Past talks

Abstract: Generative models such as Vision-Language-Action (VLA) policies are reshaping robot learning, enabling large-scale pretraining and strong generalization. Yet, these models often struggle in real-world settings due to limited physical grounding, low control frequency, and reliance on demonstrations.

 In this talk, I present a unifying perspective where reinforcement learning (RL) complements imitation learning (IL) by grounding pretrained generative policies through interaction. Building on recent work on the duality between generative models and RL in robotics , I argue that RL is essential not as an alternative, but as a refinement mechanism. I introduce the concept of transient guidance, where generative models provide sparse action priors to accelerate exploration without constraining final performance. This is demonstrated in our Vision-Language-Action Jump-Starting (VLAJS) framework, where RL leverages low-frequency VLA guidance to improve sample efficiency and sim-to-real transfer . Finally, I discuss emerging directions such as RL-based fine-tuning of generative policies and critic learning for action generation, outlining a path toward scalable, data-efficient, and physically grounded robot learning systems.

Short bio: Angelo is a third-year PhD student at IDSIA in the LEON-Robotics group under the supervision of Prof. Loris Roveda. His research focuses on reinforcement learning, imitation learning, and generative models for robot control, with particular emphasis on grounding Vision-Language-Action policies through interaction. Angelo obtained his Bachelor’s and Master’s degrees in Automation and Control Engineering from Politecnico di Milano. His current work explores how reinforcement learning can improve generative robot policies, including methods for transient guidance, RL-based fine-tuning, and the integration of structured and generative representations for scalable robot.

Abstract: Many modern datasets arise from scattered and non-uniform samples in high-dimensional spaces, where traditional grid-based methods lose efficiency and accuracy. Samplets provide a principled mathematical framework for representing such data across multiple scales, enabling both theoretical insight and computational efficiency. Inspired by wavelets but tailored for scattered data, samplets act as localized building blocks that capture information at multiple resolutions. They make it possible to compress and manipulate large numerical systems with almost linear computational cost, turning previously intractable problems into manageable ones.
This multiscale approach shines in applications such as kernel interpolation with Matèrn functions, where data are reconstructed by successively refining details across different length scales. Representing each step in samplet coordinates produces sparse, well-conditioned systems that can be solved stably and efficiently.
Samplets are also very powerful in data analysis. By observing how samplet coefficients decay within the multiresolution levels, one can measure the smoothness and detect sharp transitions within irregular signals in near-linear time. In essence, samplets bridge structure and scalability, providing a multiresolution language for understanding and processing complex, scattered data.

Short bio: Sara Avesani is a PhD student in computational science at Università della Svizzera Italiana (USI), where she has been enrolled since 2023. Her research focuses on multiscale approximation methods for scattered data, with particular emphasis on meshfree, adaptive multiresolution algorithms for detecting local smoothness classes and for the numerical solution of partial differential equations. She actively contributes to the C++ implementation of these methods within the FMCA (Fast Multiresolution Covariance Analysis) library, developing efficient and user-friendly components that combine high performance with flexible, application-oriented interfaces. Her contributions bridge theoretical advances in kernel methods and reproducing kernel Hilbert spaces with practical, scalable implementations for large-scale scientific computing. 

Abstract: Reinforcement Learning (RL) is an effective paradigm for addressing complex decision-making problems. However, traditional deep RL approaches often struggle in environments characterized by multiple interacting components. These settings introduce additional challenges in the learning procedure, such as coordination and scalability. Policies and value functions, i.e., the building blocks of RL algorithms, should be designed to account for the complex interactions within the learning environment while maintaining computational efficiency. In this context, graphs represent a natural formalization to describe systems with entities interconnected by complex relationships and dependencies. This presentation explores recent advances in graph-based methodologies for RL and investigates promising research directions.

Short bio: Tommaso Marzi is a PhD student at IDSIA USI-SUPSI, working in the Graph Machine Learning Group under the supervision of Cesare Alippi. He received a Bachelor's degree in Physics and a Master's degree in Applied Physics from the University of Bologna. His main research topic is reinforcement learning, with a particular focus on structured domains. His personal webpage is https://tommasomarzi.github.io/

Abstract: In hospital management, operating room scheduling comprises strategic, tactical, and operational decision levels, ranging from long-term resource planning to patient assignment and sequencing. In this study, we address a joint Master Surgical Scheduling and Surgical Case Assignment problem in a weekly surgery setting. The model explicitly accounts for hospital bed management, with the objective of maximizing patient priority. The problem is solved using a column-generation approach, where each subproblem is formulated as a shortest-path problem and solved using the PathWyse library.

