Lunchtime Seminar
Archive WiSe 2023/24 & SoSe 2024
Recommender Systems for Music Retrieval Tasks
Lecturer: Eva Zangerle - Researcher@DBIS
Date: 20.06.2024
Abstract:
Music is ubiquitous in today's world, and with the rise of streaming platforms, listeners have access to more music than ever before. Music recommender systems aim to help users discover and retrieve music they like and enjoy. Such user-centric retrieval approaches need to capture aspects that influence the user's perception of and preference for music. These aspects include music content (descriptors extracted from the audio signal, such as tempo or acousticness), music context (external factors describing the track or artist, such as the lyrics of a track), user characteristics (long-term descriptors of the user, such as general music preferences), and user context (short-term, dynamic factors describing the user, such as the current activity, occasion, or emotional state). User models that incorporate these aspects provide a better picture of user preferences and enable improved personalization. However, comprehensive user models and recommender systems that can incorporate rich user models are rare. In this talk, I will present comprehensive user and item models and how these models can be used in recommender systems. Furthermore, I will outline our current work on the evaluation of recommender systems and psychology-informed recommender systems
From Anthropomorphic to Zoomorphic Social Robots: Our Experiences
Lecturer: Patrick C. K. Hung - Faculty of Business and IT, Ontario Tech University Canada
Date: 13.06.2024
Abstract:
A social robot consists of a physical hardware component that interacts with humans connected through a network infrastructure as a cyber-physical computing system supported by cloud services. Human-robot interaction (HRI) is a research area that involves understanding, designing, and evaluating robots for use by or with humans. The Uncanny Valley theory describes the disturbing effect of imperfect human likenesses that have dominated HRI. Referring to the Uncanny Valley, social robots usually constitute a form of anthropomorphism. Social robots typically behave like humans or animals, such as mimicry of human/animal behavior and emotional expression, with speech, gestures, movements, and eye-gaze features. Prior research found that it is much easier for an embodied humanoid robot with emotional expression to gain users' trust to release personal information than a disembodied interactive kiosk. Emotions are essential to human cognition and behavior caused by an identifiable source, such as an event or seeing emotions in other people. During the pandemic, our research team designed and built a homemade robotic puppy with wood and mechatronics with a mechanical tail to express emotion from scratch with body language. This talk will then overview our social-technical research works from anthropomorphic to zoomorphic robots. This talk will also focus on a recent study of a guide dog robot for people with visual impairments.
Dynamic workflow scheduling in the edge-cloud continuum: Optimizing runtimes under budget constraints
Lecturer: Stefan Pedratscher - Researcher@DPS
Date: 06.06.2024
Abstract:
Scientific workflows are increasingly adopting hybrid Edge-Cloud infrastructures to benefit from the computational and storage capacity of the Cloud and the cost savings and data locality of the Edge. Workflow scheduling is one of the most challenging problems for the Edge-Cloud continuum. State-of-the-art workflow schedulers often rely on a centralized runtime system and are based on static algorithms that either focus on Cloud or Edge systems (but not both). In this work, we introduce a novel, open-source, and dynamic scheduler for scientific workflows that targets the Edge-Cloud continuum by design using fully decentralized runtime system instances. This not only reduces data transfer times but also leverages the benefits of the continuum. The proposed scheduler optimizes for runtime while adhering to a given cost limit by dynamically mapping tasks to resources and orchestrating groups of workflow tasks on runtime system instances. Furthermore, the scheduler adapts to real-time updates in task durations, accommodating for variations in resource performance, to efficiently use the cost limit and to reduce the total runtime of the workflow.
