampling [8]. However, the decreasing step size employed in SGD-AIS can slow convergence
and demands careful parameter tuning. To overcome these limitations, we propose an automatic step
size selection method using a stochastic Armijo-type line-search procedure. This approach simplifies
parameter tuning, accelerates convergence, and leverages the importance sampling distribution of SGD-
AIS. Our contributions include extending SGD-AIS with a stochastic line-search strategy and introducing
a variant for mini-batch stochastic gradients. Theoretical convergence results and experiments on ℓ2 -
regularized logistic regression and smooth hinge loss confirm the effectiveness of the proposed methods.

References
[1] F. E. Curtis and K. Scheinberg, Optimization Methods for Supervised Machine Learning: From
Linear Models to Deep Learning, arXiv preprint arXiv:1706.10207, 2017.
[2] L. Bottou, F. E. Curtis, and J. Nocedal, Optimization Methods for Large-Scale Machine Learning,
SIAM Review, vol. 60, no. 2, pp. 223–311, 2018.
[3] S. J. Reddi, A. Hefny, S. Sra, B. Póczós, and A. Smola, Stochastic Variance Reduction for Nonconvex
Optimization, Proceedings of ICML 2016, pp. 314–323.
[4] A. Defazio, F. Bach, and S. Lacoste-Julien, SAGA: A Fast Incremental Gradient Method with
Support for Non-Strongly Convex Composite Objectives, Proceedings of NIPS 2014, pp. 1646–1654.
[5] G. Franchini, F. Porta, V. Ruggiero, and I. Trombini, A Line Search Based Proximal Stochastic
Gradient Algorithm with Dynamical Variance Reduction, Journal of Scientific Computing, 2022.
[6] G. Franchini, F. Porta, V. Ruggiero, I. Trombini, and L. Zanni, A Stochastic Gradient Method with
Variance Control and Variable Learning Rate for Deep Learning, Journal of Computational and
Applied Mathematics, vol. 451, p. 116083, 2024.
[7] L. Xiao and T. Zhang, A Proximal Stochastic Gradient Method with Progressive Variance Reduction,
SIAM Journal on Optimization, vol. 24, no. 4, pp. 2057–2075, 2014.
[8] H. Liu, X. Wang, J. Li, and A. M.-C. So, Low-Cost Lipschitz-Independent Adaptive Importance
Sampling of Stochastic Gradients, Proceedings of ICPR 2020, pp. 2150–2157.
[9] S. Vaswani, A. Mishkin, I. Laradji, M. Schmidt, G. Gidel, and S. Lacoste-Julien, Painless Stochastic
Gradient: Interpolation, Line-Search, and Convergence Rates, Proceedings of NIPS 2019.

76
[10] P. Zhao and T. Zhang, Stochastic Optimization with Importance Sampling for Regularized Loss
Minimization, Proceedings of PMLR 2015, pp. 1–9.
[11] D. Bertsekas, Convex Optimization Theory, Athena Scientific, Belmont, Massachusetts, 2009.
[12] C. Tan, S. Ma, Y.-H. Dai, and Y. Qian, Barzilai-Borwein Step Size for Stochastic Gradient Descent,
Advances in Neural Information Processing Systems, vol. 29, 2016.
[13] A. Johnson, Example Book, Example Publisher, Example City, 2023.

77
 Adapting SAM2 for Few-Shot Multi-Class Semantic
Segmentation
Bernardo Forni
University of Pavia [email protected]

Segment Anything Model 2 (SAM2) has shown outstanding performance in zero-shot image and
video segmentation. We introduce a novel module to adapt SAM2 for the challenging and underexplored
task of few-shot multi-class semantic segmentation. This task involves labeling each pixel within an
image using a limited set of mask-annotated images from multiple classes. Our approach leverages a
transformer architecture that aggregates the SAM2 features of different classes, accommodating any
N-way K-shot configurations.
Furthermore, we employ a meta-learning strategy to efficiently fine-tune the entire model, thereby
improving its generalization capabilities. Our work is motivated by the demands of industrial image
segmentation, where precise segmentation is crucial for detecting semantic anomalies. We achieved
remarkable results on internal datasets.
Based on joint work with: Gabriele Lombardi, Mirco Planamente and Federico Pozzi.

