Seminars
inma |
The Departement of Mathematical Engineering organizes a series of seminars. The seminars are usually held on Tuesday from 2:00pm to 3:00pm in the Euler lecture room, Building EULER, av. Georges Lemaître 4-6, Louvain-la-Neuve (Parking 13). Be mindful that exceptions may occur; see the talk annoucements.
If you wish to receive the seminar announcements by email, please send an email to Pascale Premereur.
Master students can take this seminar for credit in either of the two semesters; see LINMA2120 for more information.
Seminars to Come
[INMA] 2026-09-15 (14:00) : LoRA-Muon: Spectral Steepest Descent on the Low-Rank Manifold
At Euler building (room A.002)
Speaker:
Cédric Simal (University of Namur)
Abstract: The Muon optimizer has been making waves as a contender for being the optimizer of choice for training large scale neural networks. Unlike other methods, it explicitly leverages the matrix structure of parameters by performing steepest descent with respect to an operator norm. Over the last year, several works have adapted Muon to various matrix manifold constraints, and we focus in particular on spaces of low-rank matrices, which is of interest for LoRA, a popular fine-tuning method for neural networks. Our method is invariant under reparametrization symmetries, and achieves learning rate transfer across rank, width, depth and factor rescaling.
[INMA] 2026-09-22 (14:00) : FUTON: Fourier Tensor Network for Implicit Neural Representations
At Euler building (room A.002)
Speaker:
Pooya Ashtari (University of Ghent)
Abstract: Implicit neural representations (INRs) encode signals as continuous functions parameterized by neural networks that map coordinates to values, rather than as grids of discrete samples. Since they are resolution-free, compact, and differentiable with respect to their input coordinates, INRs have become attractive priors for inverse problems involving incomplete or irregularly sampled measurements. INRs are typically implemented as multilayer perceptrons (MLPs) with carefully designed activation functions, but such networks can converge slowly, overfit to noise, and extrapolate poorly.
This talk presents FUTON (Fourier Tensor Network), an INR that represents a signal as a generalized Fourier series with a coefficient tensor factorized using a low-rank decomposition. The two ingredients contribute complementary inductive biases: the orthonormal, separable basis favors smoothness and periodicity, while the low-rank factorization captures the low-dimensional spectral structure of natural signals. The resulting model is shallow and multilinear and requires no learned activation functions. I will show that FUTON is a universal approximator in L²; that evaluating it through an appropriate sequence of tensor contractions reduces an otherwise exponential computational cost to a tractable level; and that it outperforms state-of-the-art MLPs in image and volume representation, trains 2–5× faster, and generalizes better to image super-resolution, denoising, and CT/MRI reconstruction.
Previous Seminars
Page 1 of 1
[INMA] 2026-09-01 (14:00) : Exploring the Impact of Memory on Network Controllability
At Euler building (room A.207)
Speaker:
Marco Peruzzo (University of Padova)
Abstract: Many natural and engineered dynamical systems can be modeled as networks consisting of a large number of interconnected, simpler dynamical units. Over the last decade, the challenge of efficiently controlling large-scale networks has driven renewed interest in the control community. Several strategies for guaranteeing network controllability have been developed, such as the optimal selection of driver nodes and the implementation of minimal changes to the network topology.
In this talk, we explore a different strategy for achieving controllability, motivated by networks whose topology cannot be easily modified, such as traffic and water-flow networks. Inspired by lifted Markov chains, we enlarge the local state space of selected network nodes, introducing memory and directionality into their dynamics while leaving the network topology unchanged. We show that this strategy can reduce the worst-case control energy by a factor exponential in the network size.
We further discuss how our strategy can render otherwise uncontrollable networks structurally controllable (i.e., controllable for almost all choices of the network parameters). We characterize networks for which different types of dynamics modifications are required and show that our approach can require significantly fewer subsystem modifications than alternative local dynamics modification strategies.
[INMA] 2026-07-16 (10:00) : Path-conditioned training: a principled way to rescale ReLU neural networks
At Euler building (room A.002)
Speaker:
Titouan Vayer (COMPACT, Inria team Irisa, Rennes, France)
Abstract: Despite recent algorithmic advances, we still lack principled ways to leverage the well-documented rescaling symmetries in ReLU neural network parameters. While two properly rescaled weights implement the same function, the training dynamics can be dramatically different. To offer a fresh perspective on exploiting this phenomenon, we build on the recent path-lifting framework, which provides a compact factorization of ReLU networks. We introduce a geometrically motivated criterion to rescale neural network parameters which minimization leads to a conditioning strategy that aligns a kernel in the path-lifting space with a chosen reference. We derive an efficient algorithm to perform this alignment. In the context of random network initialization, we analyze how the architecture and the initialization scale jointly impact the output of the proposed method. Numerical experiments illustrate its potential to speed up training.
