Open positions: Cluster of Excellence

Open PhD positions# and PostDoc positions# in the Cluster of Excellence “Bilateral AI” at TU Wien (Vienna University of Technology)

We are seeking highly motivated and talented individuals to join our dynamic research team for combining symbolic and sub-symbolic AI. The successful candidates will conduct research at TU Vienna in collaboration with our partner institutes Johannes Kepler University Linz, AAU Klagenfurt, ISTA, TU Graz, and WU Vienna.

The vision of Bilateral AI is to educate a new generation of top-quality AI scientists with a holistic view on symbolic and sub-symbolic AI methods. Training and mentoring of young researchers is a central activity, which combines groundbreaking research work with an education program. The training will be distributed over the six participating institutions.

Location:

Successful candidates will have the opportunity to work together with leading experts in the field:

 

PhD job description:

PhD students will be trained within the Bilateral AI Doctoral School. Joint seminars, scientific workshops, and compulsory courses outside the PhD students’ research fields will also be designed to encourage interdisciplinarity.
Apart from that, students will be involved in grant applications, conference organization, Bachelor and Master student supervisions, and teaching. Each student will be supervised by two experienced and internationally renowned professors with different research fields (symbolic / sub-symbolic AI). The training will also provide a career development program, advice and support for students with innovative business ideas, and workshops for presentation and soft skills.

Open PhD positions: 13

Duration: 4 years

Application Deadline:

Open until filled. Applications will be processed on a regular basis. Only full application documents will be considered.

Our project is committed to increase the proportion of academic female faculty and, for this reason, especially welcomes applications by qualified women. If applicants are equally qualified, a woman will be given preference for this position.

 

How to Apply:

If you are interested in a position, you must submit the following application documents, included in a .zip archive, to the e-mail address ta.cigolnull@ia-laretalib:

  1. letter of motivation (detailing previous research achievements, research goals, career plans);
  2. a complete CV, including a list of previous scientific expertise, awards, grants, stays abroad, attended lectures, attended summer schools, attended workshops, skills, and publications (if applicable);
  3. abstract in English of the applicant’s MSc thesis, BSc thesis, or of a research project;
  4. a complete list of completed studies and transcripts of all grades;
  5. contact details of two reference persons (at least one academic) willing to provide a recommendation letter;
  6. proof of proficiency in English (usually TOEFL/IELTS/CAE);
  7. (optional) the selection of one or more research projects and related supervisors.

Furthermore, the reference letters should be sent (there will be no invitation e-mail) with the subject “Candidate name_of_the_candidate” to ten.ia-laretalibnull@secnerefer and ta.cigolnull@ia-laretalib-secnerefer within two weeks from submission.

 

Proposed Projects (to be extended soon):

 

PostDoc job description:

Post Doctoral positions have varying periods and provide opportunities for advanced BILAI training for developing a research career in academia or industry. PostDocs undertake cutting-edge research in collaboration with BILAI partners to advance the BILAI vision in the realm of its research program. Alongside refining their technical expertise, they participate in mentoring programs and training sessions on areas such as communication and career development.
PostDocs also have the opportunity to voluntarily participate in teaching and are expected to assist in supervising master’s and doctoral students within the BILAI research program.

Open PostDoc positions: 2

Duration: 2-4 years

Application Deadline:

Open until filled. Applications will be processed on a regular basis. Only full application documents will be considered.

 

How to apply:

If you are interested in a position, you must submit the following application documents, included in a .zip archive, to the e-mail address ta.cigolnull@ia-laretalib:

  1. letter of motivation (detailing previous research achievements, research goals, career plans);
  2. a complete CV, including a list of previous scientific expertise, awards, grants, stays abroad, skills, and publications;
  3. research statement, outlining the planned research, goals and significance, detailing the connection between symbolic and sub-symbolic aspects. The research topic must either refer to one or more specific projects (see below), or it must be described how it fits in the research plan of BILAI in a broader sense.
  4. selected supervisor (among those in the projects below);
  5. contact details of three reference persons who are willing to provide a letter of recommendation
  6. proof of proficiency in English (usually TOEFL/IELTS/CAE);
  7. (optional) the selection of one or more research projects and related supervisors.

Furthermore, the reference letters should be sent (there will be no invitation e-mail) with the subject “Candidate name_of_the_candidate” to ten.ia-laretalibnull@secnerefer and ta.cigolnull@ia-laretalib-secnerefer within two weeks from submission.