References
[1] Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C.,
Gustafson, L. and Mintun, E. Sam 2: Segment anything in images and videos, arXiv preprint
arXiv:2408.00714
[2] De Marinis, P., Fanelli, N., Scaringi, R., Colonna, E., Fiameni, G., Vessio, G. and Castellano, G.
Label Anything: Multi-Class Few-Shot Semantic Segmentation with Visual Prompts, arXiv preprint
arXiv:2407.02075

78
 GANs through the Lens of Topological Data Analysis
Caterina Gallegati
University of Siena, [email protected]

B. T. Corradinia , B. Cullenb , S. Marzialic , G. A. D’Invernod , M.

Bianchinic , F. Scarsellic
a
University of Florence, b University of Pisa, c University of Siena, d SISSA, Trieste

Generative Adversarial Networks (GANs) [1] aim to produce realistic samples by mapping a low-
dimensional latent space to a high-dimensional data space by exploiting an adversarial training mecha-
nism. Despite achieving state-of-the-art results, GAN training faces significant challenges such as mode
collapse, vanishing gradients, and inefficiencies in hyperparameter tuning, relying on computationally
expensive trial-and-error methods. In addition, GANs lack a clear early stopping criterion, often leading
to resource-intensive training processes.
This work investigates GANs using Topological Data Analysis (TDA) tools [3] to gain deeper insights
into their training dynamics and generative capabilities. By employing persistent homology, we examine
the evolution of topological features during training, focusing on the convergence of the generated mani-
fold to that of real data. Through various experiments on MNIST and CIFAR-10 datasets with different
GAN models, we analyze the interplay between model architecture, training stability, and performance,
as well as characterise common issues in GANs. In particular, we show that the Wasserstein distance
between persistence diagrams, which summarise the topological features of manifolds, is a robust tool
for quantifying similarities between generated and real data, offering a novel perspective on evaluating
samples beyond conventional metrics like the Frechet Inception Distance (FID) [2]. Indeed, the FID score
is shown to be insufficient in assessing the quality of generated images, neither alone nor in combination
with the Intrinsic Dimension estimation [4]. Our results suggest that homological features provide a
suitable characterisation of the generative process that can be valuable for uncovering insights about
the structural transformations occurring during the training of a GAN. This study lays the foundation
for integrating topology-based approaches into the optimization and assessment of generative models,
potentially enabling the formulation of an early stopping criterion.

References
[1] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A. &
Bengio, Y. Generative Adversarial Networks. (arXiv,2014), https://arxiv.org/abs/1406.2661
[2] Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B. & Hochreiter, S. GANs Trained
by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium. (arXiv,2017),
https://arxiv.org/abs/1706.08500
[3] Chazal, F. & Michel, B. An introduction to Topological Data Analysis: fundamental and practical
aspects for data scientists. (arXiv,2017), https://arxiv.org/abs/1710.04019
[4] Pope, P., Zhu, C., Abdelkader, A., Goldblum, M. & Goldstein, T. The Intrinsic Dimension of Images
and Its Impact on Learning. (arXiv,2021), https://arxiv.org/abs/2104.08894

79
 CAP: Copyright Audit via Prompt generation
Daniela Gallo
ICAR-CNR and University of Salento, Italy, [email protected]
Angelica Liguori
ICAR-CNR, Italy, [email protected]

Ettore Ritacco
University of Udine, Italy, [email protected]
Luca Caviglione
IMATI-CNR, Italy, [email protected]

Fabrizio Durante
University of Salento, Italy, [email protected]
Giuseppe Manco
ICAR-CNR, Italy, [email protected]

To achieve accurate and unbiased predictions, Machine Learning (ML) models rely on large, het-
erogeneous, and high-quality datasets. However, this could raise ethical and legal concerns regarding
copyright and authorization aspects, especially when information is gathered from the Internet. Indeed,
such data may be protected by intellectual property rights, and proper authorizations for its usage should
be granted on a case-by-case basis [1]. With the rise of generative models, being able to track data has
become of particular importance. Indeed, as they require large datasets for being trained, they often rely
on data derived from different sources without being able to discriminate among public or “restricted”
sources. Consequently, they may (un)intentionally replicate copyrighted contents [2]. To this aim, we
propose Copyright Audit via Prompts generation (CAP), a framework for automatically checking if
the training set used by an ML model contains unauthorized data. Testing whether data has been used
to train an ML model is known as membership inference problem. However, different from classical
Membership Inference Attacks [3] that directly check if a given slice of information has been used in the
training phase, we cannot directly inspect the training set used by the model, as only the owner knows it.
To address this issue, CAP generates suitable keys that induce the model to reveal copyrighted content.
Additionally, training prompt generators, which rely on complex architectures like transformers, require
large computational demands. For this reason, we introduce an optimization procedure aiming to speed
up the learning process. By leveraging a generalized Pareto distribution [4], we filter out irrelevant data
based on model error, applying an 80% threshold to exclude extreme outliers. This reduces the dataset
size while preserving the most impactful samples. Extensive evaluations across four realistic IoT scenar-
ios and synthetic datasets demonstrate the effectiveness of our framework in identifying unauthorized
data with high accuracy. This work offers a robust and efficient solution for ensuring responsible and
ethical use of generative artificial intelligence models.