[INMA] 2026-07-02 (10:00) : A two-level inexact smoothing framework for nonsmooth optimization
At Euler building (room A.002)
Speaker:
Masoud Ahookhosh (University of Antwerp)
Abstract: We introduce an inexact two-level optimization framework, ItsOPT, for computing first- and second-order critical points of nonsmooth and nonconvex optimization problems. The framework consists of two interconnected levels. At the upper level, a smoothing technique—such as the high-order Moreau envelope, high-order forward-backward envelope, or high-order tensor envelope—is employed to construct a smooth approximation of the original objective function while preserving its minimizers. First- or second-order optimization methods are then applied to minimize the resulting smooth surrogate. At the lower level, the associated high-order proximal subproblems (e.g., high-order proximal, forward-backward, or tensor subproblems) are solved inexactly using subgradient-based or Bregman proximal methods. The resulting approximate solutions provide inexact evaluations of the smoothing function and its derivative information, which are subsequently used by the upper-level optimization methods. The overall complexity of the proposed framework is given by the product of the computational complexities of the upper- and lower-level procedures. By combining accelerated first- or second-order methods at the upper level with lower-level algorithms whose complexity is negligible (e.g., logarithmic in the desired accuracy), the resulting methods may achieve overall iteration complexities that improve upon existing worst-case complexity bounds. Finally, we present several concrete algorithms within the proposed framework and report preliminary numerical results demonstrating their practical performance.
[INMA] 2026-05-12 (14:00) : Supervisory Control–Driven Scheduling for Complex Manufacturing Systems
At Euler building (room A.002)
Speaker:
Patrícia Pena (Universidade Federal de Minas Gerais, Brazil)
Abstract: This talk presents methods for solving scheduling problems in manufacturing systems using Supervisory Control Theory (SCT) of Discrete Event Systems. The objective is to ensure that task sequences are logically correct and physically feasible by design while optimizing performance. The presentation outlines three strategies: i) reducing model complexity to accelerate the search for optimal solutions; ii) maximizing simultaneous resource activity as a heuristic to find solutions; and iii) using Markov Decision Processes to learn optimal scheduling policies for real-time application. Aspects of implementation in PLCs are discussed, and other related problems of interest are identified.
[INMA] 2026-05-06 (11:00) : Tensor-based Analysis of Hypergraphs and Higher-Order Network Dynamics
At Euler building (room A.002)
Speaker:
Shaoxuan Cui (University of Groningen)
Abstract: In graph-theoretical terms, an edge in a graph connects two vertices, whereas a hyperedge in a hypergraph can connect more than two vertices. From a modeling perspective, a conventional edge denotes a pairwise interaction between two nodes, while a hyperedge may denote a group-wise interaction among several nodes. A hypergraph is said to be uniform if all its hyperedges connect the same number of vertices. In algebraic graph theory, a graph is characterized by an adjacency matrix; correspondingly, a uniform hypergraph can be described by an adjacency tensor. Furthermore, a nonuniform hypergraph can be represented as a set of tensors of different orders. This structural similarity enables the extension of classical matrix analysis techniques, traditionally used for graphs and networked dynamical systems, to hypergraphs and higher-order dynamical systems by leveraging tensor properties. Specifically, we introduce novel notions of tensor irreducibility, corresponding to various forms of strong connectedness in hypergraphs analogous to the graph case. Moreover, we demonstrate that the Perron–Frobenius theorem for nonnegative tensors can be employed to analyze the stability of a class of systems evolving on hypergraphs. This tensor-based framework provides a powerful analytical tool for addressing challenges in network science, complex systems, and control theory.
[INMA] 2026-05-05 (14:00) : Designing Provably Safe Autonomous Systems Under Uncertainty
At Euler building (room A.002)
Speaker:
Sofie Haesaert (Eindhoven University of Technology)
Abstract: The deployment of autonomous aerial and ground vehicles in real-world environments presents fundamental challenges in ensuring safety and resilience under uncertainty. As autonomy increases and human oversight diminishes, we must develop formal methods that provide mathematical guarantees on system behavior in the face of stochastic disturbances and environmental variability.
This talk presents recent advances in data-driven verification and control synthesis for both linear and nonlinear stochastic models. We differentiate between abstraction-based and abstraction-free approaches and show how approximate stochastic simulation relations can be used to quantify abstractions, enabling scalable formal verification of systems with uncertainty. Building on this foundation, we address control synthesis for autonomy via a tight integration of symbolic logic, stochastic formal methods, and uncertainty quantification — paving the way toward interacting autonomous systems that are not only capable, but also provably safe.
[INMA] 2026-04-14 (14:00) : High order deterministic and stochastic optimization without evaluating the objective function
At Euler building (room A.002)
Speaker:
Philippe Toint (Université de Namur)
Abstract: We first motivate OFFO methods, that is methods for optimization without computing the objective function's value. We then show that such methods are applicable for unconstrained minimization of nonconvex functions (both in the deterministic and stochastic frameworks) and give a few examples from deep learning applications. We then move on to the case where the problem has general equality and inequality constraints, propose an OFFO algorithm for this case and analyze its global convergence rate in the deterministic case. We conclude by presenting some numerical illustration of the proposed method. This is a Joint work with S. Gratton, S. Bellavia and B. Morini.
[INMA] 2026-04-07 (14:00) : Pattern-preserving optimal control problems with increasing time-horizon
At Euler building (room A.002)
Speaker:
Matteo Della Rossa (Politecnico of Torino)
Abstract: In this seminar, within the framework of general optimal control theory, we investigate the following question: under which conditions the "structure/pattern" of optimal controls on finite horizons is preserved in the infinite-horizon problem, i.e., as the final time grows unboundedly and tends to infinity? To address this question, we introduce a notion of pattern-preserving family of optimal control problems and show how this property can be characterized via Gamma-convergence of the associated variational formulations. We also discuss important limitations and features of this approach, including counterexamples, scenarios with state constraints, and systems governed by dissipative state equations (in the sense of J.C. Willems). To illustrate the theoretical results, we present several examples and applications, with particular emphasis on a simple epidemic control problem as a real-world case study.
