Cognitive stability and flexibility are core functions in the successful pursuit of behavioral goals. While cognitive stability allows us to follow behavioral plans in an environment full of rich, distracting stimuli, cognitive flexibility enables us to adapt to relevant changes in environmental demands. We developed a simple, fast behavioral experiment that allowed us to quantify both of these abilities in individual subjects in terms of their ability to switch between two different tasks when instructed to (flexibility) and in terms of their ability to ignore salient distracting stimuli (stability).

On a biological level, these functions require the stable maintenance and flexible manipulation of representations encoding the currently relevant behavioral strategies in so-called “working memory”. These representations have to be active, in the sense that they can be read out and modified quickly, to guide and adapt behavior. This is why we developed a model of neural circuits underlying the individual behavior of subjects in this task using attractor network models, where information can be stored in stable states of neural activity. The very same basic architecture cannot only be used to store information in the stable states of the system, but also to convert continuous input signals to categorical decisions by biased transitions between such stable states. This allowed us to create a simple network implementing the working memory maintenance of the currently relevant task rule and the integration of this rule representation with individual stimuli into discrete behavioral decisions. We could fit the individual behavioral statistics of 20 subjects performing our task. Furthermore, to test the biological relevance of our model, we used it to predict the individual time courses of energy consumption of a hypothetical brain network implementing the working memory part of model. Using functional MRI data as a proxy for neural energy consumption, we found a brain network closely matching these predictions, which consisted of two regions that were known to be consistently activated by task switching and inhibition experiments, but whose computational role was still under debate. Furthermore, we could use our model to quantify the individual stability of these working memory representations of the currently relevant task rule. We found that the energy consumption of a well characterized working memory updating network correlated with the individual stability of these representations, hinting at a mechanistic role of stochastic neural dynamics for stable and flexible behavior.

Stereoscopic vision is our visual system’s ability to extract information about the three-dimensional world from two two-dimensional images. The horizontal separation of the two eyes leads to the generation of two slightly different views of the visual scene. The differences between the two images are called binocular disparities. The brain uses these to extract depth information and, therefore, faces the challenge of matching individual features in the left eye’s image with the correct ones in the right eye’s image. In healthy humans, disparities are used to generate movements of both eyes that lead to the fixation of the same object.  

We present a biologically plausible model for the autonomous learning and self-calibration of such vergence eye movements. While granting insights into the processes happening in healthy brains, it provides the basis for studying common visual processing and eye disorders like strabismus and amblyopia, as well as possible treatments.

Neuroscience is one of FIAS' fields of expertise. It's the study of the nervous system, which comprises a multitude of nerve cells, so called neurons. Grasping the basic properties and functions of neurons allows us to appreciate the challenges and advances of current neuroscience.

I will introduce the neuron by reviewing past developments of neuroscience and presenting the basic structure and function of a typical neuron. The talk will mainly target non-neuroscientists.

Stochastic volatility models (SVM) are the major tool in academia and finance to model daily returns of stocks. In these models returns are normally distributed with a time-dependent standard deviation also called volatility. SVM successfully capture many of the stylised facts observed in financial data: fat tails, volatility-clustering, and leverage. However, calibrating SVM on financial data has turned out tricky because only returns of stock prices are directly observable but not their variance or volatility, respectively. Even though calibration has been widely recognised as the main issue in the applications of SVM it has been an unsolved problem to quantify precisely the amount of information returns deliver about their underlying volatility. Our group tackled this problem in terms of Shannon’s Information theory and revealed a surprising connection between the difficulty of calibration and a century old problem in thermodynamics already stated by Boltzmann: what is the convergence rate of a perturbed thermodynamical system toward its equilibrium? Mathematics has made a tremendous progress in solving this problem during the last fifteen years and proved exponentially quick convergence. It turns out that SVM are inherently hard to calibrate for the same reason non-equlibrium thermodynamical states are barely observable: they rapidly converge toward their equilibrium.

The investigation of heavy-ion collisions at accelerator facilities aims for the understanding of the properties of matter under extreme conditions, comparable to those which prevailed in the early universe or which can be found in the dense cores of neutron stars. On the macroscopic scale, goal of the experimental and theoretical studies is to determine the phase diagram of strongly interacting matter, while on the microscopic level one is interested in the individual particle's features, as for example the generation of its mass. In my talk I will show that macroscopic and microscopic in-medium phenomena are tightly connected to specific symmetry properties of nature, their breaking and restoration. I will give a historical and conceptual introduction to symmetries in physics and illustrate the connection to other fields of research. Further I will discuss why electromagnetic observables are good probes for the medium properties (and the according symmetries) governed by the strong interaction.

Supervised learning with a deep convolutional neural network is used to identify the QCD equation of state (EoS) employed in relativistic hydrodynamic simulations of heavy-ion collisions. The final-state particle spectra ρ(pT , Φ) provide directly accessible information from experiments. High-level correlations of ρ(pT , Φ) learned by the neural network act as an “EoS-meter”, effective in detecting the nature of the QCD transition. The EoS-meter is model independent and insensitive to other sim- ulation input. Thus it provides a formidable direct connection of heavy-ion collision observable with the bulk properties of QCD.

Conservation laws written in sets of partial differential equations are a powerful mathematical tool to describe all kind of phenomena from nature up to society. Solving such sets of equations numerically is a long-standing pillar of scientific computing. However, for the next generation of computers ("Exascale computing") which we expect to see in around three years, current methods gradually fail. I want to present modern concepts to tackle this limit. We apply these schemes to the set of 58 coupled first order differential equations which represent Einstein‘s theory of General Relativity in a split of "space and time". We evolve spacetimes of black holes and neutron stars with this method to study their dynamics and properties by measuring the ejected and recently experimentally detected gravitational waves.

Heavy-ion physicists try to reproduce the Big Bang on a subatomic scale by colliding the nuclei of atoms, hoping to recreate the hot and dense matter that filled the universe when it was only microseconds old. Some of mankind's most expensive experiments are designed to perform and observe such reactions. We will have a look at how they can be modeled to provide a theory for these measurements.

Fluid dynamics and magnetohydrodynamics are excellent tools to model a broad range of physical phenomena. This talk will provide a vary basic introduction to them and it will present a few examples of their applications within different contexts, from aerospace engineering to astrophysics, with a special focus on heavy-ion collisions.

Neural networks are powerful computational tools that can solve a variety of tasks, such as recognition of handwritten characters or stock market predictions. They are also suitable for analysing how the brain processes information and have been shown to be able to perform various sequence learning tasks.

This talk will give a brief introduction to neural networks and demonstrate that a simple self-organizing recurrent neural network can learn simple grammars by generating sentences by itself.

One of the most interesting questions in physics is how the universe was created. According to the Big Bang Theory, the known universe was expanded from a very hot and dense state of matter. In order to recreate and to study matter at such extreme conditions one would need to artificially create a medium with a very high energy density. For decades heavy-ion collisions at such facilities like the LHC at Geneva, RHIC in Brookhaven and in the future also at FAIR in Darmstadt are used to create  a medium of high temperature and energy density where the protons and neutrons split into their subparticles, the quarks and gluons, which under normal conditions cannot be observed. A new phase of matter is then created which is called the Quark-Gluon-Plasma (QGP). We use the strange K* vector meson resonances, particles that are very short-lived and not seen in normal matter, in order to study the properties of the QGP as well as the dense hadronic medium. We present our results and discuss how they relate to experimental data obtained in heavy-ion collisions at RHIC and LHC energies.