Equational narrowing and multiset narrowing and their formalization in Isabelle/HOL
Lecturer: Dohan Kim - Researcher@CL
Date: 23.05.2024
Abstract:
Narrowing is a generalization of rewriting, where matching used in rewriting is replaced by unification. In this talk, I will first give an overview of narrowing, E-unifiability, and reachability. Then I will talk about equational narrowing and describe how it is applied to reachability analysis and E-unifiability problems. In particular, I will discuss an Isabelle/HOL formalization of equational narrowing and its applications to reachability analysis and E-unifiability problems. Next, I will introduce the proposed multiset rewriting and multiset narrowing and describe the difference between (ordinary) narrowing and multiset narrowing. I will explore how multiset narrowing is applied to multiset reachability analysis, reachability analysis, and E-unifiability problems consisting of multiple goals. Furthermore, I will present an Isabelle/HOL formalization of multiset narrowing and its related results on multiset reachability analysis, reachability analysis, and E-unifiability problems. Finally, I will conclude the talk by discussing potential future research directions on multiset rewriting and multiset narrowing along with their formalization in Isabelle/HOL
Expanding the scope, security and efficiency of classical symmetric primitives
Lecturer: Elena Andreeva - TU Wien
Date: 16.05.2024
Abstract:
I will start this talk by presenting a novel class of symmetric primitives, which in contrast to classical input-length-preserving block ciphers and permutations, and input-length-compressing compressing hash functions, expand the size of their inputs. Through the new designs of forkciphers and general expanding pseudorandom functions (PRFs), I will demonstrate the power of expanding primitives to enable higher security and efficiency in a wide class of applications, such as:
- Key Derivation (in Signal-like messaging protocols) and Pseudo-random Number Generation
- Encryption, Authentication and Authenticated Encryption
- IoT-to-Cloud Computation
- Lightweight-device communication
Finally, I will conclude the talk by presenting novel research directions in the area of secure theory and applications or expanding primitives in cryptography.
Towards advanced training for in-office hysteroscopy
Lecturer: Vladimir Poliakov - Volocopter GmbH, Germany
Date: 02.05.2024
Abstract:
Hysteroscopy, an essential gynecological procedure, enables minimally invasive examination and treatment of intrauterine issues, accessed through natural openings, reducing the need for incisions. Technological advancements have miniaturized instruments, thus facilitating outpatient procedures termed in-office hysteroscopy. However, this approach poses new challenges in patient comfort, necessitating tailored training. Virtual reality (VR) simulation, increasingly popular in minimally-invasive surgical training, offers high variability and visual realism, though some challenges still remain. This talk covers both the software and hardware aspects of designing a VR training platform for in-office hysteroscopy. We will explore how by focusing on generalization and reusability, we can create a system that meets the specific needs of a particular use case while also producing individual components that can be repurposed as versatile building blocks in various surgical simulation scenarios.
Landscape more secure than portrait? Zooming into the directionality of digital images with security implications
Lecturer: Benedikt Lorch - Researcher@SEC
Date: 25.04.2024
Abstract:
The rectangular layout of typical camera sensors allows photographers to choose between landscape or portrait format. The choice between landscape and portrait format appears to be purely artistic, but can in fact have significant impact for security applications that operate on image statistics. One reason for this is that many state-of-the-art methods assume that image statistics are similar in the horizontal and vertical directions, allowing them to reduce the number of features (or trainable weights) by merging coefficients. On the one hand, this artificial symmetrization tends to suppress directional properties of natural images, causing a loss of performance. On the other hand, unaddressed directionality causes learning-based methods to overfit to a single orientation. This makes them vulnerable to manipulation if an adversary chooses inputs with the less common orientation. This talk takes a comprehensive approach, identifies and systematizes causes of directionality at several stages of the image acquisition, and demonstrates its effects in three selected security applications (steganalysis, forensic source identification, and the detection of synthetic images).
Retrieval augmented generation and knowledge graphs
Lecturer: Ioan Toma - Onlim
Date: 18.04.2024
Abstract:
In this talk we will briefly introduce Large Language Models (LLMs) and Retrieval Augmented Generation (RAG), focusing on how RAG improves LLMs by adding retrieval mechanisms. This approach connects LLMs with various information sources, ensuring that the content they produce is relevant and up-to-date. Next, we'll talk about Knowledge Graphs (KGs), explaining how they organize facts as entities and relationships by following Web standards such as RDF and SPARQL. The final part of the talk will be about the combination of the two technologies. This combination provides LLMs with accurate facts from KGs, helping to make sure the information they generate is correct and reducing errors or made-up information. This talk aims to explain these technologies and their benefits, especially how their combination, which we call RAG KG, can be used to make Conversational AI systems e.g. chatbots more reliable and factually accurate.