Short bio: Sara Cambiaghi is a PhD candidate in the joint PhD program in Computational Mathematics and Decision Sciences, established by the University of Pavia, Pavia, Italy, and the Università della Svizzera Italiana (USI), Lugano, Switzerland, under the supervision of Prof. Davide Duma and Prof. Stefano Gualandi. She received both her bachelor's and master's degrees in Mathematics from the University of Milano-Bicocca (UniMiB), Milan, Italy. Her research focuses on operations research applied to healthcare.

Abstract: Self-supervised robot learning (SSRL) is a branch of deep learning that focuses on enabling robots to autonomously collect and label their own training data. Because robotic tasks require large volumes of high-frequency, dense real-world data, manual labeling by humans is often impractical. In SSRL, robots use onboard sensors to generate labeled datasets, which are then used to train deep learning models for specific tasks.

Short bio: Nicholas Carlotti is a PhD student at the robotics group under the supervision of Alessandro Giusti. His research focuses on self-supervised robot learning for perception tasks. Prior to his doctoral studies, he obtained a BS degree in computer science at the University of Bologna, and a MSc double degree program in computer science at University of Milano Bicocca and Universita' della Svizzera Italiana

Abstract: Software systems are used in virtually every aspect of modern life. For safety-critical applications, such as automotive, avionics, or medical systems, it is essential to ensure that software does not malfunction or contain significant vulnerabilities. Various approaches exist to ensure software functional correctness; however, only formal verification can provide mathematical guarantees regarding the behavior of software systems with respect to their requirements. One of the major challenges in automated software verification is the complexity of real-world programs. To make verification feasible, it is essential to abstract the system being analyzed, while preserving the essential details related to the verified properties. The goal of this work is to achieve a better balance between abstract and concrete representation of programs, utilizing the structural properties of software to improve the efficiency and performance of automated formal methods. 

Short bio: Konstantin Britikov is a PhD candidate in the Faculty of Informatics at the Università della Svizzera italiana (USI), Lugano, under the supervision of Prof. Natasha Sharygina. He earned a degree in Computer Science in Bauman Moscow State Technical University and a master's degree in Software Engineering from the Innopolis University in Russia. His research centers on Symbolic Model Checking, in particular in application to Software systems.

Abstract: Higher-order abstract Voronoi diagrams (AVDs) are a combinatorial framework unifying many concrete variants of Voronoi diagrams. Their complexity has been bounded by characterizing the unbounded edges using a class of cyclic sequences of permutations. We extend this framework by adding colors to the sites. Through a careful analysis of the resulting colored permutations, we establish tight bounds on the number of unbounded edges in higher-order color abstract Voronoi diagrams. This yields new bounds for concrete Voronoi diagrams.

Short bio: Nicolau Oliver Burwitz is a PhD candidate in the Faculty of Informatics at the Università della Svizzera italiana (USI), Lugano, under the supervision of Prof. Evanthia Papadopoulou. He earned a degree in Informatics and a master's degree in Applied Mathematics from the Universitat Politècnica de Catalunya (UPC). His research centers on Computational Geometry, currently on generalizations of Voronoi diagrams in the plane.

Abstract: Accurate state estimation is essential for the control and monitoring of nonlinear systems, where full-state measurements are rarely available. In this talk, we explore the use of Physics-Informed Neural Networks (PINNs) to learn Kazantzis–Kravaris–Luenberger (KKL) observers for autonomous nonlinear systems. The proposed framework jointly learns the lifting transformation and the contractive dynamics in a higher-dimensional space, enabling state reconstruction through a learned inverse map. We present our first validations through numerical simulations on benchmark systems, illustrating the potentials and current challenges.