 

Proposed Projects (to be extended soon):

 

Normative Autonomous Agents
[PI: Agata Ciabattoni]

The rapid expansion of AI applications, from self-driving cars to elder care robots, has highlighted the growing importance of autonomous agents. Reinforcement Learning (RL) has emerged as a powerful method for developing these agents. As they become more integrated into society, it is essential to ensure their behavior adheres to legal, ethical, and social norms while effectively accomplishing their intended tasks.
The project focuses on designing norm sensitive autonomous agents.

Background: The successful candidate should have solid knowledge about Reinforcement Learning and formal knowledge presentation, preferably deontic/temporal logics and Answer Set Programming.

 

 

 


Details on Proposed PhD Projects:

 

 

Exact and Parameterized Complexity in Causality and Beyond

[PI: Robert Ganian]

By now, there is an extensive understanding of the exact conditions under which prominent computational problems arising in fields such as Constraint Satisfaction, Graph Theory or Logic can be solved using efficient algorithms. The identification of such conditions is typically carried out via a careful analysis of the problem’s computational complexity, either in the classical or the more refined parameterized paradigm. Yet, much less is known about the necessary and sufficient conditions to guarantee efficient algorithms for the central problems in causal analysis. The aim of this project is to change this via the development of novel techniques and insights, and to apply these new tools to design more efficient algorithms and corresponding lower bounds for problems arising not only in the broad field of causality, but also in related domains such as machine training and representation learning. The PI’s past work in the area of Bayesian Networks [1] can serve as a basic proof of concept demonstrating the viability and potential of this approach.

An ideal PhD applicant for this project should have a strong background in theoretical computer science or mathematics, possess at least some knowledge of algorithms and lower bounds, and be interested in pushing beyond the limits of current complexity-theoretic techniques.

[1] Robert Ganian, Viktoriia Korchemna: The Complexity of Bayesian Network Learning: Revisiting the Superstructure. NeurIPS 2021: 430-442. https://proceedings.neurips.cc/paper/2021/hash/040a99f23e8960763e680041c601acab-Abstract.html

 

 

Interaction of Machine-Learned Knowledge and Expert Rules

[PI: Emanuel Sallinger]

Current approaches in AI that aim for integrating symbolic and sub-symbolic knowledge typically try to guide the learning of neural networks with background knowledge formulated in formal logic – e.g., many Knowledge Graph (KG) Embedding approaches) – or, conversely, enrich logical reasoning frameworks with the possibility of adding predicates learned that have been acquired by in-transparent machine learning (ML) algorithms such as neural networks. The integration of machine-learned knowledge in the form of logical rules with human expert knowledge has, however, not received much attention in the literature yet. Yet, it is a quite typical phenomenon encountered in today’s KGs, where both types of rules naturally occur in a single KG. Although they often share the same syntactic form (e.g., Datalog), their integration is non-trivial.

The main objective of this PhD project is to investigate how human expert knowledge can interact with the learning of logical rules. Broadly speaking, we foresee the following three types of interactions of machine learned (ML) and explicit knowledge (EX), including (domain-specific) expert rules and first-principle knowledge:

(1) Use EX to improve ML: Expert knowledge can be used to check the consistency of ML rules, to select and schedule suitable learning problems, to organize the learning task, to modify the data, or to bias the rule learning modules into desirable directions. (2) Use ML to improve EX: Rule learning algorithms can be used to complement and refine expert-provided knowledge, e.g., by filling in missing slots, noticing unforeseen regularities, or compiling complex reasoning processes into efficient rules. (3) Integrating ML and EX into joint reasoning processes: EX rules will typically be more reliable than ML rules, which can be modeled in a joint reasoning process in several ways.

 

 

Boosting ASP-solving via Machine Learning

[PI: Stefan Woltran]

Answer-Set Programming (ASP) is a paradigm for declarative problem solving in which solutions of problems are described in a simple first-order rule-based language that is evaluated in terms of the nonmonotonic stable-model semantics. ASP has received increasing attention and is considered a core tool in the fields of Artificial Intelligence and Knowledge Representation and is more and more used in industrial applications.

Several advanced methods in ASP require good heuristics in order to decompose the input instance in a suitable way. In this project, we shall apply ML methods in order to tackle the so-called grounding bottleneck by suitably splitting the program and decomposing rules. The foreseen research goals include

(i) applying machine-learning in context of body-decoupled grounding [1], and

(ii) in combination with decomposition techniques [2];

(iii) extend these methods to constraint ASP systems [3];

(iv) apply the resulting system to real-world problems that are currently out-of-reach for ASP systems.