Continual learning for robot manipulators
Lecturer: Sayantan Auddy - Researcher@IIS
Date: 11.04.2024
Abstract:
Continual learning for robots refers to the continuous acquisition and integration of new knowledge and skills into their existing abilities. Similar to humans who learn throughout their lives, continually learning robots can adapt to new challenges and remain effective in evolving environments. In this talk, I will share work from my PhD thesis, which proposes novel continual learning methods for robotic manipulators, with a focus on real-world applicability. The discussion is divided into two main parts. The first part examines continual learning of dissimilar tasks, encompassing fundamentally different manipulation skills with varying objectives. Specifically, I will introduce our approaches to 'Continual Learning from Demonstration' where a robot learns a sequence of manipulation skills from human demonstrations. I'll delve into the methodology, exploring aspects of efficiency and motion stability and highlighting the pivotal role of stability in enhancing continual learning performance. In the second part of the talk, I will explore continual learning in dissimilar environments. This scenario includes situations where the task objective remains consistent, but the operating environment for the robot undergoes changes, necessitating adaptation. Here, I will present our research on 'Continual Domain Randomization', a method wherein a robot is trained using continual reinforcement learning across a series of diverse simulation environments, enabling zero-shot transfer and stable performance in the real world.
Deep learning in the legal domain
Lecturer: Yiquan Wu - Zhejiang University, China
Date: 14.03.2024
Abstract:
Legal Artificial Intelligence (LegalAI) leverages AI technology to streamline legal tasks, significantly reducing paperwork for legal professionals. This talk will highlight recent advancements and research in applying deep learning to various legal tasks. It will also explore the collaboration between Large Language Models (LLMs) and domain-specific models to enhance performance in the legal domain.
Ethereum's proposer-builder separation: Promises and Realities
Lecturer: Lucianna Kiffer - ETH Zurich
Date: 25.01.2024
Abstract:
With Ethereum's transition from Proof-of-Work to Proof-of-Stake in September 2022 came another paradigm shift, the Proposer-Builder Separation (PBS) scheme. PBS was introduced to decouple the roles of selecting and ordering transactions in a block (i.e., the builder), from those validating its contents and proposing the block to the network as the new head of the blockchain (i.e., the proposer). In this landscape, proposers are the validators in the Proof-of-Stake consensus protocol, while now relying on specialized block builders for creating blocks with the highest value for the proposer. Additionally, relays act as mediators between builders and proposers. We study PBS adoption and show that the current landscape exhibits significant centralization amongst the builders and relays. Further, we explore whether PBS effectively achieves its intended objectives of enabling hobbyist validators to maximize block profitability and preventing censorship. Our findings reveal that although PBS grants validators the opportunity to access optimized and competitive blocks, it tends to stimulate censorship rather than reduce it. Additionally, we demonstrate that relays do not consistently uphold their commitments and may prove unreliable. Specifically, proposers do not always receive the complete promised value, and the censorship or filtering capabilities pledged by relays exhibit significant gaps.
Automatic hint generation
Lecturer: Jamshid Mozafari - Researcher@DS
Date: 18.01.2024
Abstract:
Nowadays, individuals tend to engage in dialogues with Large Language Models (LLMs), seeking answers to their questions. In times when such answers are readily accessible to anyone, the stimulation and preservation of human's cognitive abilities as well as the assurance of maintaining good reasoning skills by humans is crucial. This talk addresses such need by proposing hints (rather than answers or before giving answers) as a viable solution. We introduce a framework for the automatic hint generation task, employing it to construct a novel large-scale dataset, featuring approximately 160,000 hints corresponding to 16,000 questions. Additionally, we present two quality evaluation methods that measure the Convergence and Familiarity attributes of hints. To assess the quality of our dataset and proposed evaluation methods, we employed 10 individuals who annotated 3000 hints. The effectiveness of hints varied, with success rates of 81%, 47%, and 31% for easy, medium, and hard questions, respectively. Furthermore, our evaluation methods exhibit a robust correlation with annotators' results. Conclusively, our findings highlight three key insights: the facilitative role of hints in resolving unknown questions, the dependence of hint quality on question difficulty and the feasibility of employing automatic evaluation methods for hint assessment.