References
[1] Meuris, B., Qadeer, S. & Stinis, P. Machine-learning custom-made basis functions for partial differ-
ential equations. (arXiv,2021), https://arxiv.org/abs/2111.05307
[2] Li, H., Deng, G., Liu, Y., Wang, K., Li, Y., Zhang, T., Liu, Y., Xu, G., Xu, G. & Wang, H.
Digger: Detecting Copyright Content Mis-usage in Large Language Model Training. (arXiv,2024),
https://arxiv.org/abs/2401.00676
[3] Shokri, R., Stronati, M., Song, C. & Shmatikov, V. Membership Inference Attacks against Machine
Learning Models. (arXiv,2016), https://arxiv.org/abs/1610.05820
[4] Vignotto, E. & Engelke, S. Extreme value theory for anomaly detection – the GPD classifier.
Extremes. 23, 501-520 (2020,9), http://dx.doi.org/10.1007/s10687-020-00393-0
 A Low-Rank Multi-Factor Approach to Identify
Differentially Expressed Genes in Transcriptome Data
Grazia Gargano
Department of Mathematics, University of Bari Aldo Moro, Italy, [email protected]

RNA-sequencing (RNA-seq) technology provides a robust platform for transcriptome-wide analysis

of gene expression, enabling the study of transcriptional changes associated with different biological con-
ditions. A primary application of RNA-seq is the identification of differentially expressed genes (DEGs)
between different biological states (e.g., disease vs. normal, treatment vs. control), cell populations, or
time points. DEGs are defined as genes that show statistically significant differences in expression levels
or read counts across experimental conditions, reflecting changes in gene activity associated with bio-
logical processes such as disease progression, therapeutic response or developmental pathways. Despite
their critical role in transcriptomic research, further advances are needed to improve the accuracy and
efficiency of DEGs identification.

We propose a novel mathematical framework based on three-factor nonnegative matrix factorization

(tri-NMF) [1] to identify genes that exhibit differential expression under two or more distinct experi-
mental conditions. We represent gene expression data by a matrix X ∈ R+ n×m , where n is the number
of samples (e.g., patient groups, tissues, experiments, or time points) and m is the number of genes. To
compare experimental conditions, we introduce a tri-NMF-based approach, formulated as a constrained
penalized optimization task:

min Div(X|USV⊤ ) + λU P1 (U) + λS P2 (S) + λV P3 (V)

U≥0,S≥0,V≥0

where Div(·, ·) : Rn×m

+ × Rn×m
+ → R+ denotes some divergence function, which evaluates the good-
ness of fitting; U ∈ R+ n×k
, S ∈ R+ k×r and V ∈ R+ m×r are the nonnegative factors of low-rank data
representation; P1 : Rn×k → R, P2 : Rk×r → R, P3 : Rm×r → R codify regularization constraints to
enforce specific properties on the factor matrices, while λU , λS and λV are some positive regularization
parameters.

For DEGs identification, we consider the generalized Kullback-Leibler divergence as the cost function
and set k = r (k, r < min(n, m)) equal to the number of different conditions we want to compare. The
information about the sample labels is encoded in the structure of the Pfactor U. We impose U to be a
binary matrix representing sample clusters, where Uij ∈ {0, 1} and kj=1 Uij = 1. This ensures that
each sample is assigned to exactly one cluster. Imposing sparsity and orthogonality constraints on the
columns of V ensures that the extracted list of DEGs has minimal or no overlap of genes. The objective
function is minimized by using an alternating scheme with an appropriate choice of the multiplicative
update rules [2]. To compute DEGs, we define a gene score criterion based on the normalized entropy,
which is computed from the coefficients of the matrix V obtained during the factorization. We validate
our approach on synthetic data to assess its performance and robustness under controlled conditions.
Synthetic datasets are generated to simulate realistic biological scenarios, allowing us to test the model’s
ability to accurately identify DEGs.