Short bio: Clara Galimberti is a postdoctoral researcher at IDSIA within Laura Azzimonti's group, since March 2025. She received the Ph.D. degree from École Polytechnique Fédéral de Lausanne in 2024, where she worked at the intersection of control theory and neural networks, and the degree in electronic engineering from the Universidad Nacional de Rosario, Argentina, in 2018. Her research focuses on Physics-Informed Neural Networks for system identification, estimation and control. 

Abstract: Elementary Resource Constrained Shortest Path Problems (ERCSPPs) are central to optimization and transportation research. They involve finding the shortest path under resource constraints while ensuring that nodes are not revisited. The problem is strongly NP-hard and frequently arises in orienteering tasks as well as in subproblems of Vehicle Routing and Crew Scheduling. In this study, we explore whether machine learning can be used to identify promising states within dynamic programming algorithms, commonly applied to solve ERCSPPs, with the ultimate goal to improve search decisions. We start by solving 41 single-resource instances from SPPRCLIB using iterative relaxation approaches via the PathWyse library, collecting constant-time features for all generated dynamic programming states. These features allow us to build two large datasets, totaling several hundred million labels. We then apply machine learning to analyze them, uncovering patterns across successive relaxations. Using these insights, we develop normalization techniques and supervised models to identify dominating states, both within the same problem and in new instances. Finally, we use these results to propose an initial design for a data-driven dynamic programming algorithm, setting the stage for further refinements and extensions.

Short bio: Saverio Basso is a postdoctoral researcher at IDSIA, where he has been working since September 2021. He earned his PhD, MSc, and BSc in Computer Science from the University of Milan. His doctoral thesis focused on data-driven approaches for generating Dantzig-Wolfe decompositions of Mixed-Integer Problems, as well as on the design of parallel and distributed column generation algorithms to solve them. His primary research interests lie in exact and data-driven approaches to resource-constrained shortest path problems and, more broadly, in combinatorial optimization.

Abstract: A deep understanding of cause-effect relations is necessary to define effective treatment strategies. Clinical trials are the gold standard for deriving these relations; however, most times they are not feasible for several reasons: ethical, technical, and practical. In these cases, observational data offers an important alternative to recover these effects. While largely accessible, the naive application of traditional AI/ML models can lead to critically wrong and damaging conclusions. Causal inference frameworks can guide machine learning models to derive reliable, realistic, and clinically aligned conclusions based on the available data. 

Short bio: Giovanni Angelotti is a Doctoral Researcher at the Dalle Molle Institute for Artificial Intelligence USI-SUPSI in Lugano, Switzerland, focusing on neural-causal inference. His research aims to develop inference and reasoning models to enhance risk mitigation strategies and support high-stakes decision making based on real world evidence. Giovanni has an extensive expertise in applied research on electronic health records and multimodal clinical data. He has worked with data from health facilities across three continents, in both industry and academia. Prior to his graduate studies, Giovanni obtained his Master’s Degree from Politecnico di Milano in Biomedical Engineering at SPINLabs in 2017, after developing his thesis at MIT LCP in Boston. He then worked as Lead Data Scientist at Humanitas Research Hospital in Milan.
Giovanni co-organized the first two editions of ESICM Datatalk and MIT Datathon in Milan in 2019 and 2020, amongst the largest AI events in critical care in Europe at the time. Giovanni is a research affiliate at MIT Critical Data, a global ecosystem of healthcare and data science professionals advocating for AI equity in research. He is also a member of the IDSIA Imprecise Probability Group and of the Machine Learning for Bioinformatics and Personalized Medicine Group at the Swiss Institute of Bioinformatics.

Abstract: A combinatorial optimization problem aims at optimizing a function over a discrete set, subject to some constraints. Since these problems are generally very hard to solve exactly, it is common to consider relaxations: variations of the original problem that are easier to solve and provide approximate solutions. We will focus on Semidefinite Programming (SDP) relaxations, a powerful technique that has been used for approximating combinatorial optimization problems ever since the celebrated result of Goemans and Williamson (1995) for MAXCUT.  In this context, the Sum-of-Squares (SoS) hierarchy has emerged as a systematic and flexible method for constructing these relaxations. However, some key questions remain unresolved. For example, it is still unclear under what conditions SoS can be automated, meaning whether one can compute the level-d SoS relaxation in polynomial time. O’Donnell (2017) observed that the prevailing belief about the automatability of SoS using ellipsoid-based algorithms is not entirely correct. In this presentation, we will discuss about the SoS relaxations and its automatability. Further, we establish new conditions that guarantee the automatability of SoS relaxations. Lastly, we will show how these conditions lead to dichotomy results regarding the applicability of SoS for solving 0/1 problems and for optimization over 0/1 domains.