Web image formats: Assessment of their real-world usage and performance across popular web browsers
Lecturer: Benedikt Dornauer - Researcher@QE
Date: 11.01.2024
Abstract:
In 2023, images on the web make up 41% of transmitted data, significantly impacting the performance of web apps. Fortunately, image formats like WEBP and AVIF could offer advanced compression and faster page loading but may face performance disparities across browsers. Therefore, we conducted performance evaluations on five major browsers - Chrome, Edge, Safari, Opera, and Firefox - while comparing four image formats. The results indicate that the newer formats exhibited notable performance enhancements across all browsers, leading to shorter loading times. Compared to the compressed JPEG format, WEBP and AVIF improved the Page Load Time by 21% and 15%, respectively. However, web scraping revealed that JPEG and PNG still dominate web image choices, with WEBP at 4% as the most used new format. Through the web scraping and web performance evaluation, this research serves to (1) explore image format preferences in web applications and analyze distribution and characteristics across frequently visited sites in 2023 and (2) assess the performance impact of distinct web image formats on application load times across popular web browsers.
Learning and processing events in context
Lecturer: Martin Butz - Universität Tübingen
Date: 13.12.2023
Abstract:
Hierarchical, compositional models are crucial for interacting with our world in a versatile and adaptive manner. What is the structure of these models? How are they learned? Multidisciplinary evidence suggests that we segment our world into event-predictive conceptual structures and embed these events into contexts. I selectively introduce some of our recent neuro-cognitive models (Bayesian and generative recurrent artificial neural networks) along these lines and identify critical inductive learning and processing biases. These models have the potential to progressively close the gap between current conceptual models of cognition and embodied sensorimotor experiences. Moreover, these developments plot a path towards fully grounded AI.
Unmasking GNN recommenders: A comparative study of counterfactual and adversarial examples
Lecturer: Amir Reza Mohammadi - Researcher@DBIS
Date: 07.12.2023
Abstract:
Graph Neural Networks (GNNs) have emerged as prominent techniques within the field of recommendation systems. Due to the intricate nature of GNNs and their essential role in conveying recommendation outcomes to users while ensuring algorithmic fairness and minimizing biases, there is a pressing demand to enhance the interpretability and resilience of these approaches. Among the strategies aimed at achieving this, counterfactual explanation stands out as a pivotal approach, aligning its objectives closely with those of adversarial examples. Both methodologies share a fundamental goal: to alter the model's output with minimal perturbations. Within this context, our reproducibility study undertakes a comparative analysis of leading methodologies in these two domains. This talk aim to elucidate the distinctive traits of models operating in these realms, pinpointing their shared applications and potential synergies. Our ultimate objective is to uncover and explore the interconnectedness of these techniques, thereby fostering a deeper understanding of their combined utility and implications.
Machine-translation of Ladin
Lecturer: Samuel Frontull - Researcher@TCS
Date: 30.11.2023
Abstract:
The University of Innsbruck and the Ladin cultural institute "Micurá de Rü" are collaborating on a project dedicated to exploring the possibilities of machine translation of the (low-resource) Ladin language. Even in the contemporary era of ChatGPT, machine translation of smaller languages remains a challenging task. Despite the impressive progress of neural data-driven approaches and their establishment as state-of-the-art technology, their application to smaller languages is associated with significant difficulties. Traditionally, data-driven approaches require large training datasets to enable a successful generalisation and produce high-quality translations. However, the limited availability of resources and texts in these languages prevents their effective application. The advances of the last few years and the increased focus on low-resource machine translation have led to the development of various methods that circumvent the problem of limited data availability or make more efficient use of the available data. This talk discusses the applicability of these methods to the Ladin language and provides insights into the progress and results achieved so far in this research project.