This is a joint work with Nicoletta Del Buono and Flavia Esposito (Department of Mathematics,
University of Bari Aldo Moro, Bari, Italy).

References
[1] Nicolas Gillis, Nonnegative Matrix Factorization, SIAM, Philadelphia, 2020.
[2] Daniel Lee and Hyunjune Seung, Algorithms for Non-negative Matrix Factorization, Advances in
Neural Information Processing Systems (NeurIPS), Volume 13, 2001.
 Analysis of Decision-Making Styles and Personality Traits
in Women Undergoing Voluntary Termination of
Pregnancy: A Bayesian Network Approach Using bnstruct
Letizia Lorusso
School of Medical Statistics and Biometry, Interdisciplinary Department of Medicine,

University of Bari "Aldo Moro", Bari, Italy. [email protected]

In this study, we explore the application of Bayesian Networks to analyze the relationships between
the General Decision-Making Style (GDMS) test 1, the Big Five Questionnaire (BFQ) 2 and the Per-
sonality Inventory for DSM-5 (PID-5) 3, and socio-demographic characteristics of women who undergo
voluntary termination of pregnancy (VTP). Using the bnstruct package 4 for building Bayesian Net-
works, our goal is to compare the results of different alghortim applied with three scoring fuction, to
define significant patterns that can reveal the underlying dynamics of these choices, considering variables
such as personality type and decision-making aspects related to this experience. The data used comes
from a database containing socio-demographic information of 122 women, as well as their personality
and decision-making test results, with a total of 27 variables.
To this end, we construct a Bayesian network representing the probabilistic dependencies among the
variables and compare the performance of four algorithms for structure learning: Structural Expectation-
Maximization (SEM), Max-Min Parents-and-Children (MMPC), Max-Min Hill-Climbing (MMHC), and
Hill-Climbing (HC). Each algorithm employs a different approach to structure learning, and we assess
their effectiveness in identifying the most accurate causal relationships in our data 5.
Additionally, we compare the performance of the model using three main scoring methods: Bayesian
Dirichlet equivalent score (BDeu), Bayesian Information Criterion (BIC), and Akaike Information Cri-
terion (AIC). These scoring functions are employed to evaluate the quality of the model and determine
which approach provides the best representation of the data.
In general, BDeu is particularly well-suited for data with discrete variables. AIC and BIC penalizes
complexity, choosing the variables in the simpler models with good predictive ability. We evaluate these
scoring functions in the context of the four algorithms. At the first we use the MMPC alghoritm to
explore the conditional dependencies among variables, and to define the skeleton of the network, with-
out directly optimizing a global score. Then, we applied the HC and the MMHC alghoritm: the first
refines the network structure by selecting the best local changes to maximize the scoring function; the
second identifies the parent-child relationships. At the end we use the SEM alghoritm, optimizing both
the model’s structure and its parameters simultaneously, to choose the final model.
The results of this comparison provide valuable insights into which algorithm and scoring function best
capture the relationships among personality, decision-making style, and socio-demographic factors in the
context of VTP decisions.

References
[1] Di Fabio, A. (2007). General Decision Making Style (GDMS): Un primo contributo alla validazione
italiana., GIPO, Giornale Italiano di Psicologia dell’Orientamento, 8(3), 17-25
[2] Caprara, G. V., Barbaranelli, C., Borgogni, L., & Perugini, M. (1993). The Big Five Questionnaire:
A new questionnaire to assess the five-factor model. , Personality and Individual Differences, 15,
281–288. http://dx.doi.org/10.1016/0191-8869(93)90218-R
[3] Fossati, A., Krueger, R. F., Markon, K. E., Borroni, S., & Maffei, C. (2013). Reliability and
validity of the Personality Inventory for DSM-5 (PID-5) predicting DSM-IV personality disorders
and psychopathy in community-dwelling Italian adults., Assessment, 20, 689 –708
[4] Sambo, F., & Franzin, A. (2015). bnstruct: Bayesian Network Structure Learning from Data with
Missing Values (p. 1.0.15), https://doi.org/10.32614/CRAN.package.bnstruct
[5] Alberto Franzin, Francesco Sambo, Barbara di Camillo. (2017) ”bnstruct: an R package for Bayesian
Network structure learning in the presence of missing data.”, Bioinformatics, 33 (8): 1250-1252;
Oxford University Press.
 COSMONET 2.0: An R Package for Survival Analysis
Using Screening-Network Methods
Maura Mecchi
University of Basilicata, Potenza, Italy [email protected]

