Program overview with lectures, abstracts, and slides.
Time Schedule (Times in CEST)
Monday, July 3, 2023
How to find a natural grouping of a large real data set? Clustering or finding a natural grouping of a large set of objects is deeply rooted in human cognition. Our brain constantly clusters sensory stimuli in order to recognize, monitor and interpret them. Automatically clustering massive data is a challenging research problem for the following reasons. Sparse, high-dimensional and noisy data push state-of-the-art algorithms to their limits. Moreover, complex data can often be clustered in multiple meaningful ways. For instance, objects can be clustered by their shape or alternatively by their color. Each grouping represents a different view of the data set. The first part of this session introduces solutions for finding multiple alternative clusterings. Starting from the classical K-means algorithm, we will see how to find multiple subspaces with alternative K-means clusterings in moderate to high-dimensional data. To enable alternative clustering on very high-dimensional data, such as image collections, we integrate our ideas into deep learning. Our algorithm ENRC (for Embedded Non-Redundant Clustering) learns individual embedded spaces for each clustering. During the training process, we softly assign each dimension of the embedded space to the different clusterings and jointly optimize the clustering and the embedding. Results on image data show that ENRC can group the objects by color, material and shape, without the need for explicit feature engineering. This algorithm thus comes pretty close to the goal of automatically discovering multiple natural clusterings. In the second part of this session, we will introduce the result of our methods on real data from archeology. Glass beads were among the most common grave goods in the early Middle Ages, and their number can be estimated in the millions. The largest production areas of this time were in the Middle East and Southeast Asia. From there, most of the beads reached even the most remote areas of Europe. The color, size, shape, production technique and decoration of the beads are diverse. Accordingly, the previous classification systems are often subjective, complex and mostly limited to one burial field. As a result, the classifications are only conditionally applicable to glass beads from other cemeteries. In collaboration with archeologists from the Austrian Academy of Sciences, we classify the approximately 6,000 beads of the early medieval cemetery of Vienna-Csokorgasse using alternative clustering methods.. It is pioneering work as deep clustering methods have not been used on glass beads before. The session will be of mixed format. Basic concepts will be introduced in a talk followed by a tutorial-style exploration.
* The context: Knowledge graphs (KGs) have in recent years gained a large momentum both in academic research and in real-world applications. They have become a bridge between databases, artificial intelligence (AI), data science, the (semantic) web and semantic computing, linked data, and many other areas. In particular, in declarative AI, they have become a bridge between logic-based reasoning, and machine learning-based reasoning. We are going to explore this.
* The theory and systems: Languages for KGs on the one hand, and systems for KGs i.e., Knowledge Graph Management System (KGMS) on the other hand, have garnered increasing attention. Of particular importance are language and system extensions such as probabilistic reasoning, numeric reasoning, etc. - supporting various real-world applications, and the business applications that can be built using such extensions. We are going to dive into both theory and practice here, including the Vadalog system.
* The real-world applications: We focus on seeing Knowledge Graphs in action through a number of real-world and business applications, including: corporate governance, media intelligence, supply chains, collateral eligibility, hostile takeovers, smart anonymization, and anti-money laundering.
Tuesday, July 4, 2023
Günter Klambauer: Deep Learning, (slides for attendees only)
Over the last decade, machine learning and Deep Learning methods have paved their way into all kinds of computational task for molecules. The molecular machine learning research community has made strong progress in a) activity and property prediction, b) representation learning and molecular modeling, c) chemical synthesis and reaction prediction, and d) generative models for molecules. In this talk, we provide an overview over the main deep architectures, such as fully-connected, convolutional, RNN and Transformer architectures. We also provide a perspective of the recent progress in molecular machine learning, on the essential properties that our AIs should have to make a difference, and steps towards such broad AIs.
Katja Hose: Knowledge Engineering, (slides for attendees only)
Women in AI:
- Keynote Talk by Jana Eder, slides
- Panel Discussion: “AI & Data Science Career Pathways from Applied Industry to University”
- Gabriele Bolek-Fügl
- Diana Silvestru
- Dr. Jana Eder
- Prof. Katja Hose
- Carina Zehetmaier
Wednesday, July 5, 2023
In this course we will look into and critically discuss different aspects of natural language processing, with a focus on text processing. This will include both pipeline-based and end-to-end approaches, comprising classical and modern deep learning. We will address different NLP tasks and exemplarily discuss classical versus recent approaches to solving these tasks. As data is key to all levels of natural language processing, we will address topics of corpus creation both for learning and system evaluation, and talk about bias and annotation difficulties. Last but not least, we will address strategies for evaluation, including benchmark tests as well as testing in a concrete deployment context. Overall, we will give an overview of approaches to natural language processing, point out recent hot advances and present illustrating example cases.
