BI Retreat 2023

Herbert Jaeger

In order to make materials-based computation a systematic engineering discipline, we need a general formal theory (GFT) of ‘computation’ in physical systems. A GFT would amalgamate elements from dynamical systems, probability and symbolic logic. My lifetime challenge is to develop a mathematical language in which a GFT can be stated.

Stephen Trenn

Obtain suitable mathematical models of energy networks incorporating switching effects and sector couplings - Formulate and prove relevant mathematical properties and design corresponding algorithms to support analysis and design of power grids - Develop toolboxes for diagnosis, stabilization and efficient (as well as safe) operation of power grids

Julian Koellermeier

Can we establish the use of mathematical model hierarchies that bridge the gap between small and large scales to allow for analytical insights and efficient numerical simulation of fluid dynamics applications?

Celestine Lawrence

The recent success of tools like ChatGPT has clearly demonstrated how far commercial AI technology can stray away from brain-like mechanisms and yet achieve close to brain-like functionality in terms of eloquent writing. I believe that this stray away from biology is only a fleeting success and that we will ultimately be back to bio-inspired methods, not just to enhance AI but also to better understand and manage the brain (be it for a human or a worm). Thus, I wish to see a great focus on "the engineering mathematics of Bionic Intelligence (BI).

Reka Szabo

Stability of Interacting Particle Systems: Interacting particle systems are stochastic models for the evolution of complex structures arising in statistical physics, biology, economics, …. They are stable, if their long-term dynamics is not affected by a small random noise. Some sufficient criteria for stability are known; the goal is to obtain a complete classification of these models.Interacting particle systems are stochastic models for the evolution of complex structures arising in statistical physics, biology, economics, …. They are stable, if their long-term dynamics is not affected by a small random noise. Some sufficient criteria for stability are known; the goal is to obtain a complete classification of these models.

Michael Biehl

Statistical Physics of Learning: To date, the practical success of Deep Learning is not accompanied by appropriate theoretical understanding. Statistical Physics techniques facilitate the analysis of learning processes in models of large adaptive systems. Putting forward and extending this approach will contribute to the development of efficient, deep architectures and the associated learning algorithms.

Paris Avgeriou

“Software is eating the world”, in terms of all aspects of modern life depending on software running on devices, machines or the cloud. As software systems, become larger and more complex (incl. AI systems), we need the right engineering tools to govern their design and evolution, especially in critical domains.

Andrea Capiluppi

"Continuous Software Maintenance": it is well known that software maintenance is the most expensive phase of a software product's lifecycle. It is a low-morale job that eats up developers' time and that prevents them from further evolving their system. Software maintenance will be increasingly performed by satellite players (SMEs, spin-offs, universities, students); this will allow these players to gain experience, visibility, and trust, while at the same time growing a learning community around software developers.

Ayushi Rastogi

We are witnessing a generation of software systems designed for short-term needs, which generally implies profit and the needs of a privileged few. The next generation of software systems should focus on long-term goals and help restore sustainability by designing and developing software that balances business and societal interests.

Vasilis Andrikopoulos

Make software more sustainable. Understand how the design, development, and operation of software interacts with different aspects of sustainability, both internally facing (e.g. technical qualities affiliated with system longevity) and externally facing ones (e.g. minimizing its environmental impact).

Dimka Karastoyanova

Information Systems are software systems managing and processing data, as well as for the purposes of knowledge extraction and use in many application domains. Current developments require a high degree of flexibility while the systems are being in use. The scientific challenge is to compile systematically a body of knowledge about how to architect and build such systems that are capable to adapt as required by different application domains and that accounts for the diversity of available approaches and technologies for building contemporary, enterprise information systems.

Cristobal Bertoglio

To develop automatic methods to create biomechanical models of the cardiovascular system from (raw) medical images from heterogeneous sources. Doing so, we would like to improve the accuracy, robustness, cost-effectiveness and/or safety of medical diagnosis for patients.