Co-author: Antonella Iuliano

University of Basilicata, Potenza, Italy [email protected]

Network-based methods are becoming increasingly crucial in precision oncology and healthcare.
The advent of high-throughput technologies, coupled with advancements in the quantitative analysis
of biomolecular data, has created new opportunities to investigate the mechanisms driving the onset and
progression of complex diseases.
However, in this high-dimensional setting, several challenges arise. These include data heterogeneity,
limited samples relative to the number of variables, multicollinearity between variables, and the need
to integrate a priori biological information into the analysis. Equally important are the interpretation
and validation of the results, which are essential for ensuring the reliability and clinical relevance of the
findings.
Innovative statistical approaches are being developed to address some of these challenges. These
methods aim to improve the accuracy and robustness of data analysis, enabling more reliable insights
into complex biological processes and disease mechanisms. Among these, COSMONET (COx Survival
Methods based On NETworks), introduced in [1], is an R package that integrates both biologically
driven and data-driven screening techniques within a network-penalized Cox regression model. This
approach allows for more accurate identification of key biomarkers while accounting for the complex
interdependencies in biological networks (see [2, 3]). Here, we present COSMONET 2.0, an extended
version that provides a comprehensive workflow, covering the entire process from data preprocessing to
gene signature selection and survival outcome prediction. This enhanced version incorporates additional
features, such as clinical variables. It includes implementation improvements that support more robust
analysis, enabling the practical application of network-based methods to multi-omics data in survival
analysis. In addition, COSMONET 2.0 introduces new functions for data preprocessing, visualization,
survival prediction, and gene enrichment analysis, making it a powerful tool for integrating omics data in
cancer survival analysis. These enhancements enable a more comprehensive approach to understanding
the molecular underpinnings of cancer and predicting patient outcomes with increased accuracy and
reliability. Moreover, the new version of the software is significantly faster in terms of computational
costs.
We illustrate the package’s efficiency using several cancer datasets from the GDC data portal
(https://portal.gdc.cancer.gov) to evaluate its prediction accuracy under a large set of conditions. Var-
ious performance measures, including the concordance index (C-index) and other relevant metrics, are
applied to assess the package’s ability to reliably predict survival outcomes.

References
[1] Iuliano, A., Occhipinti, A., Angelini, C., De Feis, I., & Lió, P. (2021). Cosmonet: An R package for
survival analysis using screening-network methods, Mathematics, 9(24), 3262.
[2] Fan, J., Feng, Y., & Wu, Y. (2010). High-dimensional variable selection for Cox’s proportional
hazards model, In Borrowing strength: Theory powering applications–a Festschrift for Lawrence D.
Brown (Vol. 6, pp. 70-87). Institute of Mathematical Statistics.
[3] Sun, H., Lin, W., Feng, R., & Li, H. (2014). Network-regularized high-dimensional Cox regression
for analysis of genomic data, Statistica Sinica, 24(3), 1433.
 Efficiency-driven 3D CNN architectures for hyperspectral
classification
Giuseppina Monteverde
Department of Basic and Applied Sciences for Engineering - Sapienza University of Rome

[email protected]

Vittoria Bruni, Domenico Vitulano

Department of Basic and Applied Sciences for Engineering - Sapienza University of Rome

[email protected], [email protected]