[1] Viktor Besin, Markus Hecher, Stefan Woltran: Body-Decoupled Grounding via Solving: A Novel Approach on the ASP Bottleneck. IJCAI 2022: 2546-2552

[2] Markus Hecher, Patrick Thier, Stefan Woltran: Taming High Treewidth with Abstraction, Nested Dynamic Programming, and Database Technology. SAT 2020: 343-360

[3] Tomi Janhunen, Roland Kaminski, Max Ostrowski, Sebastian Schellhorn, Philipp Wanko, Torsten Schaub: Clingo goes linear constraints over reals and integers. Theory Pract. Log. Program. 17(5-6): 872-888 (2017)

 

 

Reconciling Graph Neural Networks and Two-variable Logics

[PI: Mantas Šimkus]

Recent AI research has produced two powerful yet orthogonal families of methods for modelling and performing sophisticated computational tasks on graph-structured data, which is becoming increasingly popular and prominent (see Knowledge Graphs, Graph Databases, the Semantic Web standards). The first family is the so-called Graph Neural Networks (GNNs), which is a Machine Learning technique for various classification and prediction tasks on graphs. The second family are various logics based on the so-called two-variable fragment (FO2) of first-order logic. This family, which we call 2VLs, includes several prominent formalisms designed to model and reason about graph-structured data, like verious Description Logics (DLs), W3C Web Ontology Language (OWL), and W3C Shapes Constraint Language (SHACL). While GNNs and 2VLs share the underlying graph data model and purpose, the are also complementary in several ways: (a) explainability of 2VLs vs. black box nature of GNNs, (b) exact reasoning in 2VLs vs. approximate inferences in GNNs, (c) computational intractability of 2VLs vs. efficiency of GNNs, and (d) manual representation of domain knowledge in 2VLs vs. automated learning of representations in GNNs.

The main goal is this PhD project is to obtain methods and techniques for reconciling 2VLs and GNNs, aiming to obtain hybrid frameworks that have the benefits and avoid the drawbacks of the two ingredients. Specifically, the first concrete goal is to find ways in which GNNs can be used to efficiently approximate 2VL-based logic reasoning on graph-structured data, i.e. speed up reasoning while loosing some precision guarantees. The second concrete goal is to develop methods that allow a logic reasoner to consult a GNN in order to guide the automated reasoning process. This is particularly interesting for inference problems in 2VLs that are expensive due to non-deterministic computations.

 

 

Knowledge Representation Techniques for Trustworthy Specification of Agents’ Rewards 

[PI: Mantas Šimkus]

Machine Learning techniques based on Reinforcement Learning (RL) have shown the potential to yield extremely capable AI systems (e.g., by playing Go and Atari games at a professional level), and their capabilities are likely to have a profound impact on society. A key task in the development of a RL system is to design a reward function for an AI agent. However, currently reward functions are often developed and represented in ad hoc ways (e.g. using general purpose programming languages), which makes it difficult to inspect them, to explain them and ensure that they properly align with the values and expectation of the ultimate users of the system. Moreover, as the agent’s capabilities and freedoms expand, the number of possible states of the world that need to be taken into account in the reward function explodes exponentially. This complexity makes it difficult for the developers to exhaustively and adequately penalize all dangerous agent’s actions.

The main goal of this PhD project is to develop new knowledge representation languages and associated reasoning techniques that are specifically geared towards representing and reasoning about reward functions, thus providing a safer and more convenient alternative to the ad hoc error-prone representations currently employed. The developed logic-based methods should allow to write down reward functions that can be inspected for defects by means of automated inference, as well as shared and reused by multiple parties (thus facilitating the establishment of trust).

 

 

Large Language Models and Data Curation

[PI: Georg Gottlob]

Description: Large Language Models (LLMs) such as those from the ChatGPT, Gemini, or Llama families, are good at composing well-written texts, but they are generally not good at providing precise data (unless the data are easy to find in well-known sources such as Wikipedia). For example, when asked who the top 10 competitors of a company are, often over 50 percent of the results are wrong, and this even with systems that use retrieval-augmented generation (RAG). We have thus started the Chat2Data project for automatic data curation, and developed methods and software for automatically obtaining high-quality data from, or with help of, LLMs. Chat2Data can be used to verify existing databases by identifying suspicious records, to enrich existing databases, or to build databases entirely from scratch. The underlying techniques are automated problem-decomposition techniques, intelligent LLM interrogation via automated knowledge-based prompt engineering and answer-interpretation, advanced LLM-guided Web data extraction, automated logical consistency analysis over entire prompting dialogs involving rule inference engines or theorem provers, and LLM-supported rule mining. Chat2Data is thus an innovative example for a Bilateral AI where sub-symbolic AI (the LLM) interacts with symbolic AI (the knowledge-based control mechanism around it). There are many research questions in the context of Chat2Data that call for thriving PhD research work.

