THE NOVI SAD Seminar

 

PROGRAM

---

Predavanja možete pratiti i online putem MITEAM stranice Novosadskog seminara:
https://miteam.mi.sanu.ac.rs/asset/3iuT7dhKfDxFv5kh3

---

Plan rada Novosadskog seminara za OKTOBAR 2025

---

Sreda, 08.10.2025. u 11:00, svečana sala Ogranka SANU u Novom Sadu, Nikole Pašića 6 i Online
Stepan L. Kuznetsov, Steklov Mathematical Institute of RAS; HSE University, Moscow, Russia
EXPRESSIVE POWER AND COMPLEXITY FOR LAMBEK-STYLE CATEGORIAL GRAMMARS
Lambek grammars constitute a formal grammar framework which uses substructural logical reasoning for checking grammatical correctness of sentences. As shown by Pentus (1992), the class of languages generated by Lambek grammars coincides with the class of context-free languages. The Lambek calculus itself is decidable, being NP-complete (Pentus 2006). Linguistic applications, however, suggest various extensions and variations of the Lambek calculus. In this talk, we give a survey of results on their algorithmic complexity and on the classes of languages they generate.

---

Utorak, 14.10.2025. u 11:00, svečana sala Ogranka SANU u Novom Sadu, Nikole Pašića 6 i Online
Bruno Torrésani, Institut de Mathématiques de Marseille (I2M), Aix-Marseille Université, CNRS
GRAPH-BASED WAVELETS AND SPARSE BAYESIAN LEARNING FOR ELECTROMAGNETIC BRAIN SOURCE IMAGING
Electro-encephalography (EEG) and magneto-encephalography (MEG) are non-invasive imaging modalities that provide measurements of time-dependent measurements of brain activity with high temporal accuracy and good spatial accuracy. However, the number of sensors is generally much smaller than the number of unknowns. The problem is modeled as a linear inverse problem, which is then extremely ill conditioned.
Accurate solutions can be obtained when the unknown brain activity can be represented sparsely in some specific domain. This is often done by minimizing a suitable cost function, that involves a non-smooth, parameter-dependent regularization. The parameter tuning is itself a difficult issue.
This work focuses on extended (i.e. non-focal) brain activity, and exploits sparsity in a spatial wavelet domain. Here, spatial wavelets means wavelet systems defined on the surface of the subject's cortex, following a construction by Hammond, Gribonval and Vandergheynst based upon graph Laplacian eigen-decomposition. The graph Laplacian spectrum naturally defines a (radial) frequency domain for functions defined on the graph, which can be used to construct wavelets (in the same way as designing a filter bank in signal processing).
The parameter tuning problem is circumvented using the so-called sparse Bayesian learning (SBL) approaches, which yield an optimization problem to be solved numerically. This eventually provides sparse, diagonal estimates for hyper-parameters of the prior distribution of unknowns.
All together, this results in estimates for brain activity that are sparse in the wavelet domain, but extended in the spatial domain (namely, the cortex surface). The talk will present the wavelet-SBL approach, and numerical results on simulated and real data. It will also briefly discuss alternative constructions for wavelets on the cortex surface.
Joint work with Samy Mokhtari, Jean-Michel Badier and Christian Bénar.

---