Project 1: Graph Neural Networks

Objectives:

We investigate which functions can be learned by graph neural networks.

Common graph neural networks cannot represent some seamingly simple functions on the space of all graphs. This limits the learning problems and applications they can be used for. The objective of our research is
1. to improve the understanding of what functions graph neural networks can express
2. to develop graph neural networks that can express more functions.

For this we intend to use existing concepts of theoretical computer science and graph theory in the context of graph neural networks.

Expected Results:

For (1) we intend to develop mathematical techniques that allow to more easily compare the expressivity of different graph neural networks. In our paper “Expressivity-Preserving GNN Simulation” (NeurIPS, 2023) we have proven that it suffices to represent many different graph neural networks as graph transformations with message passing which lays the foundation for this task. For (2) we intend to propose novel graph neural networks with guarantees on their expressivity and runtime.

Project 2: Estimates for Robustness/Safety/Reliability in (Distributional) Reinforcement Learning

Objectives:

To advance the state of the art in PAC (probably approximately correct) estimates in reinforcement learning (RL), in particular in view of the sharpness of the estimates. To implement various variants of such lower bounds on the performance. To compare them numerically in selected applications.

Expected Results:

In safety critical applications, (independent) estimates of the performance of estimates are among the best results that can be achieved. Due to the statistical nature of (reinforcement) learning, PAC estimates are the best results that can be achieved. However, the use of concentration inequalities in RL must address the time-dependent nature of these learning problems and many variants are conceivable, which complicates the quest for the sharpest bounds. Expected results are new and/or improved approaches to deriving estimates from various concentration inequalities and Bayesian learning; the PAC estimates (i.e. lower performance bounds) themselves, in particular for modern approaches such as distributional RL; implementations of the estimates to be used in conjunction with the Gymnasium interface. Impact is expected in safety critical applications such as medicine (intensive care) and energy systems (efficiency increase while in safe operating conditions).

Project 3: Towards Explainable and Knowledge-driven Large Language Models

Objectives:

The goal of the thesis is to investigate the inner workings of Large Language Models and other domain-specific Transformer-based architectures to gain a deep understanding of the learned representations and inference processes. This understanding shall then serve as foundation to build interpretable systems based on explicit knowledge, in particular with the goal to guarantee outputs of a certain quality and avoid so-called “hallucinations”.

Expected Results:

Methods to analyze and explain Transformer-based architectures with applications in various multimedia or cross-media domains; methods to separate natural language generation capabilities from learned knowledge representations; methods for representation and integration of knowledge bases in Transformer-based architectures; improved methods for training and inference in terms of computational efficiency and efficacy.

Project 4: Mathematical Methods in AI

Objectives:

Mathematics helps to describe and understand complex relationships modeled by AI methods. Reversely, AI methods are successfully used to solve complex mathematical problems. Specifically, the connection between statistics and ML has become much stronger over the last years, and both disciplines benefit from more flexibility of the methods and higher interpretability of the results. This SIG will strengthen CAIML with developments in this direction, jointly with other SIGs and partners of the network.

Expected Results:

In many fields of sciences, modern measurement technology leads not only to an increase of data, but also to more variety of data sources. One and the same process or phenomenon can be characterized and recorded in higher resolution, and at the same time with various different techniques (e.g. images, spectra, audio information, time/space signals, etc.). This results in the challenge to develop scalable analysis methods, and also to the task to link different data sources. This PhD project will develop and implement new methods to fully exploit the capacity of modern measurement devices, with the goal to gain deeper insights into the investigated processes. The methods will be tested with real applications from our faculty as well as from other fields (e.g. chemistry, resource management, construction engineering).

Project 5: Decision Support in Air-Traffic Control

Objectives:

Air Traffic Management (ATM) poses the most demanding planning and scheduling problems due to its real-time and safety-critical nature. Modern Artificial Intelligence (AI) methods have the potential to improve efficiency and performance while ensuring safety. The incorporation of Explainable Artificial Intelligence (XAI) holds significant potential in the field of ATM. XAI aims to demystify the opaque nature of AI decision-making, fostering transparency and trust. Imagine a system where aviation experts and controllers can readily understand the rationale behind crucial AI decisions. These advancements necessitate a robust understanding of potential benefits, with a focus on enhancing system usability and user-friendliness. Ultimately, the goal is to optimize decisions, leading to streamlined air traffic flow and reduced emissions.

The goal of the thesis is to explore and develop techniques that enhance the explainability and reliability of AI systems within the context of ATM. The research specifically addresses critical challenges related to safety, transparency, and interpretability. The automatisation potential of air- traffic control is to be investigated in particular at the example of remote towers, for which commercial demand is increasing.