Hyperspectral imaging enables the simultaneous capture of spatial and spectral information across
multiple wavelengths, yielding high-dimensional data suitable for a wide range of applications. 3D Con-
volutional Neural Networks (CNNs) can completely exploit the hyperspectral data structure through 3D
convolutional filters, which jointly extract spatial and spectral features. This process improves classifi-
cation performance by increasing intraclass variation and reducing interclass variation [1]. On the other
side, the high computational cost of deep CNN architectures — both in terms of resource consumption
and training time — when processing such high-dimensional data necessitates optimization techniques.
These can be approached through dimensionality reduction or more efficient network architectures [2].
The former reduces the input dimensionality by transforming the data into a lower-dimensional yet
representative form, while the latter focuses on streamlining the network architectures.
Two distinct approaches for enhancing hyperspectral classification efficiency using 3D CNNs are
proposed. The first method employs feature extraction, projecting the data in a proper domain and
automatically selecting relevant components in the transformed space based on the entropic normalized
information distance. This approach is an adaptive and automatic method where the number of features
to be selected is not pre-defined but automatically given [3]. The second methodology focuses on deter-
mining the filters size setting of convolutional layers in a 3D CNN, guided by Heisenberg’s uncertainty
principle. This principle inspires a rule for relating the spatial and spectral dimensions of convolutional
filters as the network depth increases, enabling the network to learn discriminative features that cap-
ture both fine spatial resolution and broad spectral characteristics [4]. The effectiveness of CNNs in
the proposed approaches is assessed using both raw and transformed input data. Both the features se-
lected by the entropy-based method and the architectures with Heisenberg-based cascaded filter setting
demonstrate a significant reduction in training time while preserving high classification accuracy. These
strategies provide solutions for processing hyperspectral data, aimed at enhancing operational efficiency.

References
[1] M. Ahmad, A. M. Khan, M. Mazzara, S. Distefano, M. Ali and M. S. Sarfraz, A Fast and Compact
3DCNN for Hyperspectral Image Classification, in IEEE Geoscience and Remote Sensing Letters,
vol. 19, pp. 1-5, 2022
[2] H. Fırat, M. E. Asker, D. Hanbay, Classification of hyperspectral remote sensing images using
different dimension reduction methods with 3D/2D CNN, Remote Sensing Applications: Society
and Environment, vol. 25, 2022
[3] V. Bruni, G. Monteverde, D. Vitulano, An Entropy Based Speed Up For Hyperspectral Data Clas-
sification Via CNNn, in 2022 12th Workshop on Hyperspectral Imaging and Signal Processing:
Evolution in Remote Sensing (WHISPERS), 2022
[4] V. Bruni, G. Monteverde, D. Vitulano, Heisenberg principle-inspired filters and size setting in
3D CNN for hyperspectral data classification, accepted for publication in 2024 14th Workshop on
Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS), 2024
 Bi-level algorithm for optimizing hyperparameters in
penalized NMF
Laura Selicato
Water Research Institute – National Research Council (IRSA-CNR), Viale Francesco de Blasio, 5,

Bari, Italy [email protected]

Over the past decade, machine learning has emerged as one of the main innovation drivers. Its
research community is expanding at an unprecedented speed, thanks to the growing need to build
accurate, reliable, and interpretable models that respond to the multitude of data generated. All the
learning algorithms require the configuration of hyperparameters (HPs), i.e., parameters that govern the
learning approach. HPs tuning is a crucial problem in the field of the learning process since the selection
of the HPs has an important impact on the final performance of the algorithm. The main goal of the
hyperparameter optimization (HPO) problem is to automate the search process, thereby improving the
generalization performance of the model and enabling a more flexible design of the underlying learning
algorithms. A reliable approach is to transform the HPO into a bi-level optimization problem that can
be solved by gradient descent techniques. The challenge is the estimation of the gradient with respect to
the HPs. In this work, we present a new mathematical framework for solving the HPO in Nonnegative
Matrix Factorization (NMF) based on bi-level techniques, focusing on penalty HPs, which turn out to be
useful to emphasize intrinsic properties in the data, such as sparsity. We design a novel algorithm, named
Alternating Bi-level (AltBi), which incorporates the HPO into the updates of NMF factors. Finally, we
provide results of the existence and convergence of solutions with also numerical experiments.
This is a joint work with Nicoletta Del Buono and Flavia Esposito (Department of Mathematics,
University of Bari Aldo Moro, Bari, Italy).
 Hybrid knowledge and data-driven approaches
for Diffuse Optical Tomography reconstruction
Alessandra Serianni
University of Milan, [email protected]

