Lunchtime Seminar
Archive WiSe 2021/22 & SoSe 2022
Hybrid AI approaches for Knowledge-Intense Manufacturing
Lecturer: Arnold Beckmann - University of Swansea
Date: 30.06.2022
Abstract:
In this talk we will introduce our vision and contributions for applying hybrid AI methods, in the context of Knowledge-Intense Manufacturing in general, and steel manufacturing and supply-chains in particular. Hybrid AI approaches employ combinations of methods and techniques from AI. We will focus roughly on the combination of statistical AI methods (like machine learning) and symbolic AI methods (like Knowledge Based Systems - Ontologies, Knowledge Graphs). Besides our general vision we will explain several results that contribute to a hybrid AI approach for manufacturing in general, and steel in particular.
Our results have been obtain in the context of the SUSTAIN project (https://www.sustainsteel.ac.uk/) and in collaboration with Tata Steel.
Robot learning from few demonstrations by exploiting the structure and geometry of data
Lecturer: Sylvain Calinon - Idiap Research Institute
Date: 23.06.2022
Abstract:
A wide range of applications can benefit from robots acquiring manipulation skills by interaction with humans. In this presentation, I will discuss the challenges that such learning process encompasses, including representations for manipulation skills that can exploit the structure and geometry of the acquired data in an efficient way, the development of optimal control strategies that can exploit variations in manipulation skills, and the development of intuitive interfaces to acquire meaningful demonstrations.
From a machine learning perspective, the core challenge is that robots can only rely on a small number of demonstrations. The good news is that we can exploit bidirectional human-robot interaction as a way to collect better data. We can also rely on various structures that remain the same within a wide range of robotic tasks. Such structures include geometrical aspects, by extending learning strategies that have been originally developed for standard Euclidean space to Riemannian manifolds. In robotics, these manifolds include orientation, manipulability ellipsoids, graphs and subspaces. Another type of structure that we study relates to the organization of data as multidimensional arrays (also called tensors). These data appear in various robotic tasks, either as the natural organization of sensorimotor data (tactile arrays, images, kinematic chains), or as the result of preprocessing steps (moving time windows, covariance features). Tensor factorization techniques (also called tensor methods) can be used to learn from only few tensor datapoints, by exploiting the multidimensional nature of the data.
Another key challenge in robot skill acquisition is to link the learning aspects to the control aspects. Optimal control provides a framework that allows us to take into account the possible variations of a task, the uncertainty of sensorimotor information, and the movement coordination patterns, by relying on well grounded control techniques such as linear quadratic tracking, differential dynamic programming, and their extensions to model predictive controllers. The formulation draws explicit links with learning techniques, as we can recast these techniques as Gauss-Newton optimization problems formulated at trajectory level (in both control space and state space), which facilitates the links to probabilistic approaches.
Lecturers' and Students' Experiences with an Automated Programming Assessment System
Lecturer: Clemens Sauerwein, Researcher at QE - University of Innsbruck
Date: 09.06.2022
Abstract:
Assessment of source code in university education has become an integral part of grading students and providing them valuable feedback on their developed software solutions. Thereby, lecturers have to deal with a rapidly growing number of students from heterogeneous fields of study, a shortage of lecturers, a highly dynamic set of learning objectives and technologies, and the need for more targeted student support. To meet these challenges, the use of an automated programming assessment system (APAS) to support traditional teaching is a promising solution. This talk addresses this trend by analyzing the experiences of lecturers and students at various universities in Austria with an APAS and its impact during the last two years
Computational models for ecological perception in virtual and augmented reality
Lecturer: Fabio Solari - University of Genoa
Date: 02.06.2022
Abstract:
A bio-inspired computational model of visual perception for action tasks is proposed to provide clues to better design virtual and augmented reality (VR and AR) systems. The proposed neural model is based on space-variant mapping, disparity, and optic flow computation by implementing paradigms of the dorsal visual processing stream. The cortical representation of the visual information is directly exploited to infer features related to the real world without devising ad-hoc computer vision algorithms. Besides artificial vision applications, the proposed model can mimic and describe human behavioral data of both motion and depth perception. By leveraging previous outcomes, we can employ the modeled perception to improve the experience in VR and AR environments: in particular, to implement a foveated depth-of-field blur that mitigates cybersickness.
Machine learning approaches for cylinder wear analysis
Lecturer: Adéla Moravová, Researcher at IGS - University of Innsbruck
Date: 19.05.2022
Abstract:
Large engines are an vital part of many fields in industry, used, for example, in maritime transport and in power generation, as backups or as a global solution of the energy distribution. Therefore, the need to extend the engine's life span and optimize its maintenance is severe, as the energy and the economic costs are to be minimized.