Oliver Lorscheid

Tropical methods are ubiquitous in mathematics and computer science and find application in the approximation of polynomial equations and as a framework for optimization problems. My main ambition lies in the development of new tropical tools, with an eye towards applications.The Riemann-Roch theorem is a 19th century result about Riemann surfaces, which was vastly generalized by Hirzebruch and Grothendieck in the 1950s. The tropical Riemann-Roch theorem is a combinatorial counterpart. I envision a cohomological understanding and generalization of this theorem in Grothendieck's sense.

Jiri Kosinka

In contrast to raster graphics (arrays of coloured pixels), vector graphics are represented using geometric primitives and colour interpolation schemes. This makes them infinitely scalable (zoomed-in/printed images never appear pixelated) and also suited for image compression. Despite recent attempts, image vectorisation, i.e., conversion from raster to vector graphics, remains an interesting challenge.

Revantha Ramanayake

Logic is the study of reasoning. Diverse scenarios demand diverse forms of reasoning. That's why so many different logics are utilised in Computer Science (and elsewhere).The formal computational properties (decidability, computational complexity) of many logics is still unknown. I want to map this landscape by developing broad theoretical methods.

Cara Tursun

Advances in deep learning have improved computer performance in visual tasks that were once deemed only possible by humans. However, the human visual system still performs remarkably due to the efficient processing and interpretation of visual stimuli. Complete emulation of human biological vision in computers is the next challenging step.

Helle Hansen

Dynamic modal logics of programs and games provide a formal basis for analyzing the correctness of complex software systems. However, their proof theory and computational properties are not yet fully understood. Cyclic proof systems for such logics have recently been studied and provide a promising approach to addressing this challenge.

Jorge Perez

Ensuring software reliability is an old but still very much open question. As large software systems consist of many diverse components, the way they interact is key to correctness. Interactions can be based on communications (as in processes sending/receiving messages) or sharing (as in processes accessing a central memory). Real programs combine these two forms of interactions, but we still lack unified foundations and tools for ensuring correctness in these relevant scenarios.

Heerko Groefsema

Although process-aware information systems allow for automation of business processes, the (re-)design of the processes themselves remains a manual error-prone task. More so when considering the increasing demand of regulatory adherence of such processes to (inter)national laws. The challenge, therefore, remains to allow for automated (re-)design [1] of formally provable regulatory compliant business processes.

Marco Grzegorczyk

As a statistician working in interdisciplinary collaboration with biologists and medical scientists, my main scientific challenge is to develop new statistical methods that can get the maximum amount of information out of the growing amount of molecular data. In particular, it will be important to develop new statistical techniques that can integrate data from different sources (`multi-omics data').

Davide Grossi

Public authorities, at all levels, are increasingly turning to technological solutions to try to boost citizens’ participation in political decision-making. But can we entrust algorithms with something as important as democracy? In particular, it will be important to develop new statistical techniques that can integrate data from different sources (`multi-omics data').

Fatih Turkmen

The data is the currency of the 21st century and its security is of utmost importance. In my research, I develop software/hardware solutions to protect data and the systems it resides in during storage, sharing and processing. This includes techniques all the way from the attack-resistant materials (e.g., memristors) used to produce chips to high-level cryptographic solutions to securely process data. Perhaps the most pressing challenge here is to secure AI systems that memorize things about (our) data during training and are left out there to be probed by the attackers to steal information about us.

Michael Wilkinson

AC/DC: Adaptive Cognition through Deep Connectivity: A core question in computer vision is how to optimally group pixels in images into meaningful structures. I will approach this problem by merging connected filters with machine learning, creating flexible framework which handles this problem in an explainable, and computationally (=energy) efficient way.

Andreea Sburlea

How does the perception of internal or external events influence our movement decisions? A professional violinist will focus on the rhythm of the music when improvising and not on the sequence of finger movements. How does the brain prioritize specific behavior by ignoring sensory irrelevant stimuli? I am interested in addressing this challenge from a unified computational, behavioral and neural perspective.