Expected Results:

It is proposed to first try to extract the strategies used by air-traffic controllers by real-world data using, for example, inverse reinforcement learning in order to understand currently employed strategies. Then modern methods such as distributional and/or deep reinforcement learning will be used to either learn strategies from scratch or to improve the extracted ones. Finally, the policies/strategies will be evaluated in order to arrive at quantitative and fair comparisons.

To address large-scale and complex planning and scheduling problems that appear in the air traffic industry, our goal is to develop innovative hybrid solution methods leveraging synergies among artificial intelligence, logic, and mathematical optimization. These methods will integrate heuristic search, reinforcement learning, supervised machine learning, constraint programming, hyper- heuristic techniques, and mathematical programming. The effectiveness of these developed methods will be evaluated using real-world planning and scheduling problems encountered in the air traffic industry.

The results of this thesis, in the context of XAI, will provide a strategic roadmap, identifying which XAI approaches could yield benefits for specific ATM systems. Additionally, during the PhD project prototypical systems will be implemented and evaluated.

Project 6: Advanced Solving Techniques for Production Planning and Scheduling

Objectives:

Production scheduling problems arise in various industrial domains, as many factories have migrated towards a highly automated production process. Finding efficient schedules in these complex environments is often challenging for human planners. Thus, there is a strong need for automated solution methods. Although state-of-the-art AI methods and optimization techniques can provide solutions in a number of applications, there is still considerable scope for basic research in the field. New and challenging problems are arising and there is a need to further automate the process of problem solving. Therefore, this PhD project aims to investigate new solving techniques for problems in this domain.

The possible directions we plan to consider include novel hybrid techniques based on AI problem-solving methods and learning techniques, innovative and general features for algorithm selection, and new heuristic techniques based on decompositions.

The PhD project will be sponsored by MCP GmbH, a Viennese company that implements industry-specific tools and planning algorithms to optimize practical production scheduling problems. MCP GmbH will further provide access to data and instances of novel production scheduling problems from the industry, that we will investigate in this project.

Expected Results:

• We expect to address new challenging production scheduling problems.
• We expect to provide novel problem-solving methods that hybridize exact techniques with learning techniques for encountered real-life problems.
• We expect to identify innovative and general features for automated algorithm selection.

Project 7: Logical methods for Deontic Explanations in Law

Objectives:

Deontic reasoning, which involves obligation and related notions, is highly important in a variety of fields — from law and ethics to artificial intelligence. The combination of deontic logic and formal argumentation provides a fruitful theoretical basis for modelling this type of reasoning. We will use deontic logic and formal argumentation to develop a formal theory of “deontic explanations”.

Deontic explanations provide reasons why some deontic notions hold and others do not. They provide answers to complex questions like “Why should a child be entrusted to its father, rather than its mother, given a specific context?” or “Should someone who follows the faith of Jehova’s Witnesses be forced to undergo a life-saving blood transfusion? Why (not)?”. By targeting the understanding and transparent presentation of reasoning processes, deontic explanations are a major concern in many fields.

Driven by case studies in law, the PhD project aims to develop logical methods to formalise and reason about deontic explanations.

Expected Results:

By integrating both preference-based and norm-based explanations, LoDEx takes up the challenge raised by Makinson (1998) and Horty (2014) of formulating a unified logical theory combining several disconnected methods from the field of deontic logic.

By means of formal argumentation and dialogues, explanations are tailored to ensure explainee comprehension with the generation of fine-tuned explanations relative to the explainee’s preconceptions and expectations.

Project 8: Developing Declarative ASP Models with Interactive LLM use

Objectives:

Answer Set Programming (ASP) is a well-known approach for declaratively modeling and solving problems in a range of applications areas. The attractiveness of ASP stems from an intuitive reading of rules, a rich landscape of language extensions and modeling constructs, and the availability of efficient solvers. However, the construction of ASP programs, and the modification of those comes with some efforts and may be cumbersome to deal with. Besides ASP aficionados, this applies in particular also users with less background on ASP. The surge of large language models (LLMs) has opened a new perspective for interactive automated programming, which is explored for different programming languages, with remarkable results for widely used languages such as Python.

Expected Results:

The aim of this project is to investigate the use of LLMs for developing ASP models for problem solving in some specific domains, such as AI optimization (e.g., scheduling or configuration). The models should be developed in an interactive process between an LLM and a human user, giving feedback so that the LLM’s capabilities will improve over time. Furthermore, methods should be developed that allow for the explanation of aspects of the model and its behavior and properties, which are conveyed in natural language style to the user.

Project 9: AI for Dementia Care

Objectives:

Artificial Intelligence (AI) and Computer Vision have shown great potential in understanding and modeling human behavior. Our primary goal is to develop AI techniques that use unobtrusive depth sensors to measure behavior, with a particular focus on behaviors related to dementia, where behavioral manifestations are strongly correlated to cognitive impairment. The overall objectives include advancing our understanding of how to measure behaviors related to cognitive status and disease progression, with the ultimate goal of developing methods that are applicable and robust in real-world situations to improve dementia-related research and clinical practice.