Short bio: Alex Bortolotti is a PhD student in the Optimization and Complexity group under the supervision of Monaldo Mastrolilli. His research interests include Sum-of-Squares optimization, algebraic and semi-algebraic proof complexity, and computational algebraic geometry. He holds bachelor's and master's degrees in pure and applied mathematics from the University of Bologna.

Abstract: The release of OpenAI’s GPT-4o introduced a highly capable multimodal model that engages users across voice, vision, and text. Among its most debated features was the deployment of a feminized voice that responded with flirtatious tone, giggles, and affective warmth, design choices that blur the boundaries between technical function and emotional interaction. Often dismissed as surface-level interface matters, such decisions reveal deeper ethical implications, highlighting the need for critical engagement with how AI systems shape human experience.

Drawing from design activism and human-computer interaction (HCI) theory, we propose a conceptual framework that identifies four distinct roles — Reformer, Change Agent, Citizen, and Rebel — as modes of intervention within AI system development. Each role provides a unique pathway for resisting harmful design norms and promoting more socially responsible computing. Through this lens, we call for a broader, interdisciplinary understanding of human-AI interaction, one that foregrounds the sociotechnical nature of AI and the imperative for collaborative, ethically grounded design practices.

Short bio: Irene Zanardi is a PhD student at USI-IDSIA in the LUXIA lab, supervised by Prof. Monica Landoni. Her research investigates how to support children's critical reflection when interacting with generative AI. With a background in Interaction Design from Politecnico di Milano, she focuses on designing interactions that make AI’s role more visible, accountable, and open to questioning.

Abstract: In combinatorial optimization, various problem-solving approaches build on a high-level framework called "branch-and-bound" algorithm. It is especially true when the underlying problem is NP-hard, and the lack of polynomial-time algorithms leaves no option other than resorting to alternatives with an exponential running time. However, even though these branch-and-bound -based methods perform surprisingly well in practice, we have a limited understanding regarding theoretical explanations of this phenomenon. In the presentation, we intend to provide a possible reasoning by revealing a connection to approximability theory; another popular framework in which algorithms can terminate as soon as they find a good enough solution.

Short bio: As a PhD student,  Koppány is affiliated with the Optimization and Complexity research group spearheaded by Monaldo Mastrolilli and Luca Gambardella. His research interests include polyhedral combinatorics, graph theory and algorithms. Prior to joining the group, he had obtained his Master's and Bachelor's in Mathematics in Hungary.

Abstract: Displays are the primary medium for conveying information and play an important role in entertainment, enabling technologies such as 3D cinema, HDR projectors, and AR/VR devices. Among these, VR is promising for enhancing cinematic experiences, telepresence, and simulations. Ideally, these devices should deliver content indistinguishable from the real world. However, achieving such realism is computationally demanding, as high-resolution content, a wide field of view, and high frame rates require significant computational resources from both software and hardware. These constraints affect the design and ergonomics of VR devices, influencing their size, comfort, and widespread adoption. Efficiency is therefore important to make these devices more portable, especially for battery-powered versions. In this talk, we focus on the power efficiency of VR display systems and introduce a technique that dynamically adjusts screen brightness based on the displayed content. Other than optimizing power consumption, this approach preserves visual quality by maintaining spatial contrast (e.g., visible details) while preventing perceptible artifacts such as flicker and temporal instability. 

Short bio: Luca Surace is a PhD student at USI-IDSIA in the Perception, Display and Fabrication group, supervised by Prof. Piotr Didyk. He works on leveraging visual perception for computer graphics applications. He did research internships at Disney Research Zurich on color grading (during PhD) and University of Plymouth on emotion recognition (during MSc). Before his PhD, he worked on the Air-Cobot project and obtained a MSc in computer science from Università della Calabria. 