[INMA] 2026-03-31 (14:00) : Envy-free divisions of cakes: recent results and open questions
At Euler building (room A.002)
Speaker:
Frédéric MEUNIER (Paris-tech )
Abstract: The envy-free cake-cutting problem asks for a way to divide a cake (identified with the interval [0,1]) among players with different tastes so that each receives a connected piece and no one envies another’s share. The Stromquist–Woodall theorem from 1980 guarantees the existence of such an envy-free division under mild assumptions. Recently, there has been a surge of interest in this problem from various perspectives — computer science, social choice theory, economics, and topological combinatorics. In this talk, we present several extensions, from these perspectives: the cake may be “poisoned,” there may be multiple cakes with joint preferences, or the cake may be discrete (as in the necklace-splitting problem). We will also discuss several challenging open questions. This talk is based on joint work with Ayumi Igarashi, Francis Su, and Shira Zerbib.
[INMA] 2026-03-24 (14:00) : Alternating low-rank optimization for solving PDEs depending on geometric parameters
At Euler building (room A.002)
Speaker:
Javier Bevia Ripoll (KU Leuven)
Abstract: We seek an efficient computation of the solutions to a PDE posed on a domain dependent on geometric parameters, as encountered in, for example, multiscale topology optimization. The domain geometry adds an extra complexity to the problem, as the parametrization is implicitly present in the domain shape, but not explicitly in the PDE coefficients. We address this by introducing a smooth change of variables that maps each parameterized domain to a fixed reference domain, yielding a PDE with analytically parameterized coefficients. The analytic dependence guarantees that the matrix containing the discretization of the solutions across parameter values has exponentially decaying singular values, so the family of solutions admits a low‑rank representation. We compute this representation with an alternating least squares (ALS) scheme and present numerical experiments for a PDE depending on one and two geometric parameters to illustrate the effectiveness of the approach.
[INMA] 2026-03-17 (14:00) : Active surface haptics for rich touch interaction
At Euler building (room A.002)
Speaker:
Zhaochong Cai (ICTEAM, UCLouvain )
Abstract: Touch is fundamental to our perception of the world and to our interaction with the physical environment. With touch, we can intuitively and effortlessly manipulate objects and control complex machines. However, most touch-based interfaces deliver only primitive tactile feedback, such as vibrations, which are a poor substitute for the richness of natural touch. This talk presents my doctoral work on the design and evaluation of active surface-haptic devices that deliver lateral force feedback to the bare fingertip. By generating controllable lateral forces using resonant traveling waves, these devices can render force fields and guide users toward targets, enabling the perception of virtual shapes and improving targeting performance. I will then briefly introduce my current postdoctoral project, which aims to connect fingertip skin deformation, mechanical modeling, and tactile afferent responses. In particular, the goal is to use imaging and microneurography to better understand how the local strain patterns at the skin surface during manipulation are transformed into neural signals.
[INMA] 2026-03-10 (14:00) : From Learning to Optimize to Learning Optimization Algorithms
At Euler building (room A.002)
Speaker:
Camille Castera (University of Bordeaux)
Abstract: Towards designing learned optimization algorithms that are usable beyond their training setting, we identify key principles that classical algorithms obey, but have up to now, not been used for Learning to Optimize (L2O). Following these principles, we provide a general design pipeline, taking into account data, architecture and learning strategy, and thereby enabling a synergy between classical optimization and L2O, resulting in a philosophy of Learning Optimization Algorithms. As a consequence our learned algorithms perform well far beyond problems from the training distribution. We demonstrate the success of these novel principles by designing a new learning-enhanced BFGS algorithm and provide numerical experiments evidencing its adaptation to different settings at test time.
[INMA] 2026-03-03 (14:00) : Consensus is a myth: Human label variation in Natural Language Inference
At Euler building (room A.002)
Speaker:
Marie-Catherine de Marneffe (CENTAL, UCLouvain)
Abstract: Recently, NLP researchers have increasingly begun to acknowledge that humans often diverge in their interpretations of various NLP tasks, and that such variation should be captured if robust language understanding is to be achieved. In this talk, I will focus on analyzing human label variation in the Natural Language Inference (NLI) task, in which, given a premise, one identifies whether a hypothesis sentence is true, false, or undetermined. For instance, if one says: “My friend often travels with a heavy suitcase”, can it be inferred that “My friend often travels with a light suitcase”? I will examine the various sources of NLI label variation and investigate whether or not they can be captured by current LLMs, arguing that, in the presence of variation, labels without explanations are not sufficiently meaningful.
[INMA] 2026-02-24 (14:00) : A passivity-based perspective on distributed optimization and its acceleration
At Euler building (room A.002)
Speaker:
Ivano Notarnicola (University of Bologna)
Abstract: This talk revisits the classical gradient method for unconstrained optimization through the lens of control theory. By explicitly interpreting the gradient method as a feedback interconnection between a linear dynamical system and a static nonlinearity associated to the cost function gradient, we uncover an underlying control-theoretic structure. Within this framework, linear convergence is established using arguments from passivity theory. This system-theoretic viewpoint further reveals that the gradient method can also be interpreted as a feedback-controlled static nonlinearity, opening the door to an unconventional interpretation of accelerated schemes. Finally, the same passivity-based tools are also applied to consensus optimization problems, yielding a unified framework for the design and analysis of distributed gradient methods and their acceleration.