Cecilia Salgado

Algebraic geometry error correcting codes on higher dimensions: Algebraic Geometry Codes is a fast-evolving area in the intersection of pure mathematics and computer science: on one side number theory and algebraic geometry, and on the other, coding theory and information theory. While the theory of curve codes is well developed, the theory of surface codes is still lacking a uniform treatment, and higher dimensional variety codes have barely entered the picture. The development of the latter to a stage similar to the one-dimensional case is one of my main research challenges.

Bart Verheij

Artificial intelligence systems should be designed for responsible behavior to allow for a proper collaboration between humans and machines. For instance, AI systems should be able to communicate in a correct and reliable way, should be able to provide meaningful explanations for their behavior, and should be able to follow values and norms guiding their actions. The design of responsible AI systems requires a variety of AI methods, including knowledge representation, machine learning, natural language processing, formal reasoning, interaction design and cognitive robotics. Progress can be expected by combining mathematical, computational and AI approaches. The focus of this challenge is to work on the synergy between various AI methods in order to arrive at hybrid systems in which humans and machines can successfully and safely collaborate.

Zoe Christoff

Similarity is one of the main drivers of social networks dynamics: more similar agents are more likely to connect (homophily), and connected agents are more likely to become more similar (social influence). Similarity is also the main driver of recommendation algorithms. Understanding how these two types of similarity filters interact is crucial to understanding how information, opinion, and behavior are shaped by online tools.

Hamidreza Kasaei

Nowadays, robots work well in predefined settings, but any changes in the environment require complex, time-consuming, and expensive robot re-programming by human experts. My goal is to develop data‑efficient, continual learning techniques that allow robots to learn new concepts over time and safely interact with non‑expert users in open‑ended fashions.

Bart Besselink

Modular control of cyber-physical systems Challenge: Engineered systems such as high-tech equipment and smart energy systems comprise a large number of components, including physical as well as "cyber" components. Aimed at addressing the complexity and heterogeneity of such systems, a key challenge is to develop a control theory that is inherently modular, i.e., based on considering components independently while simultaneously guaranteeing desired system behavior.

J. D. Cardenas Cartagena

Safe artificial intelligence in cyber-physical systems: Artificial intelligence (AI) is reaching environments beyond the digital world. Indeed, AI supports everyday tasks in cyber-physical systems vital for society, such as robotics, electric grids, or healthcare monitors. Therefore, developing methods to guarantee human safety while interacting with AI becomes critical for a sustainable and human-centric industry 5.0.

Raffaella Carloni

The Robotics Lab develops robotic systems that are intended to physically interact with uncertain dynamic environments and to cooperate with humans. The main focus is on mechanical designs and model-based/model-free control architectures of novel bio-inspired compliant actuation system as solutions to medical and engineering challenges.

Jelmer Borst

Cross the bridge from neural spiking to human cognition, that is, explain how the firing from neurons in the brain can lead to the goal-directed behavior that we see around us every day.

Jacolien van Rij-Tange

Having computational systems learn language as human people do, which is different from chatGPT in that it will be able to understand fundamental aspects of human language including metaphors, sarcasm, and language jokes based on much less dataThe applications range from improving automatic language processing to facilitating language education.

Fokie Cnossen

AI systems are not 100% accurate and usually non-transparent. So how should we design AI systems so that human users of AI systems put appropriate levels of trust in such systems, so neither too much nor too little trust? A multi-disciplinary approach is needed to guide the design of AI systems.

Niels Taatgen

To construct a cognitive architecture of human intelligence that can learn new tasks through experience and instruction. This architecture should be able to exhibit transfer learning (using knowledge from one task for another), and have applications in the Machine Learning and in personalized education and training.