- Statement of the Digital Humanism Initiative on Generative AI and Democratic Sustainability
- Summary of the Digital Humanism Summit
Nysret Musliu, Lucas Kletzander, Florian Mischek: Artificial Intelligence for Optimization, slides, video
Optimization problems arise in a variety of areas in industry, business, engineering, health care, etc. In this tutorial, we will give an overview of different AI methods and application areas/problems, such as planning and scheduling, timetabling, routing problems, and other optimization problems. In the first part of the tutorial, we will discuss various methods based on AI techniques for solving such problems. These topics include solver-independent modeling, constraint programming strategies, heuristic methods, and hybrid techniques. In the second part of the tutorial, we will present methods that use machine learning techniques for automatic algorithm selection and heuristic algorithm design. We will demonstrate the application of the presented techniques on several real-world application domains.
Stöckl im Park
Prinz-Eugen-Straße 25, 1030 Wien
Thursday, July 6, 2023
Gerhard Friedrich, Martin Gebser: Logic for Declarative Problem-Solving and Its Applications, slides (part 1), video (part 1), slides (part 2), video (part 2)
Answer Set Programming (ASP) is a knowledge representation and reasoning paradigm that has become popular for declarative problem-solving. The basic idea is to represent a complex application problem with a logic program such that specific interpretations, called answer sets, correspond to problem solutions. Powerful off-the-shelf ASP systems, such as clingo, DLV and IDP, automate the problem-solving process by first grounding a general problem encoding relative to an instance given by facts, and then performing Boolean constraint solving to compute (optimal) answer sets. The application areas of ASP include a variety of domains ranging from artificial intelligence, databases, and mathematical and scientific fields to industrial use cases. For instance, the clingo system has been utilized for radio spectrum reallocation in the first-ever incentive auction conducted by the Federal Communications Commission, which in 2016 yielded about 20 billion dollars in revenue. Likewise, the DLV system has been deployed as a core tool in enterprise software for e-medicine, e-tourism, intelligent call routing, and workforce management. Last but not least, the IDP system has been harnessed for interactive configuration in the banking sector. This tutorial introduces the basic syntax and semantics as well as the expressive first-order modeling language, which makes ASP attractive for representing and solving complex application problems. We illustrate the declarative problem-solving process on example applications, where we particularly highlight the proficient usage of optimization features that are central constituents in various application scenarios of ASP.
Reinforcement learning is the type of machine learning where an agent interacts with an environment, receives rewards, and learns policies that maximize the sum of all rewards. It is a general concept with many applications, e.g. in decision making, robotics, gaming, and medicine. Reinforcement-learning methods solved Go and power ChatGPT. The lecture starts with an introduction to reinforcement learning and its classical algorithms. Then recent techniques such as deep reinforcement learning and distributional reinforcement learning are discussed, as well as the use of reinforcement learning in large language models. Finally, applications are presented.
Panel Discussion: “Enabling sustainable profits AND serving society with AI”
- Georg Trausmuth - Frequentis
- Lukas Weinwurm - IMMOunited GmbH
- Nysret Musliu - Christian Doppler Laboratory for Artificial Intelligence and Optimization for Planning and Scheduling, TU Wien
- Moderator: Florian Michahelles - TU Wien
Friday, July 7, 2023
Robert Legenstein: Brain-Inspired Computation and Learning, (slides for attendees only)
The quest of Artificial Intelligence research is to build artefacts that mimic the cognitive capabilities of the human brain. Starting with the McCulloch-Pitts neuron exactly 80 years ago, the brain has always provided inspiration for systems and algorithms in AI research. In this tutorial, I will ask the question how AI research we can make use of our knowledge about the brain. I will discuss the current knowledge about the organisation of computations in brain. Next, I will present efforts that attempt to model such computations and how these models give rise to new machine learning approaches. One immediate application of this research is 'neuromorphic' hardware, that is, hardware that implements computation and learning based on principles borrowed from the brain. I will discuss the advantages and challenges of such neuromorphic systems.
In reaction to raising concerns about the ethicality of AI systems, dozens of proposals for ethical frameworks for artificial intelligence (AI) have been published. The frameworks present a case of principlism that is similar to the situation in medical ethics. This talk discusses whether principlism can solve the problem of ethical AI and whether frameworks are a practical tool for the AI designer. I present a systematic overview of practical approaches to make AI systems ethical, explain why they are hardly sufficient, and propose future directions for research.
Coffee breaks are scheduled at around 10:30-11:00 and 15:30-16:00.