The FIAS ombuds person visits the FIGSS seminar.

Who would win: 1 bowling ball or 1000 ping pong balls? In this talk, I will investigate the hadronic equivalent of this question.  Up and down quarks are the basic constituents of the particles we think of as "everyday matter". However, in relativistic nuclear collisions, heavier flavors of quarks appear, produced out of vacuum in particle-antiparticle pairs. Because of their large mass, they punch through the surrounding medium, coming out  of it confined in heavy hadrons. While the initial evolution is thoroughly under investigation, less attention has been paid to the later stage when these hadrons rescatter within the dilute hadron gas. In this talk, I construct a toy model for the late evolution, systematically adding complexity and comparing to basic observables. 

When nuclei collide at nearly the speed of light, they create tiny fireballs so hot and dense that quarks and gluons, usually trapped inside protons and neutrons, can move freely. One clear sign of this extreme state, known as the quark–gluon plasma, is the enhanced production of strange quarks.

Interestingly, similar effects also appear in much smaller collisions, where a full plasma is not expected. This suggests that the underlying cause may already be present in the strong color fields connecting quarks during the collision. As these fields overlap and intensify, they can pull new strange quark–antiquark pairs from the vacuum.
By modeling this process in the SMASH framework and comparing it to hydrodynamic descriptions and experimental data, we study how strong color fields may give rise to plasma-like behavior even before the quark–gluon plasma fully forms.

In this talk, I will present a deep learning approach for identifying

events related to quark–gluon plasma (QGP) formation in the Compressed Baryonic Matter (CBM) experiment at FAIR. Using data from PHSD and UrQMD transport models, convolutional neural networks (CNNs) were trained to distinguish QGP and non-QGP events and to reconstruct key collision parameters. The method achieves up to 95% accuracy on simulated data and maintains over 80% efficiency after reconstruction effects, demonstrating its potential for real-time QGP triggering in CBM.

We would like to cordially invite you to the event "Life Beyond University" in the FIAS Lecture Hall (0.100) on Friday, 28 November 2025 at 14:15. Our guest will be Dr. Karla Viviani Allebrandt (Boehringer Ingelheim) who will give insights into the career at industry after university from a personal perspective particularly targeting PhD candidates and postdocs. After the talk, there will be ample time for discussion as well as coffee and cookies.

Life Beyond University is organized by Dr. Harald Berchthold and Prof. Dr. Martin Grininger (Goethe University) in collaboration with the graduate school FIGSS at FIAS. The two events this semester will take place at FIAS on 28 November and 30 Januar.

Infancy can be a powerful source of inspiration for autonomous robots. But connecting human development and robotics requires understanding how learning arises from early sensorimotor experiences. In this talk, I present a new framework that retargets infant motion onto humanoid robots to emulate first-person sensorimotor experience. From a single video, our method reconstructs the infant’s body configuration by extracting its skeletal structure and estimating the full 3D pose of each frame. The reconstructed motion can then be mapped onto multiple humanoids, including developmental platforms: the physical iCub robot and the virtual simulators pyCub, EMFANT and MIMo. Replaying the retargeted motions on these embodiments produces realistic multisensory streams including proprioception (joints and muscles), touch, and vision. The retargeting achieves sub-centimeter accuracy and enables a rich multimodal analysis of infant development as well as expert-level automated annotation of behaviors. This framework provides a unique window into the infant's sensorimotor experience, offering new tools for robotics, developmental science, and early detection of neurodevelopmental disorders.

Understanding invariances and biases in visual recognition requires stimuli where nuisance factors can be controlled independently. In this study, we introduce a controllable 3D stimulus dataset built in Blender comprising 100 semi-realistic meshes aligned with ImageNet categories. The dataset is designed to be compatible with off-the-shelf pretrained vision models while enabling systematic manipulation of scene parameters. For each object, we generate large parametric stimulus sets by sampling a multi-dimensional scene space (e.g., distance and illumination) and evaluate a range of pretrained architectures. We propose a regression-based analysis that maps model outputs to scene parameters, yielding interpretable sensitivity profiles at the level of models and objects. We extend the same framework to internal activations, quantifying parameter modulation at the level of individual units to identify neurons whose responses are selectively influenced by specific scene factors. Finally, we introduce a stimulus-optimization procedure that searches the same parameter space to estimate preferred stimuli for each model–object pair, and we use these preference profiles to define a functional-similarity measure and induce a hierarchy over models.

We would like to cordially invite you to the event "Life Beyond University" in the FIAS Faculty Club (4th floor) on Friday, 30 January 2026 at 14:15. Our guest will be Dr. Umar Jan (European Operations Sabinsa Europe GmbH) who will give insights into the career at industry after university from a personal perspective particularly targeting PhD candidates and postdocs. After the talk, there will be ample time for discussion as well as coffee and cookies.

Life Beyond University is organized by Dr. Harald Berchthold and Prof. Dr. Martin Grininger (Goethe University) in collaboration with the graduate school FIGSS at FIAS.

Cryo-electron microscopy permits imaging of biomolecules at

nanometer scale. Thousands to millions of 2D projections of individual
particles are collected and averaged to produce a high resolution 3D
reconstruction. With this averaging process, the structural
heterogeneity captured across projections is lost. In this endeavour, we
leverage simulation based inference in order to retrieve hypothesized
structural variability directly in the individual projections. This
framework is applied to the rho-type GABA receptor biological system and
through synthetic data, the requirements of characterizing a transient,
rare conformational state of the system are identified.