[INMA] 2026-02-17 (14:00) : Newcomers Seminar
At Euler building (room A.002)
Section 1: Fleet Rebalancing: Scalable Feedback Optimization for Autonomous Mobility-on-Demand via null-space projection
Speaker:
Arthur Mélot (UCLouvain)
Abstract: While state-of-the-art Model Predictive Control (MPC) approaches for Autonomous Mobility-on-Demand (AMoD) achieve scalability through flow-based modeling , they remain computationally burdened by iterative solvers , the need for a pre-computed static equilibrium to track and also a non-negligible prediction horizon to ensure stability. This presentation introduces a horizon-free feedback optimization algorithm that also eliminates the need for offline equilibrium computation. By treating the economic cost minimization as a dynamic feedback process, our approach steers the fleet directly toward the optimal operating point in real-time and does not track a pre-computed reference signal. The preliminary results provide a negligible execution time, outperforming the execution time of standard MPC. This project is still in progress and all the results provided are preliminary results.
Section 2: Mechanical determinants of tactile perception in the human fingertip
Speaker:
Viktoriia Kozadaeva (UCLouvain)
Abstract: Human tactile perception enables discrimination of features with microscale precision. This remarkable sensitivity arises from dense innervation of mechanoreceptors within the fingertip that transduce small skin deformations into neural signals. While neural innervation of the fingertip has been widely studied, the mechanical mechanisms enabling the microscale sensations remain underexplored. My research project aims to establish a relationship between psychophysical performance and mechanical determinants underlying microscale tactile perception. Specifically, we examine how fingerpad deformation relates to perceptual detection of small surface features.
Section 3: Reinforcement Learning for Petri-Net based Discrete Event Systems: Application to Aircraft Maintenance Scheduling
Speaker:
Margot Devillers (UCLouvain)
Abstract: Efficient aircraft maintenance scheduling is critical to maximizing fleet availability while ensuring regulatory compliance and cost efficiency. The problem involves grouping periodic tasks into maintenance projects and deciding when to ground aircraft under strict deadline and resource constraints. The combinatorial nature of these decisions makes classical optimization approaches challenging at fleet scale.
Building on a Colored Petri Net (CPN) digital twin of the aircraft maintenance operations, this work formulates the scheduling process as a discrete-event dynamical system and investigates the use of Reinforcement Learning (RL) to address it. An RL agent is currently being developed to interact with the CPN environment and learn fleet-level policies that balance aircraft utilization and operational feasibility over a finite planning horizon.
[INMA] 2026-02-10 (14:00) : The Ellipsoidal Separation Machine
At Euler building (room A.002)
Speaker:
Antonio Frangioni (Università di Pisa)
Abstract: We propose the -- to the best of our knowledge -- first fully functional implementation of the "Separation by a Convex Body" approach first outlined in [Grzybowski et al., Optimization Methods and Software, 2005] for classification, separating two data sets using an ellipsoid. A training problem is defined that is structurally similar to the Support Vector Machine (SVM) one, thus leading to call our method the Ellipsoidal Separation Machine (ESM). Like SVM, the training problem is convex, and can in particular be formulated -- via a set of not entirely obvious reformulation tricks -- as a Semidefinite Program (SDP). However, practical classification tasks produce rather large SDPs, solving which by means of standard SDP approaches (be them IP-or first-order based) does not scale. As an alternative, a nonconvex formulation is proposed that is amenable to a Block-Gauss-Seidel approach alternating between a much smaller SDP and a simple separable Second-Order Cone Program. For the purpose of the classification approach the reduced SDP can even be solved approximately by relaxing it in a Lagrangian way and updating the multipliers by fast subgradient-type approaches. A characteristic of ESM is that it necessarily defines "indeterminate points", i.e., those that cannot be reliably classified as belonging to either one of the two sets. This makes it particularly suitable for Classification with Rejection (CwR) tasks, whereby the system explicitly indicates that classification of some points as belonging to one of the two sets is too doubtful to be reliable. We show that, in many datasets, ESM is competitive with SVM -- with the kernel chosen among the three standard ones and endowed with CwR capabilities using the margin of the classifier -- and in general behaves differently. Thus, ESM provides another arrow in the quiver when designing CwR approaches, although more work would be needed to scale it to really large datasets.
[INMA] 2026-02-03 (14:00) : IRKA Is a Riemannian Gradient Descent Method
At Euler building (room A.002)
Speaker:
Petar Mlinarić (University of Zagreb)
Abstract: Large-scale systems frequently arise in applications involving partial differential equations or network dynamics. Model order reduction seeks to replace a large-scale system with a reduced-order model, enabling faster simulations with minimal loss of accuracy. The Iterative Rational Krylov Algorithm (IRKA) is a well-known method for model order reduction of linear time-invariant systems, originally formulated as a fixed-point iteration. In this talk, we show that IRKA can be interpreted as a Riemannian gradient descent method with a fixed step size on the manifold of rational functions of fixed degree. This geometric perspective motivates the application of other Riemannian optimization techniques to the same problem. We illustrate the effectiveness of these approaches through numerical examples.
