Seminar on October 20: "Geometry as a Guide: Building AI That Understands Shape"
by Roveda Loris
Dear colleagues,
just a reminder for the planned seminar on October 20 by Prof. Simone
Melzi from UniMiB, with title "Geometry as a Guide: Building AI That
Understands Shape".
The seminar will be in room C2.09 at 11am.
Please find below additional details about the seminar.
See you there,
Loris
--
*Abstract:*
Shape matching is a core task in many fields such as computer graphics,
robotics, medical imaging, and geometric learning. Accurately
establishing correspondences between 3D shapes enables applications
ranging from surgical planning to robotic grasping and manipulation. In
this talk, I present a data-driven perspective where understanding
geometry and learning shape correspondences become mutually reinforcing
goals. By targeting tasks such as shape matching, deformation, and
representation refinement, we show that machine learning captures
geometric features, while at the same time, geometry itself guides
learning towards robust and interpretable models.
We explicitly exploit geometric structures within data-driven pipelines,
showing that localized deformation cues can be learned without relying
on global shape encodings. This integration of geometric priors into
neural architectures leads to more efficient and interpretable systems
suited for real-world applications at scale.
Going further, we explore the emergent structure within learned
representations and attention patterns in Transformer-based models for
point cloud matching. These insights enable us to reduce training
overhead and improve robustness by integrating geometry-aware mechanisms
directly into the model design. Our findings pave the way for future
research that unifies geometric priors and deep learning to build AI
systems that not only learn from shape but understand it, enabling
smarter tools in design, medicine, and robotics.
*Short Bio: *
Simone Melzi is an Associate Professor at the University of
Milano-Bicocca, in the Department of Informatics, Systems and
Communication (DISCo). He was a Post Doctoral researcher in Computer
Science at the Sapienza University of Rome in the GLADIA group led by
Emanuele Rodolà, at the École Polytechnique in the team of Prof. Maks
Ovsjanikov (6 months) and at the Università Degli Studi di Verona
(2018-2019). He received his PhD in Computer Science at Università Degli
Studi di Verona (2018) and graduated in math at the University of Milan
"La Statale" (2013). He received The Marie-Curie Individual Fellowships
H2020-MSCA-IF-EF-ST-2019, the BE-FOR-ERC 2020 grant from the Sapienza
University of Rome, the EG-Italy PhD thesis Award 2018 and the
Eurographics Young Researcher Award 2023. He is a member of the Junior
Fellow of Eurographics and a scholar of the European Lab for Learning
and Intelligent Systems (ELLIS). He works on geometry processing, 3D
shape analysis and artificial intelligence.
--
---------------------------------------------------------------------
Associate Professor, Loris Roveda
email:loris.roveda@supsi.ch
webpage:
https://orcid.org/0000-0002-4427-536X
https://scholar.google.it/citations?user=3un_pPgAAAAJ&hl=it&oi=ao
https://www.scopus.com/authid/detail.uri?authorId=56031943800
Istituto Dalle Molle di studi sull’Intelligenza Artificiale (IDSIA)
Scuola Universitaria Professionale della Svizzera Italiana (SUPSI)
Università della Svizzera italiana (USI)
Politecnico di Milano, Mechanical Department
---------------------------------------------------------------------
2 months, 1 week