This master thesis investigates specific part of the large engines; its cylinders. During the operation, the cylinder accumulates wear, which leads to suboptimal distribution of lubricant, which may grow into a severe damage to the engine. RGB images and corresponding depth maps of the surface of the cylinders are analyzed in both un- and supervised approaches with various techniques. Lastly, a neural network predicts a depth map on a microscopic scale from a single surface image.
Future-proofing the online order: The promise of law and the perils of regulation
Lecturer: Matthias Kettemann, Institut für Theorie und Zukunft des Rechts - University of Innsbruck
Date: 12.05.2022
Abstract:
Who decides what we can say online? Elon Musk? Russian and Chinese authorities? European rules? States with their laws play an important role, but since the early 2000s online platforms with their private rules and community standards have become increasingly important in deciding what's allowed online and what isn't. Through their rules, and their automated decision-making engines, including their recommender algorithms, they have emerged as a high-powered communication rule-maker and -enforcer. Is this a sustainable model? What can we do to ensure a sustainable online order where public values play a role? Together we will enquire into the potential of different models to "redemocratize" online spaces. .
Extrapolation in robot skill generalization
Lecturer: Hector Villeda, Researcher at IIS - University of Innsbruck
Date: 05.05.2022
Abstract:
Imitation learning approaches achieve good generalization within the range of the training data, but tend to generate unpredictable motions when querying outside this range. We present a novel approach to imitation learning with enhanced extrapolation capabilities that exploits the so-called Equation Learner Networks (EQLN). Unlike conventional approaches, EQLN use supervised learning to fit a set of analytical expressions that allows them to extrapolate beyond the range of the training data. We augment the task demonstrations with a set of task dependent parameters representing spacial properties of each motion and use them to train the EQLN. At run time, the features are used to query the EQLN and generate the corresponding robot trajectory. The set of features encodes kinematic constraints of the task. We validate the results of our approach on manipulation tasks where it is important to preserve the shape of the motion in the extrapolation domain. Our approach is also compared with existing state-of-the-art approaches, either in simulation and real setup environments.
Knowledge graph-based hard drive failure prediction
Lecturer: Tek Raj Chhetri, Researcher at STI - University of Innsbruck
Date: 28.04.2022
Abstract:
The hard drive is one of the important components of a computing system, and its failure can lead to both system failure and data loss. Therefore, the reliability of a hard drive is very important. Realising this importance, a number of studies have been conducted and many are still ongoing to improve hard drive failure prediction. Most of those studies rely solely on machine learning, and a few others on semantic technology. The studies based on machine learning, despite promising results, lack context-awareness such as how failures are related or what other factors, such as humidity, influence the failure of hard drives. Semantic technology, on the other hand, by means of ontologies and knowledge graphs (KGs), is able to provide the context-awareness that machine learning-based studies lack. However, the studies based on semantic technology lack the advantages of machine learning, such as the ability to learn a pattern and make predictions based on learned patterns. Therefore, in this paper, leveraging the benefits of both machine learning (ML) and semantic technology, we present our study, knowledge graph-based hard drive failure prediction. The experimental results demonstrate that our proposed method achieves higher accuracy in comparison to the current state of the art.or.
Context matters: An example of country context in music recommender systems
Lecturer: Christine Bauer - Utrecht University
Date: 07.04.2022
Abstract:
Algorithmic curation has been adopted by online music platforms to assist listeners in navigating the huge catalogs of music recordings. An ideal music recommender system is meant to propose “the right music, to the right user, at the right moment”. What happens if a music recommender does not consider country-specific variations in music taste? And what happens if so? Besides showcasing country-specific differences in music taste, I will demonstrate performance differences of recommender approaches with and without considering the country context.
Continual learning from demonstration of robotic skills
Lecturer: Sayantan Auddy, Researcher at IIS - University of Innsbruck
Date: 31.03.2022
Abstract:
Methods for teaching motion skills to robots focus on training for a single skill at a time. Robots capable of learning from demonstration can considerably benefit from the added ability to learn new movements without forgetting past knowledge. To this end, we propose an approach for continual learning from demonstration using hypernetworks and neural ordinary differential equation solvers. We empirically demonstrate the effectiveness of our approach in remembering long sequences of trajectory learning tasks without the need to store any data from past demonstrations. Our results show that hypernetworks outperform other state-of-the-art regularization-based continual learning approaches for learning from demonstration. In our experiments, we use the popular LASA trajectory benchmark, and a new dataset of kinesthetic demonstrations that we introduce in this paper called the HelloWorld dataset. We evaluate our approach using both trajectory error metrics and continual learning metrics, and we propose two new continual learning metrics. Our code, along with the newly collected dataset, is available at https://github.com/sayantanauddy/clfd.