Requirements: Good background knowledge on machine learning and LLMs; excellent programming skills in languages such as Python, systems development skills; good logical thinking, high creativity, and ability to design new algorithms.

Desirable:  knowledge of or interest in  rule-based reasoning, computational logic,  rule mining, and/or database technology; knowledge of middleware such as, for example, Apache Flink.

 

 

Neurosymbolic Algorithmics

[PI: Stefan Szeider]

This project seeks to integrate symbolic computational reasoning techniques (such as SAT, MaxSAT, #SAT, QBF, and CP) with cutting-edge machine learning approaches, including generative methods like transformer models and predictive techniques like reinforcement learning. The goal is to develop algorithms that surpass traditional methods in efficiency. Beyond creating these ML-enhanced solving techniques, the project will also focus on designing and analyzing benchmark sets for symbolic solvers.

The project requires a good understanding of both areas (Symbolic Solving and ML) that we will combine. So ideally, the candidate has experience in both areas, but we will also consider candidates with a strong background in one of the areas and are willing to gain expertise in the other. Excellent programming skills in languages such as Python and the ability to design efficient algorithms are required.

Indicative references:

[1] Hai Xia, Stefan Szeider: SAT-Based Tree Decomposition with Iterative Cascading Policy Selection. AAAI 2024: 8191-8199. https://doi.org/10.1609/aaai.v38i8.28659

[2] Florentina Voboril, Vaidyanathan Peruvemba Ramaswamy, Stefan Szeider: Realtime Generation of Streamliners with Large Language Models. Arxiv, August 2024. https://arxiv.org/abs/2408.10268

[3] Johannes Klaus Fichte, Daniel Le Berre, Markus Hecher, Stefan Szeider: The Silent (R)evolution of SAT. Commun. ACM 66(6): 64-72 (2023). https://doi.org/10.1145/3560469

 

 

AI Alignment and Dialogues

[PI: Kees van Berkel]

Description: AI Ethics has many perspectives, one of which is AI alignment: Ensuring that AI behaves conform human values and principles. Some immediate challenges are: how to express values and normative principles in AI systems? What is the logical structure of reasoning with norms, values? And, how to resolve conflicts between norms and values? In this project, these questions are investigated. In particular, we investigate how to make formal reasoning with norms and values more transparent, to enhance the explanatory power of existing formalisms, with the aim of increasing understanding and trust. The focus will be on the development dialogue models, that bring together symbolic and sub-symbolic methods. The long term aim of such models is to generate explanations via an interactive argumentative exchange between a human and a given formal system

 

 

Learning-Based Scheduling, Planning, and Search Data Structures

[PI: Nysret Musliu]

Planning, scheduling, and other combinatorial problems arise in a variety of areas of industry, business, and engineering. Search (indexing) problems are at the core of efficient databases and information retrieval. A common aspect of the two settings is that practical implementations require massive amounts of time-intensive, instance-specific fine-tuning to perform well. Machine learning techniques are a very promising approach for these domains. We plan to investigate novel approaches for integrating symbolic AI methods and learned components to solve problems in scheduling, planning, combinatorial optimization, and search data structures. Besides these solution frameworks, we will also investigate the explainability of solutions. Symbolic methods will include techniques based on constraint programming, SAT, and discrete data structures, while learning-based methods will include classical, deep learning, and reinforcement learning-based approaches.

 

 

Domain Knowledge for Neurosymbolic Reasoning and Learning

[PI: Magdalena Ortiz]

In the last few years, we  have witnessed the emergence of formalisms that combine the power of neural networks for tasks like perception and pattern recognition, with high-level symbolic reasoning with rules. Some such successful neurosymbolic approaches incorporate probability distributions over atomic facts—which capture the output of a neural network—into existing (probabilistic) logic-based rule languages. DeepProbLog does this on top of probabilistic Datalog [1], and neurASP proposes a similar extension of Answer Set Programming [2].