[INMA] 2026-01-20 (14:00) : Opinion Dynamics with Nonlinear Interaction: From Robust Clustering to Environmental Coupling
At Euler building (room A.002)
Speaker:
Anthony Couthures (Université de Lorraine)
Abstract: Social opinion and environmental states are deeply coupled: collective human behavior impacts the environment, while the state of the environment feeds back into public opinion. In this talk, I will present a mathematical framework to analyze these interactions, moving from standard consensus models to coupled socio-environmental dynamics.
First, I will introduce a generalized framework for multi-agent opinion dynamics with nonlinear interactions. Unlike classical linear consensus models, nonlinear communication allows for the emergence of rich behaviors beyond simple agreement. We will establish a sharp threshold linking network connectivity (algebraic connectivity) and interaction nonlinearity (Lipschitz constant) that dictates the transition from global synchronization to persistent polarization. Using Input-to-State Stability (ISS) theory, I will also discuss the robustness of these polarized clusters against external influence.
In the second part, I will couple this opinion model with an environmental resource variable. By analyzing the system on the synchronization manifold, we identify the role of the "attention parameter" β: representing the weight agents place on environmental feedback versus social influence. Through bifurcation analysis, I will demonstrate how varying this parameter triggers fundamental qualitative changes, specifically Pitchfork bifurcations (leading to bistability and polarization) and Hopf bifurcations (leading to recurrent cycles of environmental collapse and recovery).
[INMA] 2025-12-16 (14:00) : Geometry of low-rank tensors
At EULER (room A.002)
Speaker:
Simon Jacobsson (KU Leuven)
Abstract: Morally, a manifold is a set where notions from calculus are well-defined. For example, the set of n-by-n orthogonal matrices and the set of n-by-n symmetric positive definite matrices are manifolds. Knowing that a set of matrices or tensors is a manifold allows us to use a host of calculus tools to do numerical analysis on that set. For example, many constrained optimization problems can be formulated as unconstrained manifold optimization problems. Algorithms for these can then make use of manifold gradients and analogues of straight lines called geodesics.
We consider the set of fixed-rank tensors. When the rank is sufficiently low, then (almost) any tensor in this set is related to (almost) any other tensor via a change of basis. We explain how this relation induces a manifold structure, and show how relevant quantities like gradients and geodesics can be computed efficiently. More precisely, we identify the set as a quotient of Lie groups. We also discuss how the manifold perspective can be used to integrate tensor differential equations.
[INMA] 2025-12-09 (14:00) : Matroids are equitable
At EULER (room A.002)
Speaker:
László Végh (University of Bonn)
Abstract: We show that if the ground set of a matroid can be partitioned into k≥2 bases, then for any given subset S of the ground set, there is a partition into k bases such that the sizes of the intersections of the bases with S may differ by at most one, settling a conjecture by Fekete and Szabó from 2011.
In the talk, I will present the surprisingly simple proof, as well as some extensions and applicaitons in fair division. I will also give an overview of related questions on matroid basis exchanges. This is based on joint work with Hannaneh Akrami, Roshan Raj, and Siyue Liu.
[INMA] 2025-12-02 (14:00) : Expanding BGP Data Horizons
At EULER (room A.002)
Speaker:
Cristel Pelsser (INGI/ICTEAM/UCL)
Abstract: BGP data collection platforms as currently architected face fundamental challenges that threaten their long-term sustainability: their data comes with enormous redundancy and yet dangerous visibility gaps. GILL is a new BGP routes collection platform that can collect routes from at least an order of magnitude more routers compared to existing platforms while limiting the increase in human effort and data volume. GILL’s key principle is an overshoot-and-discard collection scheme: Any AS can easily peer with GILL and export their routes. GILL offers a lossy compression algorithm that only stores the nonredundant routes as well as lossless compression leveraging redundancy in BGP attributes, in our new bgproutes.io platform. Our new mode of data selection and delivery enables to improve BGP data analysis such as topology mapping, AS ranking, and forged origin hijack detection.
We have built such a detector. DFOH is a system designed to detect forged-origin hijacks across the entire Internet. Forged-origin hijacks are typically malicious BGP hijacks where attackers manipulate the AS path of BGP messages to make them appear as legitimate routing updates. DFOH is particularly useful because the proposed BGP extensions for cryptographically verifying the validity of AS paths (e.g., BGPSec or ASPA) are challenging to deploy widely. With DFOH, operators can quickly and confidently determine when their traffic is being hijacked.
[INMA] 2025-11-25 (14:00) : A constrained Lie group approach to the modeling of dynamic mechanical systems
At EULER (room A.002)
Speaker:
Olivier Bruls (University of Liège)
Abstract: This talk addresses general-purpose geometric modeling methods for a wide class of mechanical systems which includes robotic systems, biomechanical systems, deployable space structures, automotive vehicles, or industrial machines. These systems are represented as a set of rigid and flexible bodies whose dynamics is restricted due to the presence of kinematic joints and contact conditions.
It is well-known that the motion of an isolated rigid body can be conveniently represented on the special Euclidean group SE(3). In the first part of the talk, I will show that this SE(3) representation can be extended to model deformable structures, such as rods, shells or more complex 3D flexible bodies. The Lie group framework can then be exploited for the construction of geometrically-consistent spatial discretization schemes and offers to the possibility to write the equations of motion in local frames (and not in an inertial frame).