Exploring Factors in a Crossroad Dataset Using Cluster-Based Association Rule Mining
Lecturer: Mahtab Shahin - researcher at DPS and TTÜ Infotehnoloogia Maja (ICT)
Date: 24.03.2022
Abstract:
Investigating the contributory factors in crossroad accidents is a high-priority issue in the traffic safety analysis. This study exploits a method based on association rules to analyze these contributory factors. Using data about one year of crossroad traffic accidents in Isfahan, Iran, 63 and 156 association rules are generated for non-serious and serious accidents, respectively. The results show that both accident severity levels are associated with head-to-the-side collisions and the spring season.
The frequency of non-serious accidents is about 38% higher than that of serious accidents. However, the association analysis results show that serious accidents are associated with more influencing factors than non-serious. Seat belt usage and road surface condition are additional decisive factors for serious accidents but not so for non-serious. The association analysis reveals that many influencing factors (such as traffic lights and the existence of a traffic enforcement camera) exhibit effects only under some specific circumstances (e.g., the peak of traffic).
Differentiable Inductive Logic Programming
Lecturer: Stanislaw Purgal, Researcher at CL - University of Innsbruck
Date: 10.03.2022
Abstract:
Inductive Logic Programming (ILP) is a form of symbolic machine learning which learns a logic program from background knowledge. In 2017 a method of doing ILP through backpropagation was proposed, though is was very weak. Recently, we have found a few ways of improving this method, which I am going to present.
Expected Cost Analysis of Probabilistic Programs
Lecturer: Jonas Schöpf, Researcher at TCS - University of Innsbruck
Date: 03.02.2022
Abstract:
Static analysis of programs is an active research area which is concerned with developing methods for finding meta-information about programs such as runtime bounds, termination or memory safety. In recent years, programs with probabilistic behavior became more important and thus the interest in finding suitable analysis methods has grown. We present a modular cost analysis method which is implemented in our tool ecoimp. With this method we are able to derive upper bounds on the expected cost of probabilistic programs. It is one of the first methods which can analyze a probabilistic program in a modular way and ecoimp is a few magnitudes faster than similar tools. A recent extension to (partly) recursive programs and constraint solving using semi-definite programming gives much more expressive power while slowing down our analysis. However, we are able to give an automatically derived upper bound of a probabilistic version of the quickselect algorithm.
Results from MIDSISE project: analysis of security curricula and industrial needs with NLP
Lecturer: Irdin Pekaric, Researcher at QE - University of Innsbruck
Date: 27.01.2022
Abstract:
Given the ongoing "arms race" in cybersecurity, the shortage of skilled professionals in this field is one of the strongest in computer science. The currently unmet staffing demand in cybersecurity is estimated at over three million jobs worldwide. Furthermore, the qualifications of the existing workforce are largely believed to be insufficient. We attempt to gain detailed insights into the nature of the current skill gap in cybersecurity. To this end, we correlate data from job ads and academic curricula using two kinds of skill characterizations: manual definitions from established skill frameworks and "skill topics" that are automatically derived by text mining tools. Our analysis shows a strong agreement between these two analysis techniques and reveals a substantial undersupply of non-technical cybersecurity skills such as compliance, management, and certification. Based on the results obtained with the new analysis techniques, we provide recommendations on curricula development in cybersecurity to decrease the identified skill gaps
From Internet of Things Platforms to Web of Things User Agents
Lecturer: Andreas Harth - University of Erlangen-Nuremberg
Date: 20.01.2022
Abstract:
Many current enterprise IoT projects are implemented with IoT platforms, which follow a centralised architecture. Sensor readings are transmitted to a cloud, where the data is processed and aggregated. The cloud is also responsible for controlling actuators. Such a setting works for systems that are managed by a single organisation. But future IoT systems involve devices and data from multiple organisations, which makes a centralised architecture difficult to implement, both from a technical perspective and from an organisational perspective. The talk starts with an overview of centralised Internet of Things platforms and outlines their shortcomings. The talk then introduces hyperlinks as a means for decentralised linking and discovery on the web, and sketches systems that sense and actuate within an agent architecture in a decentralised setting.
Do Cookie Dialogs Work? Measuring the Effect of Privacy Preference Signals
Lecturer: Maximilian Hils, Researcher at SEC - University of Innsbruck
Date: 13.01.2022
Abstract:
Since the inception of the General Data Protection Regulation (GDPR) in 2018, many websites now feature cookie dialogs which provide users with the opportunity to deny data processing based on consent. This is at odds with AdTech vendors, who have monetary incentives to build ad profiles of all users. Naturally, a key question is whether the displayed dialogs are just shown to simulate compliance, or whether third parties are actually compliant and stop processing personal data. This talk presents work-in-progress that tries to answer this question using both manual and automated measurements.