Undecidability of polynomial termination
Lecturer: Fabian Mitterwallner - Researcher@CL
Date: 23.11.2023
Abstract:
In termination analysis of term rewrite systems one of the oldest techniques uses polynomial interpretations over the natural numbers. While many other related techniques have been proven to be undecidable over the years, this was unknown for polynomial termination. It has been conjectured to be undecidable in the literature, but we only recently found a proof confirming the conjecture. We showed the result via a reduction from Hilbert's 10th problem. Additionally we could prove that related and more restricted properties like incremental polynomial termination, polynomial termination over the real and ration numbers, and linear polynomial termination are also undecidable. In this talk I will discuss the technique of polynomial termination itself, as well as explain the rough structure and idea behind the undecidability proof.
New approaches in symmetric cryptography for privacy preserving computation
Lecturer: Arnab Roy - Researcher@SEC
Date: 16.11.2023
Abstract:
In the past years significant progress was made in privacy preserving cryptography making zero-knowledge proof (ZKP), multiparty computation (MPC) and homomorphic encryption (HE) practical or nearly practical for real-world applications like secure key storage, anonymous cryptocurrencies, private machine learning etc. Block ciphers and hash functions are symmetric key (SK) primitives that play an important role for providing data security in this newly emerging and dynamic area of modern cryptography and its applications. However, SK primitives once developed for classical applications where data in transit is secured (e.g. TLS), pose an efficiency bottleneck for this new class of cryptographic applications where data under computation has to be secured. In this talk I will discuss the performance requirements of SK primitives that are dictated by privacy preserving cryptography, and the new approaches in SK cryptography to solve the above-mentioned inefficiency bottleneck. I will describe how new types of SK primitives can significantly improve efficiency compared to their classical counter-parts like AES and SHA-2. I will show examples of such novel SK primitives, their efficiency gain (over classical primitives) and usage in real-world applications. I will also explain how the design and cryptanalysis of these new types of SK primitives pose interesting and new mathematical challenges that have not been encountered thus far and have not been decisive in developing their classical counter-parts.
Runtime monitoring and safety assurance for cyber-physical systems
Lecturer: Michael Vierhauser - Researcher@QE
Date: 09.11.2023
Abstract:
The domain of software-intensive systems in general and Cyber-Physical Systems (CPS) in particular has drawn considerable attention from both industry and academia in recent years. CPS differ from traditional software systems as they include both software and hardware components, with feedback loops where physical processes and computations affect each other, and where humans are involved in the decision-making process. This shift from traditional software systems towards pervasive systems where software, hardware, and the human actors controlling them are deeply interwoven, has a significant impact on how these systems are designed, implemented, tested, and maintained, posing new challenges and opportunities to researchers. Particularly, small (semi-)autonomous vehicles such as unmanned aerial vehicles (UAVs) or small autonomous robots have gained widespread use, with frequent interaction with humans, such as UAVs engaged in search-and-rescue flights, introducing various safety concerns that need to be mitigated. To mitigate against these threats, precautionary measures need to be taken to ensure that a CPS adheres to its specified requirements and operates within its predefined safety envelope. One crucial aspect for addressing these issues is runtime monitoring, i.e., collecting, and analyzing diverse runtime properties for detecting and mitigating potential safety and security issues. This talk explores tools, methods, and frameworks to create digital representations of a running system, using runtime models, and to ensure that the system adheres to its requirements. Furthermore, it discusses opportunities for leveraging testing, combined with runtime monitoring to create a monitoring platform, and how simulation environments and new paradigms such as Human-on-the-Loop and Human Machine Teaming affect the behavior and monitoring needs of these systems.