The goal of this project is to incorporate such probabilistic rules with ontologies of explicit, structured knowledge written in description logics (DLs). Such ontologies have been successfully combined with tradicional rule formalisms like Datalog and ASP [3,4],  enabling advanced reasoning with complex domain knowledge.  The goal of this project is to do the same with neurosymbolic rule languages like DeepProbLog and neurASP.  We will develop algorithms for knowledgeable learning, inference and prediction in these neurosymbolic knowledge bases. Naturally, we will have not just one formalism a toolkit of them, resulting from different choices of DLs, different classic of predictive rules, and different restrictions on their interactions. We will therefore investigate the usefulness of different variants for different kind of use cases.

[1] Giuseppe Marra et al. “From statistical relational to neurosymbolic artificial intelligence: A survey”. In: Artif. Intell. 328 (2024), p. 104062.

[2] Zhun Yang, Adam Ishay, and Joohyung Lee. “NeurASP: Embracing Neural Networks into Answer Set Programming”. In: IJCAI. ijcai.org, 2020, pp. 1755–1762.

[3] Boris Motik and Riccardo Rosati. “Reconciling description logics and rules”. In: J. ACM 57.5 (2010), 30:1–30:62.

[4]  Labinot Bajraktari, Magdalena Ortiz, and Mantas Simkus. “Combining Rules and Ontologies into Clopen Knowledge Bases”. In: AAAI. AAAI Press, 2018, pp. 1728–1735.

 

 

Visual Analytics & Dynamic Network Visualizations

[PI: Silvia Miksch]

Knowledge Graphs (KG), causal representation learning, causal relationships and reasoning, etc., capture graph-based structures.

Visualization/Visual Analytics (VA) methods assist humans in exploring and comprehending complex and multi-variate data and information. VA intertwines the capabilities of computers and humans by combining automated analysis with interactive visualization methods enriched by cognitive and perceptual principles to solve data analysis tasks. Network visualization provides meaningful representations of graph-structured data, facilitating users’ understanding of the connections, insight gathering, and detection of unexpected patterns.

The objective of this project is to examine the utility of Visual Analytics methods, particularly dynamic network visualizations and event-based visualizations, in enabling collaborative exploration, quality assessment, and curation of KGs as well as the exploration and communication of causal relationships and reasoning.  Our aim is to assess the impact of our contributions within the healthcare sector.

[1]    Aigner, W., Miksch, S., Schuman, H., Tominski, C.: Visualization of Time-Oriented Data, Human-Computer Interaction, Second Edition, Springer, London, UK, 2023, ISBN: 978- 1-4471-7527-8. DOI: 10.1007/978-1-4471-7527-8. [Online]. Available: https://timeviz.net.

[2]    Beck, F., Burch, M., Diehl, S., Weiskopf, D.: A Taxonomy and Survey of Dynamic Graph Visualization, Computer Graphics Forum (CGF), 36(1):133–159, 2017. DOI: 10 .1111/cgf.12791.

[3]    Filipov, V., Arleo, A., Miksch, S.: Are We There Yet? A Roadmap of Network Visualization from Surveys to Task Taxonomies.
Computer Graphics Forum (CGF), 42(6):e14794, 2023. DOI: 10.1111/cgf.14794.

 

 

Integrating Logical Rule Languages with Sub-Symbolic Techniques

[PI: Thomas Eiter]

Logical rule languages such as Prolog and Answer Set Programs come with powerful reasoning engines for symbolic reasoning tasks. However, there are only few extensions that support perceptional reasoning tasks, which build on sub-symbolic reasoning, dealing also with associated uncertainty.

One obstacle is the mismatch between the underlying computational paradigms, which use different reasoning engines. Together with the black-box nature of neural atoms, used to access neural computation in logic programs, as in NeurASP, DeepProblog, or SLASH, the integration becomes opaque; bridging from symbolic terms to vectorized representations as used by neural networks, known as grounding, may be hidden or blurred.

Uncertainty would be natively supported by probabilistic programming.

We target in this project richer rule languages with model-based semantics, featuring stable negation in a seamless integration with subsymbolic AI, with a better understanding of the grounding issue, in connection with a dynamic world. Further aspects are novel and efficient reasoning methods, addressing the computational challenge of uncertainty in reasoning, and forming rule bases and updating them. The latter should be done by improved learning and rule acquisition methods, with possible user interaction in a transparent and accessible manner.

Background: The successful candidate should have solid knowledge about formal knowledge representation, preferably logic programming, logic and theory, and solid programming skills. Knowledge and skills about subsymbolic AI, specifically neural networks, is desired.

 

For more positions within the Cluster of Excellence, please visit: Jobs – Bilateral AI

Comments are closed.