1. Develop AI techniques that use unobstrusive depth sensors to measure the behavior of nursing home residents, with a particular focus on individuals with dementia.
2. Explore the application of computer vision deep learning models and different types of input data to optimize behavioral measurement methods.
3. Investigate differences between real-world data collected from individuals with dementia and controlled laboratory settings with healthy participants, with the goal of adapting models for robust performance in real-world conditions.

Create depth-based models that can account for individual variation and demonstrate generalization across domains, focusing on strategies for effective domain generalization and the role of self-supervised learning.

Expected Results:

By advancing the development of state-of-the-art computer vision models, we anticipate significant progress in processing depth image sequences. This effort aims to create privacy-preserving models with robust capabilities for detecting care needs and disease monitoring, thereby contributing to an enhanced quality of care and increased autonomy for individuals with dementia.

1. Achieve state-of-the-art results in the development of methods to measure the behavior of dementia patients by investigating different combinations of deep learning computer vision models and input data types for depth image sequences.
2. Provide insight into the differences between real-world and laboratory collected data for individuals with dementia, leading to improved model performance in practical settings and effective adaptation for adequate domain generalization. Develop depth-based models capable of detecting the need for assistance in people with dementia, using the lavatory setting as a use case, and develop effective methods of interacting with the user to guide the user in case of need.

Project 10: Declarative and Hybrid AI in Financial Knowledge Graphs

Objectives:

Knowledge Graphs are a particularly fertile ground for AI models that utilize declarative knowledge, as well as for those that combine symbolic and sub-symbolic AI models into hybrid AI systems.

A domain for which this is of special importance is that of finance and economics: Financial and economic data is on the one hand very interlinked, often forming graphs. Examples of this are companies and their interconnections, such as ownership, supplier/customer relationships, etc. On the other hand, both symbolic as well as sub-symbolic models are of critical importance: Declarative, symbolic models are critical to capture aspects of the domain where domain knowledge is available, often with the additional requirement of explainability and 100% certain matching with regulations, guidelines, etc. Sub-symbolic models are critical to capture unknown parts of the domain, such as certain types of illicit transactions, or other activities not directly known.

In this project our specific objectives are to develop Knowledge Graph methodologies for the financial and economic domain. Our aim is to focus in particular on the temporal domain, with financial data and reasoning having critical dependence on the aspect of time. On the side of declarative AI, our objective is to develop logical reasoners based on Vadalog that are able to handle such temporal, financial Knowledge Graphs as well as the scale of data encountered here. On the side of hybrid AI, we aim to include Knowledge Graph embeddings (KGEs), Graph Neural Networks (GNNs) and/or Large Language Models (LLMs), as required by the specific research questions.

Expected Results:

• Knowledge Graphs in the domain of finance and economics, with particular importance of the aspect of time
• Declarative AI approaches that are particularly effective for reasoning over financial and economic data, in particular with temporal reasoning capabilities and scalability

Effective Hybrid AI approaches including Datalog+- and Vadalog on the one hand, and the required mix of KGEs, GNNs and LLMs

Project 11: Integrating Large Language Models in Automated Constraint Programming Optimization

Objectives:

The main objective of this project is to combine the power of Large Language Models (LLMs) for automated Constraint Programming (CP) optimization. More specifically, the plan is to use LLMs to come up with streamlining constraints, implied (redundant) constraints, symmetry-breaking constraints, and dominance-breaking constraints. This task will require prompt engineering. The correctness of the LLM response will be automatically verified and checked with constraint-reasoning approaches, whereas the influence on the CP solving will be evaluated on training data. The verification and performance evaluation results will be fed back to the LLM.

Expected Results:

The demonstrated code-understanding capability of LLMs and the encouraging results of using LLMs for invariant generation (see, e.g., Wu et al., arXiv:2310.04870), let us expect that related techniques could also be applied in CP Optimization.

Project 12: AI-supported Higher Education Learning Spaces: AI-Enabled Teaching Strategy Optimization

Objectives:

This research targets higher education, focusing on AI-driven tools to enhance teaching strategies in diverse learning environments. The project aims to develop AI systems proficient at analysing higher education spaces and recommending adaptable teaching approaches for optimised student engagement.

Designed AI tools should guide professors using data on university classroom layouts, environmental factors, and instructional methods. The methods utilise large language models by providing instructions in natural language (e.g., prompting techniques). They would suggest teaching methodologies best suited for specific spaces, considering factors like room acoustics, lighting, and spatial design to maximise learning outcomes in higher education settings.

Collaborative efforts between educators, AI specialists, and design experts will shape an ecosystem where professors adjust their teaching methods dynamically. This research addresses the gap between instructional approaches and physical spaces in universities, fostering an aligned, interactive, and effective learning environment.

