2026's Seminars
Page 1 of 11
Next
[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.
[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] 2026-09-01 (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.
[INGI] 2026-07-02 (13:00) : 2 INGI seminars in one (J. Vanderhaeghen & A. Fierens)
At Shannon (Maxwell - a.105)
Speaker :
Amaury Fierens (ICTEAM)
,
and
Juliette Vanderhaeghen (ICTEAM)
Abstract : By Juliette Vanderhaeghen:
Towards Personalized Tumor Growth Modeling with Physics-Informed Neural Networks
Abstract:
Cancer remains a major global health challenge, highlighting the urgent need for more personalized treatment strategies. Mathematical models, in particular PDE-based approaches, provide a powerful framework to describe tumor dynamics. However, classical numerical methods can be computationally expensive and require careful calibration, while purely data-driven models often lack interpretability and demand large amounts of data.
Physics-Informed Neural Networks (PINNs) offer a promising alternative by embedding physical laws directly into the training process, enabling both efficient simulations and the inference of patient-specific parameters from limited data.
In this presentation, I present our current work on the use of PINNs to model tumor growth dynamics. First, I will demonstrate how PINNs can reconstruct tumor evolution and estimate key parameters from partial synthetic observations. Then I will present their application to multicellular tumor spheroids as a controlled experimental system. These spheroids generate spatially and temporally resolved data under varying treatment conditions, making them a valuable testbed for model validation.
By Amaury Fierens:
Self-Alignment and Reranking for Large-Scale Medical Coding in Open-Source Clinical NLP
Abstract:
Medical coding is a challenging large-scale linking problem: clinical mentions must be mapped to highly ambiguous terminologies containing millions of possible codes, such as UMLS. While large language models are increasingly used in clinical NLP, they are not yet a fully satisfactory solution for this task, especially when reliability, traceability, terminology coverage, and precise code selection are required.
The presentation is structured in three parts. First, I will introduce the complexity of medical coding and why it remains difficult for both automated systems and clinicians. Second, I qill discuss an effective technical approach based on self-alignment pretraining with contrastive learning, as in SapBERT, combined with candidate reranking through cross-encoders, rule-based signals, and hierarchical information. Finally, drawing on the open-source work carried out this year by two master’s students, I will discuss how effective NLP components can be integrated into practical tools that support medical annotation and coding workflows.
Towards Personalized Tumor Growth Modeling with Physics-Informed Neural Networks
Abstract:
Cancer remains a major global health challenge, highlighting the urgent need for more personalized treatment strategies. Mathematical models, in particular PDE-based approaches, provide a powerful framework to describe tumor dynamics. However, classical numerical methods can be computationally expensive and require careful calibration, while purely data-driven models often lack interpretability and demand large amounts of data.
Physics-Informed Neural Networks (PINNs) offer a promising alternative by embedding physical laws directly into the training process, enabling both efficient simulations and the inference of patient-specific parameters from limited data.
In this presentation, I present our current work on the use of PINNs to model tumor growth dynamics. First, I will demonstrate how PINNs can reconstruct tumor evolution and estimate key parameters from partial synthetic observations. Then I will present their application to multicellular tumor spheroids as a controlled experimental system. These spheroids generate spatially and temporally resolved data under varying treatment conditions, making them a valuable testbed for model validation.
By Amaury Fierens:
Self-Alignment and Reranking for Large-Scale Medical Coding in Open-Source Clinical NLP
Abstract:
Medical coding is a challenging large-scale linking problem: clinical mentions must be mapped to highly ambiguous terminologies containing millions of possible codes, such as UMLS. While large language models are increasingly used in clinical NLP, they are not yet a fully satisfactory solution for this task, especially when reliability, traceability, terminology coverage, and precise code selection are required.
The presentation is structured in three parts. First, I will introduce the complexity of medical coding and why it remains difficult for both automated systems and clinicians. Second, I qill discuss an effective technical approach based on self-alignment pretraining with contrastive learning, as in SapBERT, combined with candidate reranking through cross-encoders, rule-based signals, and hierarchical information. Finally, drawing on the open-source work carried out this year by two master’s students, I will discuss how effective NLP components can be integrated into practical tools that support medical annotation and coding workflows.
Page 1 of 11
Next
Seminars
Subscribe (ICS)
INMA Contact Info
Mathematical Engineering (INMA)
L4.05.01
Avenue Georges Lemaître, 4
+32 10 47 80 36
secretaire-inma@uclouvain.be
Mon – Fri 9:00A.M. – 5:00P.M.
JOIN US