Palestras e Seminários

03/08/2026

08:30

Auditório Luiz Antônio Fávaro

Palestrante: Tom Hanika

Responsável: Elias Salomão Helou Neto (elias@icmc.usp.br)

Salvar atividade no Google Calendar

Abstract/Resumo:
Modern machine learning increasingly operates on data that is high-dimensional, structured, and geometrically complex. A central obstacle is the curse of dimensionality: as dimension grows, features concentrate near their means, which impairs discrimination. Hence, learning becomes harder. Yet a basic question remains surprisingly subtle: how can we quantify this complexity in a principled way?

This talk presents a geometric answer, grounded in an axiomatic framework that ties the curse of dimensionality to concentration of measure. We discuss how this approach is situated among established notions of intrinsic dimension, such as manifold-based estimators, and which distinct quality of the data it captures. The framework yields a well-behaved intrinsic dimension function that is actually computationally feasible; it scales to graphs with over one hundred million nodes. Applications to graph neural networks and discriminability-based feature selection illustrate the approach, before we close with a vision extending this perspective from geometry to topology, towards foundations for non-metric data.

Author bio:
Tom Hanika is Professor at the University of Hildesheim, where he co-directs the Information Systems and Machine Learning Lab, and has twice held visiting professorships at Humboldt-Universität zu Berlin. Trained as a mathematician at TU Dresden, he received his doctorate in computer science from the University of Kassel in 2019. His research develops mathematical foundations for trustworthy machine learning, bringing geometry, order theory and topology together with algorithmic methods: intrinsic dimension and high-dimensional concentration phenomena, the extraction of interpretable logical theories from learned models through probably approximately correct (PAC) learning, and topological data analysis for data that carries no natural metric.

CONECT WITH US
 

© 2026 Instituto de Ciências Matemáticas e de Computação