This research aims to improve higher education by using AI’s adaptability to optimise teaching strategies within varied spatial contexts, enrich student engagement, foster innovative pedagogy, and elevate educational outcomes in higher education institutions.

Expected Results:

AI tools that offer tailored recommendations for professors based on specific spatial characteristics, resulting in more effective teaching strategies aligned with diverse classroom environments.

Project 13: Supporting Senior Long-term Care Environment Design through Knowledge Graph-Driven Architectural Design

Objectives:

This research project aims to improve/support senior long-term care facilities design using knowledge graph technology. By intricately integrating elderly resident profiles, healthcare workflows, and industry best practices within a robust knowledge graph, the study aims to develop an intelligent system. This system will use advanced algorithms to suggest optimal spatial layouts and architectural configurations that meet the unique needs of senior residents and caregivers.

The developed system would analyze individual preferences, healthcare necessities, and comfort requisites through the intricate web of data in the knowledge graph. Its primary function will be to generate design recommendations prioritizing comfort, accessibility, and safety for senior adults within long-term care facilities.

This project aims to improve architectural design methodologies for senior adult long-term care facilities by emphasizing the utilization of knowledge graphs. By using this technology to propose tailored architectural solutions that align precisely with the distinct requirements of senior residents, the research aims to redefine long-term care environments for senior populations. Ultimately, the goal is to create living spaces that promote well-being and enhance senior adults' quality of life.

Expected Results:

Using knowledge graph technology, develop a set of personalized architectural recommendations based on resident profiles and healthcare data. These recommendations would prioritize comfort and safety, ultimately transforming care environments to improve senior residents' well-being and quality of life significantly.

Project 14: Atomistic simulations of electrochemical interfaces

Objectives:

The project aims at developing transferable neural-network force fields to simulate the atomic structure of metal-liquid interfaces. Methods should be developed to enable active learning and the inclusion of long-range interactions.

Expected Results:

There is only limited atomistic understanding of the structure of liquids at corrugated metallic surfaces. With special focus on metals and solvents relevant for the electrochemical reduction of CO2, this will be systematically investigated within the present project. To address the challenges posed by the necessary length and time scales, the development of transferable neural-network force fields capable of handling chemically complex multicomponent systems is necessary. To accomplish this, active-learning strategies will be developed to improve on databases of the chemical constituent and on pretrained foundation models. Additionally, to ensure that the models are transferable in the presence of charge, long-range interactions will be included based on equivariant message-passing networks.

Project 15: Microscopic derivation of effective lattice model Hamiltonians for long-range interacting atoms

Objectives:

The project aims to provide a functional renormalization group (fRG) based scheme for an improved construction of effective low-energy models whose precise determination is crucial for a quantitative theory. The fRG accounts for the renormalization of the bare interaction due to virtual transitions of particles from the low-energy bands in consideration to higher energy ones. For systems with short- range interactions, this has been shown to lead to effective multi-body interactions which can have an important impact on emergent collective behavior and the associated low-energy phase diagram. We will perform a systematic analysis of the interaction structures arising from the interplay between long-range interactions and virtual transitions / renormalization effects for paradigmatic setups.

On the methodological side, we will also make use also of the information-theoretic formulation of the fRG that allow us to reformulate the renormalization group flow as a variational problem. This enables the development of new numerical techniques and establishes a systematic connection between neural network methods and renormalization group flows of conventional field theories.

Expected Results:

Development of a quantum field-theoretical scheme for effective lattice models of long-range interacting Rydberg atoms and investigation of their physical behavior.

Project 16: AI for Communication in Hybrid Classic Quantum Systems

Objectives:

The goal of this project is to foster utilization of ML and AI for the integration of hybrid systems in the post Moore era. Quantum computers are feasible for computation of very specific computational tasks that can benefit from the native 3D modelling of the problem field. However, decomposition of an existing applications into suitable parts that can benefit from the execution on the quantum machine is an unsolved task. In this project we want to explore the AI methods for mapping (i) digital data representation into the quantum state; to (ii) analyse the worst case loading time and to (iii) develop methods to ensure low latency during the data transformation process.

The first goal is to develop theoretical concepts for the transformation of digital data representation into the quantum state by utilizing state of the art methods like hamiltonian evolution ansatz encoding or amplitude encoding. The second goal is to verify some of the prominent applications like molecular dynamics application or quantum machine learning with realistic data processing formats (e.g., data streaming).

Expected Results:

• Theoretical foundation to understand data transformation mechanisms in a hybrid quantum classic computational continuum.
• Benchmarks for various system relevant metrics applicable to data transformation in a hybrid quantum classic computational continuum.

In situ tests for real live applications in a hybrid quantum classic computational continuum.