| Date: | Friday, Mar. 27 |
|---|---|
| Time: | 11:00 |
| Location: | Seminarraum 109, Engehalde, E8 |
Our guest speaker is Radu Timofte from the University of Wurzburg.
You are all cordially invited to the CVG Seminar on March 27th, 2026 at 11:00 am CEST
Computational photography and low-level computer vision are research areas with significant impact on both academia and industry. This talk reviews trends in computational photography and imaging. We will cover neural image signal processors (ISPs), advances in image restoration and enhancement, and image domain mapping problems. Our focus will be on challenges, findings, quality of the results, complexity and readiness of the solutions for real-world applications.
Radu Timofte is a Full Professor (W3) for AI and Computer Vision at the University of Wurzburg. Previously, he worked at ETH Zurich as a postdoc (2013-2016) and lecturer and research group leader (2016-2022). He earned his PhD degree from KU Leuven, in 2013. He serves(d) as an associate editor for top journals: CVIU, IEEE TPAMI, Elsevier Neurocomputing and SIAM Journal on Imaging Sciences. He regularly serves(d) as an (Senior) Area Chair/SPC for top vision and machine learning venues: ICCV, ECCV, CVPR, IJCAI, NeurIPS, AAAI, ICLR, ICML. Radu Timofte is part of the organizing teams of ICIP'26 (Tampere) and ECCV'28 (Bucharest). He and his team received multiple awards, including a 2022 Alexander von Humboldt Professorship Award and a 2021 Romanian Academy Award. He is co-founder of Merantix, co-organizer of NTIRE, CLIC, AIM, MAI, and AIS events, member of IEEE, CVF, and an ELLIS Fellow. His current research interests include augmented perception, mobile AI, multimodal learning, and image/video manipulation.
| Date: | Friday, Nov. 21 |
|---|---|
| Time: | 14:45 |
| Location: | N10_302, Institute of Computer Science |
Our guest speaker is Tim Steinert from the University of Bern.
You are all cordially invited to the CVG Seminar on November 21st, 2025 at 2:45 pm CEST
We study the incorporation of equivariances into vector-valued GPs and more general classes of random field models. While kernels guaranteeing equivariances have been investigated previously, their evaluation is often computationally prohibitive due to required integrations over the involved groups. In this work, we provide a kernel characterization of stochastic equivariance for centred second-order vector-valued random fields and we construct integration-free equivariant kernels based on the notion of fundamental regions of group actions. We establish data-efficient and computationally lightweight GP models for velocity fields and molecular electric dipole moments and demonstrate that proposed integration-free kernels may also be leveraged to extract equivariant components from data.
Tim Steinert is a PhD student in statistics at the University of Bern, working in the research group of Prof. David Ginsbourger. His research focuses on kernel design and inference for Gaussian process models, with applications to equivariant modeling of molecular data. He holds a Master’s Degree in Applied Mathematics from ETH Zurich.
| Date: | Friday, Nov. 7 |
|---|---|
| Time: | 14:45 |
| Location: | N10_302, Institute of Computer Science |
Our guest speaker is Giuseppe Bruno from the Institute of Mathematical Statistics and Actuarial Science (IMSV) at the University of Bern.
You are all cordially invited to the CVG Seminar on November 7th, 2025 at 2:45 pm CEST
While transformers have revolutionized machine learning, a fundamental understanding of how they construct internal representations remains a central challenge. This talk will present a recent theoretical framework that models the evolution of tokens as a mean-field interacting particle system, with network depth interpreted as time. The resulting mathematical description of the token distribution shows that, under certain regimes, tokens self-organize into clusters across multiple timescales, creating structure from initially random states. This mechanism offers a potential explanation for how meaning emerges in these models, while uncovering links to classical mathematical equations and other machine learning paradigms, and raising several open problems.
Giuseppe Bruno is a PhD student at the Institute of Mathematical Statistics and Actuarial Science (IMSV) at the University of Bern, working in the research group of Prof. Andrea Agazzi. His research explores the mathematical foundations of machine learning, with a specific focus on interacting particle systems and the theory of transformer models. He holds a Master's Degree in Mathematics from the University of Pisa.
| Date: | Friday, Oct. 3 |
|---|---|
| Time: | 14:45 |
| Location: | N10_302, Institute of Computer Science |
Our guest speaker is Eloy Mosig from the University of Pisa.
You are all cordially invited to the CVG Seminar on October 3rd, 2025 at 2:45 pm CEST
In this talk, we study the quantitative convergence of trained shallow neural networks to their associated Gaussian processes in the infinite width limit. While previous work has established qualitative convergence under broad settings, precise, finite-width estimates remain limited, particularly during training. We provide explicit upper bounds on the quadratic Wasserstein distance between the network output and its Gaussian approximation at any positive training time, demonstrating polynomial decay with network width. Our results quantify how architectural parameters, such as width and input dimension, influence convergence, and how training dynamics affect the approximation error. This is joint work with Andrea Agazzi and Dario Trevisan.
Eloy Mosig is a PhD student at University of Pisa which is currently visiting Professor Andrea Agazzi's team at IMSV in Bern. His main research interests lie at the intersection of probability theory, machine learning and applied topology. He holds a Master's degree from University of Bologna.
| Date: | Friday, Jun. 27 |
|---|---|
| Time: | 14:45 |
| Location: | N10_302, Institute of Computer Science |
Our guest speaker is Timo Blattner. He will present his Master Thesis.
You are all cordially invited to the CVG Seminar on June 27th, 2025 at 2:45 pm CEST
Recently, it has been shown that all mammal brains fold in a similar fashion, following the same mechanical model of folding. However, cetaceans remain outliers, having a systematically more folded brain than expected. A current hypothesis suggests that this is due to the increase in ambient pressure on the brain when these species dive, but this remains to be shown. Reconstructing these cortical surfaces is extremely difficult due to their high degree of folding and has never been done accurately before. We present a novel cortical surface reconstruction method, based on a few-shot learning of 2D expert manual tracings in each scan, to segment the full 3D image. From the segmentation, we reconstruct the white matter surface and displace it to the pial surface using a diffeomorphism. We successfully reconstruct the brains of 3 non-cetacean and 4 cetacean brains. We investigate the number of labeled slices needed for training a model to accurately reconstruct the cortical surface, and benchmark our method in humans. We show that these models can be used to label unseen scans of anatomically similar species, eliminating the need for manual labor. Our measurements support the validity of this pressure hypothesis.
Timo Blattner is a Master's student in Computer Science at the University of Bern. During his studies, he worked part-time as a research assistant in the Neuroradiology Department at the University Hospital of Bern, where he focused on deep learning-based segmentation and neuro-morphometric measurements aimed at improving clinical diagnostics. His research sparked international collaborations with partners in the UK and Brazil, allowing him to broaden his knowledge from clinical applications to the wider field of comparative neuroscience and the foundational scaling of brain morphology.