Resource-aware time-critical application scheduling in the edge-cloud continuum
Lecturer: Zahra Najafabadi - Researcher@DPS
Date: 19.10.2023
Abstract:
The rapid expansion of time-critical applications such as health monitoring, self-driving cars, intelligent traffic control, and emergency services, with substantial demands on high bandwidth and ultra-low latency pose critical challenges for Cloud data centers. To address time-critical application demands, the computing continuum emerged as a new distributed platform that extends the Cloud toward nearby Edge and Fog resources, substantially decreasing communication latency and network traffic. However, the distributed and heterogeneous nature of the computing continuum with sporadic availability of devices may result in service failures and deadline violations, significantly negating its advantages for hosting time-critical applications with diverse demands and lowering users’ satisfaction. Additionally, the dense deployment and intense competition for limited nearby resources in the computing continuum pose resource utilization challenges. To tackle these problems, we investigated the problem of time-critical application placement with constraint deadlines and various demands in the heterogeneous computing continuum with three main contributions:
A multilayer resource partitioning model for placing time-critical applications to minimize resource wastage while maximizing deadline satisfaction;
An adaptive placement for dynamic computing continuum with sporadic device availability to minimize resource wastage and maximize deadline satisfaction;
A proactive service level agreement-aware placement method, leveraging distributed monitoring to enhance deadline satisfaction and service success.
Virtual payment channel networks in cryptocurrencies
Lecturer: Lukas Aumayr - TU Wien
Date: 12.10.2023
Abstract:
Permissionless cryptocurrencies like Bitcoin are revolutionary but come with limitations - notably, they can handle only an extremely limited number of transactions per second. Enter Payment Channel Networks (PCNs), which let two users exchange numerous transactions with a minimal blockchain footprint. Imagine it as setting up a temporary tab with a friend, recording the final result instead of each tiny transaction. But PCNs are not perfect. They often require the use of intermediaries for routing payments, which means added fees and potential privacy concerns. Furthermore, their design primarily supports payments, leaving out numerous other fascinating blockchain applications. In this talk, we will dive into these challenges and introduce virtual channels (VCs) - a novel approach designed to address these limitations. VCs allow users to bypass intermediaries with temporary, off-chain channels and can host a wider range of applications, thus providing a cheap and generic solution for having scalable applications on Bitcoin and other cryptocurrencies.
Cyclic unification: a step towards cyclic automated reasoning
Lecturer: David Cerna - Czech Academy of Science
Date: 05.10.2023
Abstract:
Proof theory studies systems of rules useful for enumerating valid statements within a target logic. Given an expressive enough proof system, we can translate a ‘mathematical’ proof into a fully formal logic proof explicitly describing every reasoning step. Like mathematical proofs, there may be a variety of formal proofs ending with the same statement. Such proofs are critically important to mathematical practice, providing deep insights into the object of interest. A natural question is, “Can we compare formal proofs?” Gentzen’s Hauptsatz provides a method for transforming a proof by simplifying the contained lemmas and even eliminating the lemmas altogether, resulting in an analytic proof, i.e. a proof without detours. Effective construction of an analytic proof provides us with Herbrand’s Theorem, a key result providing the foundation for the resolution method and most modern automated reasoning technology. Contemporary investigations into proof transformation led by A. Leitsch and M. Baaz turned this relationship around to show that resolution-based theorem provers allow for effective proof transformation. However, the transformation methods mentioned above falter in the presence of an inductive argument. While there exist methods transforming proofs containing induction, they cannot reach the analytic proof or a representation of it. Studying proof transformation in the presence of inductive arguments is the central task of the GACR-FWF international project PANDAFOREST. We study resolution-based proof transformation for proofs containing induction to improve our understanding of automated inductive reasoning. This talk presents our recent work on inductive analogs of classical first-order syntactic unification. We show that it is possible to decide whether a given recursively-defined infinite term unifies with a given finite term. The proof depends on a simple and well-understood unification procedure and shows that there are only finitely many constraints to consider modulo variable renaming. Thus, evoking the pigeonhole principle is enough to close the argument. Our construction closely resembles certain streams, processes, and memory models; thus, we see the possibility of application beyond the scope of the project.