Diffuse Optical Tomography (DOT) is a non-invasive medical imaging technique which employs Near-
Infrared (NIR) light to recover the spatial distribution of optical coefficients in biological tissues. Due to
the limited availability of boundary measurements and the intense light scattering, DOT reconstruction
is a severely ill-posed problem [1]. Recently, the success of deep learning methods has shifted the focus
of tomographic imaging from purely knowledge-driven to data-driven approaches.
In this contribution, we propose a hybrid approach that combines model-based and deep learning tech-
niques. Our idea is to leverage Graph Neural Networks (GNNs), that -once trained- we use as a fast
forward model that solves partial differential equations, into an iterative optimization-based method for
solving the inverse problem. Due to the severe ill-conditioning of the reconstruction problem, we also
learn a prior over the space of solutions using an autoencoder-type neural network which maps the latent
code to the estimated physical parameter, that is passed to the GNN to obtain the prediction. The latent
code is finally optimized to minimize the difference between the recorded and predicted data.
By optimizing the latent code, we constrain the solution space to the manifold learned by the generative
model. In order to add greater structure and meaning to the latent space, we learn a compact and
non-degenerate intrinsic manifold basis [2] and the rank of the covariance matrix of the latent space is
implicitly minimized [3], while encouraging better reconstructions.

References
[1] A. Benfenati, G. Bisazza, P. Causin, A Learned SVD approach for Inverse Problem Regularization
in Diffuse Optical Tomography, arXiv preprint arXiv:2111.13401, (2021)
[2] K. Flouris, E. Konukoglu, Canonical normalizing flows for manifold learning Proceedings of the
37th International Conference on Neural Information Processing Systems, (2024), pp. 27294 - 27314
[3] J Mounayer, S Rodriguez, C Ghnatios, C Farhat, F Chinesta, Rank Reduction Autoencoders–
Enhancing interpolation on nonlinear manifolds, arXiv preprint arXiv:2405.13980, (2024)
Spatial Informed Hierarchical Clustering for Hyperspectral
Imagery via Total Variation
Gaetano Settembre∗,a , Nicoletta Del Buonoa , Flavia Espositoa
Nicolas Gillisb
[a] Department of Mathematics, University of Bari Aldo Moro, Italy
[b] Faculty of Engineering, University of Mons, Belgium

{gaetano.settembre, nicoletta.delbuono, flavia.esposito}@uniba.it,

[email protected]

Hierarchical clustering algorithms offer powerful tools for hyperspectral image analysis, reflecting
the inherent hierarchical structure of materials within images. Despite their potential, existing mod-
els often neglect critical image properties, such as the spatial similarity and proximity of neighboring
pixels. Building on the H2NMF algorithm proposed in [1], which employs a rank-two nonnegative ma-
trix factorization for binary cluster splitting, we propose two key improvements to enhance clustering
performance.
Firstly, we refine the estimation of the basis matrix W . While the original approach relies on the
successive projection algorithm, we employ more robust and advanced variants such as the smoothed
successive projection algorithm (SSPA) and the smoothed vertex component analysis (SVCA) [2]. These
methods address the limitations of the pure pixel assumption by better identifying the vertices of the
convex hull of the data, even in noisy conditions.
Secondly, we incorporate Total Variation (TV) regularization [3] into the objective function to im-
prove the estimation of the coefficient matrix H. This regularization exploits the spatial structure within
hyperspectral images, promoting smoother and spatially coherent solutions while preserving critical edge
information. The new objective function is defined as:
r
X
min ∥X − W H∥2F + λ ∥SH(ℓ, :)∥1 ,
H≥0
ℓ=1

where X represents the original hyperspectral image, W is derived from the aforementioned methods.
In our case r = 2 and S ∈ RK×n is a sparse matrix encoding pixel neighborhood relationships such that
S(k, i) = 1 and S(k, j) = −1 for some k if pixels i and j are neighbors. We solve this new optimization
problem using an iterative gradient-based approach.
Several experiments are conducted on different real remote sensing hyperspectral datasets (e.g.,
Cuprite, Urban, Samson, etc) to evaluate the convergence curve of the algorithm, and then the effective-
ness of our new proposed clustering method.