Predicting Temporal Validity of Text
Lecturer: Adam Jatowt, Researcher at DS and DiSC - University of Innsbruck
Date: 16.12.2021
Abstract:
Knowing whether and how long information remains valid is important in various applications including user state tracking in social network services and in chatbot conversations, as well as is beneficial for deep story understanding. However, such an inference remains still a difficult problem for machines as it often requires temporal common-sense reasoning. We propose a novel task, Temporal Natural Language Inference, designed to determine the temporal validity of text content, which is inspired by natural language reasoning in NLP. The task requires inference whether an action expressed in a sentence is still ongoing or has been rather completed (hence whether the sentence still remains valid or rather has become invalid) in view of additional information provided in the form of either supplementary sentence or of elapsed time period. We created a large-scale dataset for this task and develop effective models for its solution.
Formalization of the first order theory of rewriting
Lecturer: Alexander Lochmann, Researcher at CL - University of Innsbruck
Date: 09.12.2021
Abstract:
The first-order theory of rewriting is a decidable theory for finite left-linear right-ground rewrite systems. We present a formally verified variant of the decision procedure for the class of linear variable-separated rewrite systems. This variant supports a more expressive theory and is based on the concept of anchored ground tree transducers. The correctness of the decision procedure is formalized in Isabelle/HOL.
Hit Song Prediction
Lecturer: Michael Vötter, Researcher at DBIS - University of Innsbruck
Date: 02.12.2021
Abstract:
Hit song prediction is a subfield of the music information retrieval research field. In general hit song prediction is a popularity prediction task that is modeled in different ways. The overall aim is predicting the popularity of a given song before or shortly after its release. This is of particular interest for the music industry as well as for individual artists to tailor songs towards success. From a scientific standpoint two general questions arise. (a) What influences the success of a piece of music and (b) how can we create models that are able to predict the popularity of a song? Our focus lies on the second overarching question. To answer this question in different ways of modeling the task range from binary classification (hit/popular and non-hit/unpopular) to regression that aims to predict e.g. the top position in charts or streaming metrics such as listener counts. To achieve this, different types of features and models are used. We distinguish two categories of song features: (a) internal (audio) features and (b) external features that are related to the song that cannot be derived from the song itself. Further, the community uses a wide variety of models ranging from linear models to neural networks to utilize mainly internal features to tackle the task. One major obstacle for research in this field is the availability of publicly available data. Our recently published dataset is one of the few publicly available datasets that is large in size containing numerous features for a wide range of non-exotic music and hence tries to overcome this problem.
FaaSification and Serverless Functions Abstraction, Orchestration, and Distribution
Lecturer: Sashko Ristov, Researcher at DPS - University of Innsbruck
Date: 07.10.2021
Abstract:
Function-as-a-Service (FaaS) is a popular technology in recent years to develop and run cloud applications. Developers build and deploy their codes as serverless functions and the entire underlying platform and infrastructure is completely managed by cloud providers. However, FaaSification (conversion of existing monoliths to the FaaS architectural style) is challenging because serverless functions are isolated from each other. Further on, in order to exploit all benefits that FaaS paradigm offers, but also to overcome FaaS limitations, serverless functions can be orchestrated in a workflow, which may allow the runtime system to instantly spawn numerous functions. This talk will present our recent Dependency-Aware FaaSifier (DAF) that simplifies the creation of serverless functions, as well as techniques how to abstract, model, orchestrate, and distribute them across multiple FaaS providers.
Inpainting with particle hydrodynamics
Lecturer: Viktor Daropoulos, Researcher at IGS - University of Innsbruck
Date: 22.10.2020
Abstract:
Image inpainting refers to an interpolation technique used to reconstruct a damaged or incomplete image by exploiting available image information. Restoring the missing image data, in a visually plausible manner, is a challenging task since it is an ill-posed inverse problem. The main goal of this work is to perform the image inpainting process using the Smoothed Particle Hydrodynamics (SPH) technique, a meshfree approach, by exploiting a set of sparsely distributed image samples. Spatial and data optimization is performed and various isotropic and anisotropic kernels are assessed both on random and spatially optimized inpainting masks.
3D human pose estimation in the wild
Lecturer: Stefan Fogarollo, Researcher at IGS - University of Innsbruck
Date: 25.11.2021
Abstract:
Human pose estimation is a long-standing problem in the computer vision research area. The problem consists of retrieving the pose and the orientation of human body. Current standard pipelines assume to have training data in the form of images, camera information and 3D annotations. Such information can be only retrieved in a controlled settings and therefore current pipelines do not work well in the wild. The goal of the thesis is to conceive (and evaluate) a model that takes as input some images of the same human body and outputs its pose in the world coordinates. This model would be able to work in the wild where access to camera information is infeasible. With respect to the described goal, we successfully implemented and tested the method in the Human3.6M dataset.