Catherine Sibert

My research aims to combine the theories and methods of computational modeling with the techniques of cognitive neuroscience with the goal of creating a new generation of architectural cognitive models inspired by the structure and functionality of human brains. These models could provide a framework for more human-usable Al systems.

Marieke van Vugt

I am in the field of Computational Psychiatry. The big challenge for the next 3 years is to improve our understanding of how we can use cognitive tasks to assess the mental state of an individual person. There is research to suggest that on a group level, there are differences between healthy individuals and those suffering from psychiatric problems such as depression, but can we use those to detect the decline in mental health before it becomes a full-fledged depression? A related question is how different data streams, ranging from subjective judgments of cognitive and emotional state to behaviour on cognitive tasks to ambient sensors of e.g., heart and space to characterize a person’s mental state.

Kerstin Bunte

Experts desire to know how their data can inform them about the natural processes being measured. We aim to unite the predictive power of machine learning and the explanatory power of modeling, to develop transparent and interpretable techniques. The power of novel hybrid methods will be demonstrated for applications in medicine and engineering

Stephen Jones

Large language models like Google Translate make mistakes that people don’t, when complex sentence structures or negative words like “without” make a mission-critical contribution to meaning. Understanding what happens when these words and structures are heard or read by humans helps us understand the brain better and will improve AI.

Alef Sterk

It is well-known that deterministic dynamical systems can be unpredictable. Predictability is traditionally computed for a system as a whole. Developing new mathematical methods for quantifying the predictability of specific events, such as high wind speeds in meteorological models, is a pressing challenge.

Gilles Bonnet

A large varieties of phenomena can be described by a complex structure with a geometric flavour, e.g.: 1) ``Geometric graphs``: model many type of networks (communication, electrical, ...) 2) ``Tessellations``: model of an aggregate material; or partition of a large dimensional space for signal compression purposes;3) ``Polyhedron``: space of feasible solutions of linear optimization problems.My challenge is to describe as accurately as possible these structures when they are constructed through a random procedure. This entails computing mean and variance asymptotics, central limit theorems and large deviation principles.

Holger Waalkens

The N-body problem is one of the paradigm systems in Dynamical Systems Theory. In the coming years I want to carry over results from the classical Newtonian case to more general interactions including charges and a relativistic setting.

Roland van der Veen

Knots are robust, flexible 3D shapes. Can we use and compute with them effectively and as a language to understand more complex phenomena? Find a simple topological foundation for the quantum physics techniques used in knot theory to facilitate applications in 3D and 4D geometry, (fluid) dynamics, algebra and logic.

Nikolay Martynchuk

Integrability of geometric flows: Invariant geometric flows are interesting PDEs used in mathematics and computer science. Many of such flows are integrable, but the mechanism for integrability is not well understood. I'd like to shed light on this question and on connections of such flows to finite-dimensional integrability.

Marcello Seri

Provide a comprehensive description of the classical-quantum correspondence in singular geometries, in particular in the sub-Riemannian context, and their potential applications. Laplace-Beltrami operators in this setting are singular and geodesic flows are very strange, requiring radically new ideas to understand and characterize the relation between the two.

Juan Peypouquet

First order algorithms are central in optimization and big data analysis. Inertial schemes, which use information from previous steps, highly improve their numerical performance, under certain conditions. For equilibrium problems (economics, games, mechanics), only the simplest inertial methods have been studied. The combination of inertia with restarting techniques, geometric analysis, optimal control and switched systems tools, is likely to produce much faster algorithms.

Tamas Gorbe

Symmetries & Conservation Laws - the interplay between integrable systems and gauge theories. I'd like to explore and exploit emergent links between gauge theories (the mathematical underpinning of particle physics) and integrable systems. This could lead to new special functions and connections between different realms of physics.

Monday 5 June

Location

• Computing and Cognition
• Systems, Data and Society
• Geometry and its Applications

• Lijster
• Specht
• Buizerd

Tuesday 6 June

Location