Abstract: In recent years, technological advancements have led to the generation of vast amounts of genetic data, offering new opportunities for analysis and providing unprecedented insights into genomics, particularly in relation to diseases. Despite this progress, significant challenges remain, especially in the study of complex disorders. Unlike monogenic diseases, which are caused by mutations in a single gene, complex diseases involve multiple genetic loci, with genetic contributions to disease development often being non-trivial. Understanding the intricate role of genetics in these conditions remains a critical task. In this context, AI methods are implemented to identify meaningful patterns in genetic data. This talk will explore how AI can be applied to the study of both monogenic and complex diseases, advancing our understanding of how genetics contribute to their development.

Short bio: Francesco Gualdi is currently a Postdoctoral Researcher at the Istituto Dalle Molle di Studi sull’Intelligenza Artificiale (IDSIA) in Lugano, Switzerland. With a background in Medical Biotechnology, his research is focused on analyzing genetic and medical data to gain deeper insights into complex disorders. He specializes in the development and application of AI algorithms to investigate the genetic underpinnings of human conditions. Francesco also holds a diploma in Classical Guitar from the Conservatory of Ferrara.

Abstract: Detecting when a neural sequence model does “interesting” computation is an open problem. Next token prediction loss is a poor indicator: Low loss can stem from trivially predictable sequences that are uninteresting, while high loss may reflect unpredictable but also irrelevant information that can be ignored by the model. We propose a new metric: Measuring how well a model can predict its own future hidden states. We do this by introducing the prediction of hidden states (PHi) layer, which serves as both an information bottleneck and a prediction module. This layer is architecture-agnostic and can be either trained from scratch or added post-hoc to large pre-trained models. By incentivizing the model to keep its hidden states both predictable and compressed, it encodes unpredictable information only when it is expected to be relevant for subsequent tokens. We argue that complex and interesting tasks are those containing unpredictable yet relevant information, and we show empirically that PHi loss–in contrast to the next token prediction loss–correlates with task complexity.In particular, we show that, for in-context learning of formal languages, our metric predicts the description length in both small Transformers/RNNs and large pre-trained language model

Short bio: Vincent Herrmann is a Ph.D. student in Jürgen Schmidhuber’s group at the Dalle Molle Institute for Artificial Intelligence Research (IDSIA) in Lugano, Switzerland. His research interests include artificial creativity, curiosity and open-endedness. Vincent received a Master of Arts in Music Informatics from the University of Music Karlsruhe and a Master of Arts in Piano Performance from the University of Music and Performing Arts Stuttgart.

Abstract: Can the intricacies of a dynamical system be understood not only from its input/output patterns but also by observing the behavior of similar systems? This talk aims to explore this question by introducing the concept of similarity between systems through the lens of Meta-Learning and System Identification. We will present a novel paradigm for System Identification: a meta-model that leverages the power of the transformer architecture to represent a class of dynamical systems. This approach opens new possibilities for understanding and modeling complex systems by drawing insights from similar ones. The presentation aims to encourage an interactive discussion on potential future research directions, exploring applications, theoretical advancements, and challenges. Attendees will also be introduced to relevant perspectives from various research domains, providing resources for deeper exploration of this emerging and interdisciplinary topic.

Short bio: Matteo is a PhD student in the LEON group at IDSIA, working under the supervision of Marco Forgione and Dario Piga. He holds a degree in Mathematical Engineering from the Polytechnic University of Turin. Matteo's research focuses on machine learning and system identification. Before starting his PhD, he gained valuable research experience through internships at BCAM in Bilbao and CERN in Geneva. Outside of academia, Matteo enjoys cooking, playing volleyball, and cheering for the Napoli football team.

Abstract: Intermittent time series are common in applications such as retail chain supply and spare parts replenishment. They are characterised by a large number of zeros; it is therefore necessary forecasting both the occurrence and the level of the demand. Most approaches for forecasting intermittent time series only provide point forecasts. In this talk, some properties of the main probabilistic forecasting models for intermittent series will be discussed, along with the assumptions that characterize them. Our model, Tweedie GP, which aims to overcome some critical issues, will also be presented. This model exploits a predictive distribution from the Tweedie family, which is zero inflated and has long tails, while the structure of dependencies between different timestamps is learned through a Gaussian process (GP), a model of Bayesian nonparametric statistics. Finally, the main forecast evaluation methods for intermittent series will be discussed, and will be used to compare the performance of the models discussed above.