References
[1] Gillis, N., Kuang, D. & Park, H. Hierarchical Clustering of Hyperspectral Images Using Rank-Two
Nonnegative Matrix Factorization. IEEE Transactions On Geoscience And Remote Sensing. 53,
2066-2078 (2015,4), DOI: 10.1109/TGRS.2014.2352857.
[2] Nadisic, N., Gillis, N. & Kervazo, C. Smoothed separable nonnegative matrix factorization. Linear
Algebra And Its Applications. 676 pp. 174-204 (2023,11), DOI: 10.1016/j.laa.2023.07.013.
[3] Rudin, L., Osher, S. & Fatemi, E. Nonlinear total variation based noise removal algorithms. Physica
D: Nonlinear Phenomena. 60, 259-268 (1992,11), DOI: 10.1016/0167-2789(92)90242-F.
 Empowering Clinicians with Explainable AI: Predicting
Mortality Risk in MAFLD with Counterfactual Analysis
Paolo Sorino, Domenico Lofù, Rossella Donghia, Caterina Bonfiglio,
Gianluigi Giannelli, and Tommaso Di Noia
Address [email protected], [email protected],

[email protected], [email protected],
[email protected], [email protected]

Metabolic Dysfunction Associated with Fatty Liver Disease (MAFLD) represents a paradigm shift
in liver disease classification, moving from the concept of a “non-condition” to an inclusive diagnostic
entity. Introduced in 2020, MAFLD diagnosis is based on the presence of hepatic steatosis along with
one of three metabolic conditions: overweight or obesity (Subtype 1), metabolic dysregulation in lean
individuals (Subtype 2), or diabetes mellitus (Subtype 3). As MAFLD is increasingly recognized as a
public health concern, there is an urgent need for innovative approaches to improve early detection
and management. In this context, Machine Learning (ML) has emerged as a game-changing technology
in modern clinical practice, offering the capability to extract actionable insights from complex, high-
dimensional datasets. By leveraging sophisticated algorithms, ML enables clinicians to address critical
challenges such as early disease diagnosis, accurate risk stratification, and the development of personalised
treatment strategies, making ML an indispensable tool for tackling multifaceted health problems such as
MAFLD. To address the early identification of high-risk patients, we developed MORIX, an artificial
intelligence-based framework for predicting mortality risk in individuals with MAFLD. The study cohort
consisted of 1, 675 subjects (543 females and 1, 132 males) aged > 30 years, diagnosed with MAFLD
and recruited between May 2005 and January 2007 from the National Institute of Gastroenterology,
IRCCS ‘S. De Bellis’ in Castellana Grotte (Italy). The cohort was observed until December 31, 2023.
Using this dataset, which included anthropometric and biochemical parameters, we applied Recursive
Feature Elimination (RFE) with a Random Forest (RF) model to select the most relevant features. These
features were then used to train and evaluate five machine learning algorithms—Random Forest (RF),
eXtreme Gradient Boosting (XGB), Support Vector Machine (SVM), Multilayer Perceptron (MLP), and
Light Gradient Boosting Machine (LGBM)—using a 5-fold cross-validation approach. Among the tested
models, RF demonstrated the highest performance, achieving an accuracy of 83%, with a precision
and recall of 83% for mortality prediction, and an F1 score of 0.83. The Area Under the ROC Curve
(AUC) was 0.88, confirming the RF model’s ability to effectively distinguish between high- and low-risk
patients. In comparison, XGB and SVM achieved slightly lower accuracies of 82% and 80%, while MLP
and LGBM showed weaker results overall.
In addition, explainability was a core component of the MORIX framework. Explainable Artificial
Intelligence (XAI) techniques, specifically Shapley Additive exPlanations (SHAP), were applied to the
RF model to make the decision-making process transparent. SHAP values revealed that age and blood
glucose were the most critical predictors of mortality, providing clinicians with clear insights into the
model’s decision-making process.
Furthermore, MORIX includes a counterfactual analysis feature, enabling clinicians to simulate
“what − if ′′ scenarios. For instance, modifying biochemical parameters, such as cholesterol or weight,
allows users to observe how these changes influence the predicted mortality risk. This capability offers
actionable insights, supporting targeted interventions to improve patient outcomes.
To ensure accessibility, we developed a user-friendly web application that integrates the trained RF
model. This application enables healthcare professionals to input new patient data, receive real-time
mortality risk predictions, and access detailed explanations of the model’s decisions.
In conclusion, MORIX exemplifies how ML can bridge the gap between complex data and practical
clinical applications. By combining robust predictive performance with explainable AI and counter-
factual analysis, MORIX offers a valuable tool for clinicians to make informed, data-driven decisions.
Its integration into clinical workflows has the potential to enhance patient care by identifying high-risk
MAFLD patients early and providing actionable insights into improving outcomes. Future work will
focus on expanding the dataset to include additional clinical variables and exploring the use of Deep
Learning (DL) to further enhance model performance.