In the second part of the talk, I will address the treatment of bilateral constraints, which model kinematic joints, leading to a formulation of the equations of motion as a differential-algebraic equation (DAE) on a Lie group. Geometric time discretization methods for such DAE will then be discussed. Notice that unilateral constraints, which model contact conditions, can also be considered by adapting the formulation of measure differential inclusions and nonsmooth time integration schemes to the Lie group settings. Finally, a few numerical examples will be presented to illustrate the generality of the proposed framework.
[INMA] 2025-11-18 (14:20) : Data-Driven Methods for Formal Verification and Synthesis of Dynamical Systems
At Euler building (room A.002)
Speaker:
Sadegh Soudjani (Max Planck Institute)
Abstract: Ensuring the safe and reliable behavior of dynamical systems under uncertainty is a fundamental challenge in control and verification. In this talk, I will discuss recent advances in data-driven and certificate-based approaches for the formal verification and synthesis of such systems. I will first present results on necessary and sufficient certificates for reachability in stochastic systems, which provide exact characterizations of when a target set can be reached with probability one. With this, we have closed the long-standing question of characterizing certificates that are both necessary and sufficient. I will then show that the common practice of fixing a template for computing such certificates results in losing completeness: there are polynomial systems that do not admit polynomial certificates. Building on this foundation, I will discuss one of the results from our EIC SymAware project on data-driven approaches for distributionally robust control in multi-agent systems subject to logical and temporal constraints. By leveraging samples from uncertain environments, these methods enable the synthesis of controllers that are robust to distributional uncertainty while satisfying high-level behavioral specifications utilizing knowledge of the behavior of other agents in the system. Credit for the works being discussed in the talk also goes to my collaborators and hard-working students.
[INMA] 2025-11-17 (14:00) : Regularized block coordinate descent methods: Complexity and applications
At a.002
Speaker:
Ernesto Birgin (Universidade de São Paulo, Brazil)
Abstract: In this work, we propose block coordinate descent methods for bound-constrained and nonconvex constrained minimization problems. Our approach relies on solving regularized models. For bound-constrained problems, we introduce methods based on models of order $p$, which exhibit asymptotic convergence to $p$th-order stationary points. Moreover, first-order stationarity with precision $\epsilon$ with respect to the variables of each block is achieved in $O(\epsilon^{-(p+1)/p})$; while first-order stationarity with precision $\epsilon$ with respect to all the variables is achieved in $O(\epsilon^{-(p+1)})$. For nonconvex constrained minimization, we consider models with quadratic regularization. Given feasibility/complementarity and optimality tolerances $\delta>0$ and $\epsilon>0$ for feasibility/complementarity and optimality, respectively, it is shown that a measure of $(\delta,0)$-criticality tends to zero; and the number of iterations and functional evaluations required to achieve $(\delta,\epsilon)$-criticality is $O(\epsilon^{-2})$. Numerical experiments illustrate the effectiveness of our methods. We apply the first method to solve the Molecular Distance Geometry Problem, while the second method is used to enhance heuristic approaches for the Traveling Salesman Problem (TSP) with neighbors, a variant of the classical TSP problem where regions in the plane must be visited instead of cities. The case where regions are described by arbitrary (nonconvex) polygons is considered.
[INMA] 2025-11-04 (14:00) : Automated algorithm analysis for time-varying optimization: tracking and regret bounds
At Euler building (room A.002)
Speaker:
Fabian Jakob (University of Stuttgart)
Abstract: Time-varying optimization problems arise across many disciplines from engineering to online learning. The development of efficient algorithms can be quite impactful, as application domains include e.g. power grid systems, mobile robotics, and portfolio optimization. The performance of such algorithms is typically assessed in two ways: (1) how well they track time-varying minima, and (2) how large their accumulated suboptimality is when the objective functions are not known in advance, also known as regret. However, deriving tracking or regret guarantees is often tedious and highly problem-specific, requiring involved and ad-hoc analyses.
This talk addresses this issue. We present a novel framework for computer-aided analysis of first-order optimization algorithms for strongly convex and smooth objectives. The framework builds on casting first-order algorithms as dynamical systems and using Integral Quadratic Constraints (IQCs) for their analysis. We recap the concept of IQCs and present an extension to the time-varying setting, which allows us to model temporal variations as disturbances acting on the algorithm dynamics. Based on this, we show how tracking and regret certificates of an algorithm can be obtained as the solution of a semidefinite program and demonstrate numerically how the choice of algorithm affects the performance and sensitivity to time-variations.
[INMA] 2025-10-29 (14:00) : Internal Optimization Seminar
At Euler building (room A.002)
Speaker:
No author specified
Abstract: This is a weekly seminar which explores cutting-edge research and applications in mathematical optimization, spanning theory, algorithms, and real-world problem-solving. Each session features talks from leading researchers, practitioners, or graduate students, covering various topics. Attendees include mathematicians, engineers, and data scientists — all united by a passion for optimization. Open to all. No registration required.
[INMA] 2025-10-21 (14:00) : Over-Approximation Methods for Safe Control: Lifting and Preview Information
At Euler building (room A.002)
Speaker:
Aspeel, Antoine
Abstract:
Ensuring safety in the control of nonlinear systems is a central challenge in control theory. Over-approximations address this by replacing a deterministic nonlinear system with a nondeterministic (piecewise) linear one, enabling the use of control synthesis techniques with formal guarantees for the original dynamics.