Short bio: Stefano Damato is PhD student at IDSIA USI-SUPSI, under the supervision of Giorgio Corani and Dario Azzimonti. He holds a Master's degree in Mathematics from Universitá degli Studi di Torino. His research interests include machine learning, Bayesian statistics, and forecasting. His non-research interests include cats and tennis.

Abstract: In the field of Numerical Analysis, it is common practice to solve an arbitrary mathematical problem through the implementation of a numerical algorithm on a computer. Since not all data of a problem can be represented exactly on a computer as floating-point numbers, it can happen that the algorithm starts with rounding errors already from the initial set of input that then propagate throughout the process. Consequently, it is important to study the numerical stability of an algorithm, that is, its behaviour with respect to the propagation of the errors that occur during the arithmetic operations executed by the computer. In this talk, we focus on the barycentric interpolation problem, both in the univariate and the bivariate setting. In the first case, we theoretically discuss the numerical stability of all algorithms that implement a barycentric rational interpolant, providing conditions under which it is possible to know a priori whether they are stable. In the second case, we focus on the barycentric interpolant defined on a planar polygon, and this leads to the study of the numerical stability of generalized barycentric coordinates, particularly the mean value coordinates. 

Short bio: Chiara is pursuing a PhD in Informatics at USI, where she works on the "Barycentric Interpolation" project under the supervision of Prof. Kai Hormann. Her research centers on Numerical Analysis, with a strong focus on the stability and robustness of algorithms. She applies her theoretical knowledge to practical applications, exploring and comparing methods in areas like function approximation and computer graphics. Chiara holds a Bachelor's and Master's degree in Mathematics from Università della Calabria and has completed a research internship and Erasmus program at Universitat de València.

Abstract: In this work, we critically examine the trend toward rapid and seamless human—AI interactions, focusing instead on alternative forms of engagement aimed at fostering user empowerment, skill development, and appropriate reliance on AI for responsible decision-making. This paradigm shift centers on the core concepts of 'programmed inefficiencies' and 'frictional protocols’, which involve design elements aimed at slowing down the interaction with AI to promote more thoughtful behavior. This talk will discuss design principles that balance efficiency with engagement through user studies in the medical field, as well as presenting novel methods for revealing and assessing over-reliance and under-reliance on decision support systems.

Short bio: Chiara Natali is a PhD student in Informatics at University of Milano-Bicocca and Research Fellow at IDSIA USI-SUPSI via the Swiss Government Excellence Scholarship scheme. She is also a lecturer for the Human-Computer Interaction and the Interaction Design courses at University of Milano-Bicocca (2023-2024). Her research draws from her interdisciplinary background in Political Philosophy: in particular, she focuses on the ethical challenges of human-AI interaction and eXplainable AI, particularly the socio-technical issues behind biased decision-making. Her work was awarded as the "Best Doctoral Consortium Paper" at CHItaly23 and "Best Paper" at xAI24. 

Abstract: In this talk we explore advancements in polynomial optimization using sum-of-squares (SOS) hierarchies and semidefinite programming (SDP), which offer strong relaxations for various hard optimization problems such as stability and chromatic number of a graph. Despite their potential and some result under specific assumptions, the computational complexity of these methods is still not completely understood. In fact, there are polynomials that admit SOS representations, but every such a representation requires exponentially large coefficients, preventing the ellipsoid method to run in polynomial time. Our goal is to determine whether SOS approximations for copositive programs can be computed in polynomial time, placing a strong emphasis on computational tractability. We show that certain SOS relaxations for copositive programs can be computed in polynomial time.

Short bio: Marilena is a Ph.D. student in Theoretical Computer Science at IDSIA USI-SUPSI, working in the Algorithms and Complexity group under the supervision of Professor M. Mastrolilli. Her research focuses on the computational complexity of sum-of-squares bounds for copositive programs and the bit complexity of sum-of-squares proofs. Marilena holds a Master's degree in Theoretical and Applied Mathematics from the University of Bologna and a Bachelor's degree in Mathematics from the University of Udine. Prior to her Ph.D., she gained experience as a technology consultant at Accenture and as a quantitative risk analyst at EY.

