Questo sito utilizza solo cookie tecnici per il corretto funzionamento delle pagine web e per il miglioramento dei servizi.
Se vuoi saperne di più o negare il consenso consulta l'informativa sulla privacy.
Proseguendo la navigazione del sito acconsenti all'uso dei cookie.

Prossimi seminari del Dipartimento di Matematica

ore 14:00

presso Seminario II

Speaker's website

fisica matematica

probabilità

Federated Learning (FL) [1] trains a shared model across a population of clients that never expose their local data, replacing centralized empirical risk minimization with distributed optimization over heterogeneous, non-IID sources. The difficulty is mathematical rather than computational: the federation's behavior is governed by the structure of interactions among clients, since conflicting update directions, latent sub-populations, and adversarial or faulty participants emerge from how local data distributions relate. This talk reads these phenomena off the geometry and spectral content of client data. Interactions and clustering. I first introduce FedGWC [3], encoding pairwise client similarity in an interaction matrix built by applying a Gaussian reward to clients' empirical-loss processes. Clustering becomes the analysis of an interacting system, where clients act as particles whose mutual affinities induce homogeneous coalitions. We prove convergence of the Gaussian-weight estimators and introduce the Wasserstein Adjusted Score for cluster cohesion under class imbalance. From detection to clustering. The second part builds on the Wavelet Scattering Transform (WST) [2], a non-expansive, deformation-stable representation summarizing each client by a privacy-preserving spectral embedding computed locally before training. Waffle [4] is a supervised offline detector flagging malicious clients from these embeddings; removing the need for attack labels leads to WASA [5], an unsupervised denoising-autoencoder variant scoring deviations from a learned benign manifold and selecting clients via a Boltzmann-Gibbs sampling rule. Towards zero-shot clustering. I close with ongoing work on zero-shot clustered FL, where cluster assignment is computed offline from a client's spectral signature, with provable guarantees and no additional communication. References: [1] B. McMahan, E. Moore, D. Ramage, S. Hampson, B. Aguera y Arcas. Communication-Efficient Learning of Deep Networks from Decentralized Data. AISTATS, PMLR 54:1273-1282, 2017. [2] S. Mallat. Group Invariant Scattering. Communications on Pure and Applied Mathematics, 65(10):1331-1398, 2012. [3] A. Licciardi, D. Leo, E. Fani, B. Caputo, M. Ciccone. Interaction-Aware Gaussian Weighting for Clustered Federated Learning. Proc. 42nd Int. Conf. on Machine Learning (ICML), PMLR 267:37642-37666, 2025. [4] A. Licciardi, D. Leo, D. Carbone. Wavelet Scattering Transform and Fourier Representation for Offline Detection of Malicious Clients in Federated Learning. IEEE Internet of Things Journal, 2026. doi:10.1109/JIOT.2026.3671698. [5] A. Licciardi. WASA: Wavelet Scattering Autoencoders for Unsupervised Offline Detection of Malicious Clients in Federated Learning. Accepted, IEEE Int. Conf. on Omni-layer Intelligent Systems (COINS), 2026.

Luglio

Venerdì

Maria Angelica Cueto

Tritangent planes to space sextic curves: a tropical viewpoint

nel ciclo di seminari: SEMINARIO DI ALGEBRA E GEOMETRIA

ore 12:00

presso - Aula Da Stabilire -

algebra e geometria

A classical result due to Clebsch from the mid-nineteenth century confirms that every complex space sextic curve (given as an intersection of a quadric and a cubic surface in projective 3-space) has exactly 120 tritangent planes. In this talk we will show how to use combinatorial methods arising from tropical geometry to revisit this classical problem and perform the analogous count over the reals and extensions thereof. This is joint work with Yoav Len, Hannah Markwig and Yue Ren (arXiv:2512.24277 and arXiv:2605.19905).

ore 11:15

presso Aula Seminario VIII piano

sistemi dinamici

Settembre

dal giorno
02/09/2026
al giorno
04/09/2026

Conference

ASK 2026 Conference in Geometric Analysis and PDEs

da mercoledì 02 settembre 2026 a venerdì 04 settembre 2026

The goal of the ASK 2026 - Geometric analysis and PDE conference is to bring together young researchers from various fields of Mathematical Analysis, specifically the geometric analysis of Partial Differential Equations (PDEs), linear and nonlinear PDEs, dispersive PDEs, microlocal analysis, regularity theory of PDE solutions, as well as the calculus of variations. The interdisciplinary nature of the event aims to foster interaction and collaboration among mathematicians specializing in the aforementioned fields, through invited talks by researchers from Italian and international universities, and through contributions from young scholars. Finally, to encourage networking, the schedule includes dedicated time for both formal and informal discussions. For over two years, the ASK group has been organizing periodic seminars and previously hosted another conference in December 2024.

Settembre

Mercoledì