The presentation will cover two recent approaches that reduce conservatism in over-approximations. Lifted over-approximations represent the system in higher dimension, providing additional degrees of freedom. Over-approximations with preview reinterpret the approximation error as input-dependent preview information, leading to policies that depend jointly on the state and the error. The resulting concretization problem—recovering a valid input for the true system from such a policy—is formulated as a fixed-point equation, enabling efficient computation.
[INMA] 2025-10-16 (13:00) : Malware Detection with Machine Learning: Challenges and Perspectives
At Shannon room, Maxwell building
Speaker:
Serena Lucca (ICTEAM)
,
and
Samy Bettaied (ICTEAM)
Abstract: This presentation explores some challenges and perspectives of applying machine learning to malware detection, drawing insights from two complementary studies. The first investigates the surprising effectiveness of simple models—such as One-Rule and AdaBoost—on state-of-the-art malware detection datasets, revealing that a small subset of dominant features often drives classification performance, leading to minimal differences between simple and deep learning approaches. The second study provides a systematic comparison of tabular and graph-based feature representations under unified conditions, evaluating their trade-offs in computational cost, detection accuracy, and robustness to adversarial attacks. Together, these works question the common assumption that complex models or sophisticated feature types always yield superior results, and instead highlight the importance of understanding feature dominance, dataset biases, and practical constraints in real-world malware detection.
[INMA] 2025-10-15 (14:00) : Internal Optimization Seminar
At Euler building (room A.002)
Speaker:
No author specified
Abstract: This is a weekly seminar which explores cutting-edge research and applications in mathematical optimization, spanning theory, algorithms, and real-world problem-solving. Each session features talks from leading researchers, practitioners, or graduate students, covering various topics. Attendees include mathematicians, engineers, and data scientists — all united by a passion for optimization. Open to all. No registration required.
[INMA] 2025-10-14 (14:00) : Newcomers seminars (PhDs)
At Euler building (room a.002)
Section 1: Low-rank PSD matrix completion
Speaker:
Sophie Lequeu (PhD UCLouvain/INMA)
Abstract: While convexity is traditionally seen as essential for solving optimization problems, many nonconvex ones pose no significant issues in practice. Moreover, simpler nonconvex formulations are generally more compact and amenable to parallel solving, even when an equivalent convex formulation exists. This justifies the interest in studying the global landscape of selected optimization problems, in order to determine the characteristics that distinguish problems with benign nonconvexity from those with non-benign nonconvexity.
In this research project, we will study the Burer–Monteiro nonconvex reformulation of the low-rank PSD matrix completion problem. For sparsity patterns corresponding to chordal graphs, a classical result guarantees the existence of a low-rank solution, encouraging the use of this reformulation. The goal is to determine conditions under which this formulation has no spurious local minima, by studying characteristics of the sparsity pattern.
Section 2: Novel deep learning architectures for the detection of the stochastic gravitational wave background.
Speaker:
Antonin Oswald (PhD UCLouvain/INMA)
Abstract: We propose to advance the understanding of the stochastic gravitational wave background by proposing new machine learning architectures specifically dedicated to processing correlation matrices signal representations. These architectures will heavily rely on the geometry of the manifold of positive-definite matrices, to which the correlation matrices representing classically the signal rely. We will explore several directions, aiming to account for time- and frequency dependency in the correlation representations. We will then use our proposed novel architectures for denoising correlation matrices in view of subsequent SGWB detection.
Section 3: Path-complete reinforcement learning
Speaker:
Lea Ninite (PhD UCLouvain/INMA)
Abstract: Reinforcement Learning (RL) has achieved remarkable success in complex control tasks, yet its lack of theoretical guarantees limits its use in safety-critical systems. In RL, the Q-function satisfies a decrease condition similar to that of a Lyapunov function, but this relation is typically enforced through heuristic, data-driven updates, which hinders robustness and interpretability. In contrast, Path-Complete Lyapunov Functions (PCLFs) offer a systematic and combinatorial framework for encoding stability through sets of local decrease conditions on a graph structure.
This PhD project aims to bridge these two paradigms by developing a Path-Complete Reinforcement Learning (PCRL) framework, introducing a path-complete relaxation of the Bellman equation. As a first step, we focus on computing an upper bound of the value function for arbitrarily switched linear systems using path-complete graphs, where each node encodes a quadratic function satisfying Bellman-like inequalities. Preliminary results show that this construction can yield tight upper bounds on the true value function, highlighting the potential of path-complete methods to bring theoretical structure to learning-based control.
[INMA] 2025-10-08 (14:00) : Internal Optimization Seminar
At Euler building (room A.002)
Speaker:
No author specified
Abstract:
This is a weekly seminar which explores cutting-edge research and applications in mathematical optimization, spanning theory, algorithms, and real-world problem-solving. Each session features talks from leading researchers, practitioners, or graduate students, covering various topics. Attendees include mathematicians, engineers, and data scientists — all united by a passion for optimization. Open to all. No registration required.