Abstract: In this study, we investigate the conformational dynamics of Omphalotin A using both classical molecular dynamics (MD) simulations and Metadynamics. By integrating these methods, we aim to obtain a comprehensive view of Omphalotin A’s structural behavior, addressing sampling limitations inherent to traditional MD. Our findings provide key insights into the molecule's conformational dynamics in both hydrophilic and hydrophobic environments, offering a deeper understanding of its biological activity and guiding future experimental and computational research.

Short bio: Filip is a PhD student in Andrea Danani's biophysics group at IDSIA. He holds a degree in Biomedical Engineering from the Polytechnic University of Turin, and his primary research areas include molecular modeling and molecular dynamics.

Abstract: 

Within a prediction task, Graph Neural Networks (GNNs) use relational information as an inductive bias to enhance the model's accuracy. As task-relevant relations might be unknown, graph learning approaches have been proposed to learn them while solving the downstream prediction task. In this paper, we demonstrate that minimization of a point-prediction loss function, e.g., the mean absolute error, does not guarantee proper learning of the latent relational information and its associated uncertainty. Conversely, we prove that a suitable loss function on the stochastic model outputs simultaneously grants (i) the unknown adjacency matrix latent distribution and (ii) optimal performance on the prediction task. Finally, we propose a sampling-based method that solves this joint learning task. Empirical results validate our theoretical claims and demonstrate the effectiveness of the proposed approach. 

Short bio: 

Alessandro Manenti is a second-year Ph.D. student at IDSIA (USI - SUPSI). After receiving a B.Sc. in Physics and an M.Sc. in Physics of Complex Systems, he began his Ph.D. under the supervision of Professor Cesare Alippi. His main research interests include Graph Deep Learning, with a particular focus on applying these techniques when the topological structure is unavailable and generalizing these methodologies to higher-order relations, such as hypergraphs. 

Abstract: 

Causal opacity denotes the difficulty in understanding the ''hidden'' causal structure underlying a deep neural network's (DNNs) reasoning. This leads to the inability to rely on and verify state-of-the-art DNN-based systems, in high-stakes scenarios. For this reason, causal opacity represents a key open challenge at the intersection of deep learning, interpretability, and causality. This work addresses this gap by introducing Causal Concept Embedding Models (Causal CEMs), a class of interpretable DNN models whose decision-making process is causally transparent by design. The results of our experiments show that Causal CEMs can: (i) match the generalization performance of causally-opaque models, (ii) support the analysis of interventional and counterfactual scenarios, thereby improving the model's causal interpretability and supporting the effective verification of its reliability and fairness, and (iii) enable human-in-the-loop corrections to mis-predicted intermediate reasoning steps, boosting not just downstream accuracy after corrections but also accuracy of the explanation provided for a specific instance.

Short bio: 

Pietro Barbiero is PostDoc at USI Software Institute and Research Fellow in the Artificial Intelligence Group at the Computer Laboratory, University of Cambridge. He studied mathematical and computer engineering at Politecnico di Torino (Italy), and obtained his PhD in Artificial Intelligence at the Computer Laboratory, University of Cambridge. His research focuses on interpretable and neural symbolic AI.

Gabriele Dominici is currently a first-year PhD student under the supervision of Marc Langheinrich at USI. He completed his Master's degree in computer science and machine learning at the University of Cambridge, and his Bachelor's degree in computer science at the Università degli Studi di Udine. His research is dedicated to enhancing the interpretability of AI models, with a focus on understanding their decision-making processes and enabling human steering models' behaviour.

Alberto Termine is researcher at IDSIA and PostDoc in AI Ethics at EPFL, College of Humanities. He studied Logic, Philosophy of Science and Cognitive Science at San Raffaele University (Milan, Italy), and obtained his PhD at the LUCI lab, Department of Philosophy, University of Milan, with a thesis on Probabilistic Model Checking. His research focuses on causal and counterfactual methods in explainable AI, model-checking and verification of trustworthy autonomous systems and the epistemology and ethics of machine learning in science. 

Abstract: 

The Premurosa project, funded by the EU, has made significant use of machine learning to examine data and offer individualized insights on various aspects of medicine and biology. This includes the use of spiking neural networks to predict toxicity, bioaccumulation, and enzyme activity, as well as data fusion to interpret biomarkers or characterize biomaterials, and anomaly detection in proteomics.

Short bio: 

Educational background in medical sciences, economics, and engineering, alongside prior experience as a data analyst and research assistant. Through my academic training I have developed a specialized skill set in analyzing biological signals and applying machine learning techniques to biomedical data. I am deeply interested in utilizing artificial intelligence to streamline processes, anticipate outcomes, and facilitate decision-making in the field of medicine.

Abstract: 

The Steiner tree problem (STP) in graphs is an extensively studied challenge in combinatorial optimization. However, to the best of our knowledge, limited attention has been given to exploring the integrality gap across different formulations. Building on related efforts in the context of the Travelling Salesman Problem, we share our ongoing work focused on the computation of the exact integrality gap for the Steiner Tree Problem. We start by introducing a novel formulation that is improving for the metric STP, and then we move to some preliminary results on the exact integrality gap computation.

Short bio: 

Ambrogio Maria Bernardelli is about to start his third year as a Ph.D. student at the University of Pavia in a joint program with USI after receiving a MSc and a BSc in Mathematics in Pavia. His main research interests concern combinatorial and stochastic optimization, having worked in particular on stochastic scheduling problems, stochastic energy dispatching problems, MIP training of neural networks, and integrality gap problems.

Abstract: 

Deep learning algorithms have found increasing applications in the biomedical domain. However, they often manifest as inscrutable black-box models, seldom harnessing the wealth of prior knowledge available for the tasks at hand. Addressing this limitation requires the incorporation of such prior knowledge into our algorithms. During this talk, I will explore this concept by delving into two illustrative examples, one leveraging a physical model and one a physiological model. The focal point of the discussion will be the first example, which draws from my doctoral research where physics-based tomographic models have been integrated with deep learning-based regularization by alternating between model-based updates and data-driven image regularization. At the end of the talk, I will offer a glimpse into our ongoing research at IDSIA on EMG-based speech synthesis. Here, the deep learning architecture itself capitalizes on prior knowledge derived from a physiologically inspired latent space model of speech articulation.

Short bio: 

Stefano van Gogh is a postdoctoral researcher at IDSIA since August 2023. He earned his PhD at ETH Zürich, while conducting research at the Paul Scherrer Institute. His doctoral thesis focused on the development of hybrid tomographic reconstruction algorithms, combining data-driven techniques with physics-based models, designed for a cutting-edge breast imaging technology called Grating Interferometry CT. Before obtaining his PhD, he received a MSc in Biomedical Engineering and a BSc in Health Sciences and Technology, both from ETH Zürich. Stefano's primary research interests center around the seamless integration of data-driven algorithms with mechanistic models (physical and physiological), in the realm of biomedical engineering.

Abstract: 

In the modern digital landscape, technological solutions are undergoing a paradigm shift towards inclusivity and interactivity to cater to diverse audiences. Despite this progress, individuals with intellectual disabilities continue to face obstacles when accessing and engaging with technology and communication. This talk explores the challenges that people with intellectual disabilities encounter when interacting with traditional environments or inaccessible technologies. Drawing on best practices and innovative examples, I will showcase the transformative role that accessible technology can play in breaking down these barriers using cutting-edge technologies, such as Multisensory Experiences, Artificial Intelligence, and Social Robots. If designed and used appropriately, these technologies not only empower individuals with intellectual disabilities to explore the world at their own pace but also create opportunities for meaningful and personalized learning experiences.

Short bio: 

Leandro is a Ph.D. Candidate in Informatics at Università della Svizzera italiana - USI (Switzerland) with Doctoral Exchange Program in Computer Science at Queensland University of Technology - QUT (Australia). He holds an M.Sc. in Computer Science from Federal University of Rio Grande do Sul - UFRGS (Brazil) and a B.Sc. in Computer Science from Federal University of Pelotas - UFPEL (Brazil) with an exchange program at University of Porto - U.Porto (Portugal). His current research focuses on exploring the use of cutting-edge technologies to improve the experience of people with intellectual disabilities. He is mainly interested in Human-Computer Interaction, Accessibility, User Experience, Inclusion, and Education.