[INMA] 2025-10-07 (14:00) : Communication-efficient distributed optimization algorithms
At Euler building (room A.002)
Speaker:
Laurent Condat (King Abdullah University of Science and Technology (KAUST))
Abstract:
In distributed optimization and machine learning, a large number of machines perform computations in parallel and communicate back and forth with a server. In particular, in federated learning, the distributed training process is run on personal devices such as mobile phones. In this context, communication, that can be slow, costly and unreliable, forms the main bottleneck. To reduce it, two strategies are popular: 1) local training, which consists in communicating less frequently; 2) compression. Also, a robust algorithm should allow for partial participation. I will present several randomized algorithms we developed recently, with proved convergence guarantees and accelerated complexity. Our most recent paper “LoCoDL: Communication-Efficient Distributed Learning with Local Training and Compression” has been presented as a Spotlight at the ICLR conference in April 2025.
[INMA] 2025-09-30 (14:00) : Toward Resilient Operation of Large-Scale Cyber-Physical Human Systems
At Euler building (room A.002)
Speaker:
Ahmad Al-Dabbagh (University of British Columbia)
Abstract:
Examples of large-scale cyber-physical human systems are many in society, such as in manufacturing and energy applications. The systems rely on a high degree of coupling between their cyber, physical, and human components, where operational information is exchanged between the components using communication networks. The involved coupling and the communication networks introduce vulnerabilities which jeopardize the reliability and security of the systems. This seminar provides an overview of decision-making methods for control and monitoring of large-scale cyber-physical human systems, using control theory and artificial intelligence, while focusing on practical challenges related to cybersecurity, fault diagnosis, and alarm management.
[INMA] 2025-09-23 (14:00) : Newcomers seminars (PhDs)
At Euler building (room a.002)
Section 1: Random Embeddings for Deep Learning: Improving Scalability and Generalization
Speaker:
Roy Makhlouf (PhD UCLouvain/INMA)
Abstract: Increasingly powerful processing units have led to a dramatic surge in the number of parameters of deep neural networks (DNNs), for which training comes with heavy computational costs. As a consequence, it is essential to develop more scalable algorithms for training DNNs.
This work aims to explore the possible benefits of low-dimensional embeddings as a dimensionality reduction technique for DNN training. Instead of considering the entire parameter space, the idea is to restrict training to a lower-dimensional subspace, thereby significantly reducing computational cost. This approach is motivated by prior numerical results, which suggest that training overparameterized neural networks within a subspace of very small dimension still allows to achieve a high test accuracy.
Building on my Master's thesis results, we will first conduct a deeper investigation of random Gaussian embeddings for DNN training under the assumption that the training loss exhibits anisotropic variability. That is, when it varies very slowly along some directions and possibly much faster along others. This setting often occurs in overparameterized neural networks training, where not all parameters influence the training loss equally. We will then consider more structured embeddings known as sparse embeddings, which are closer to techniques already used in deep learning. Finally, we will look at how random embeddings can help avoid sharp spurious minima, a class of minima expected to harm model generalization.
Section 2: Analysis of Hidden Convexity in Neural Networks and Transformers: Toward More Efficient and Robust Deep Learning.
Speaker:
Adeline Colson (PhD UCLouvain/INMA)
Abstract: While nonconvexity is traditionally viewed as a challenge in optimization, many machine learning models exhibit a phenomenon known as benign nonconvexity, where nonconvex formulations are surprisingly tractable and often more scalable than their convex counterparts. The research project aims to understand and leverage this phenomenon to identify models that are both expressive and efficient to train, develop strategies to escape spurious local minima, and propose new formulations with benign nonconvexity across domains. Hidden convexity refers to a convex structure that is not apparent in the original nonconvex problem. By reformulating the problem using local optimality conditions (first and second order), one can analyze it via a convex program. This allows global properties, like optimality of local minima, to be inferred from local conditions.
Section 3: Scalable Control Design for Networked Systems: Coordination through Local Cooperation.
Speaker:
Jonas Hansson (Lund University,Sweden)
Abstract: In this talk, I will present a compositional framework for consensus and coordination, with applications to vehicular formations. The approach, called serial consensus, constructs high-order protocols by cascading simpler first-order dynamics, which makes stability transparent and enables scalable performance guarantees such as string stability. I will also discuss extensions to nonlinear settings, where the framework accommodates constraints such as saturation and time-varying topologies. Altogether, the results show how distributed controllers based only on local relative measurements can ensure robust and scalable coordination in large-scale networks.
[INMA] 2025-09-16 (14:00) : Safety in the Face of Uncertainty: When is a Scenario Decision-Making Algorithm Safe?
At Euler building (room A.002)
Speaker:
Guillaume Berger (UCLouvain)
Abstract:
Making risk-aware decisions in the face of uncertainty is a central problem in many applications of engineering such as autonomous transportation, energy planning, medical devices, etc. Indeed, in these applications, failures or errors come at a high cost, so that it is important to bound the probability of such events. Nevertheless, this problem is often very challenging in practice because the probability distribution of the uncertainty is often unknown to the decision maker, which must thus make decisions in a black-box way. Scenario decision-making is a powerful data-driven approach to risk-aware decision-making, consisting in drawing samples (called scenarios) of the uncertainty and making a decision based on these samples. A key question is when such scenario-based decisions have a low risk. In this talk, I will review the main techniques from the literature for providing such bounds on the risk, and will show that they are incomparable, in that none is more general (i.e., non-vacuous on a larger class of problems) or less conservative than the other. I will then present a more general bound, inspired by the connection between scenario decision-making algorithms, set operators and VC theory. Finally, I will demonstrate the usefulness of the new bound on problems from scenario optimization.
Page 1 of 1
Please visit seminars archive for the whole list.