Alexander Kastius

In the past years, reinforcement learning (RL) gained much attention in related research communities and public media. It has been shown that it can be used for various types of problems to automate the solution of decision processes. Due to its universality, many other research areas can benefit from the tools provided by RL. To allow everybody in the data-related research groups to assess whether RL algorithms might open new opportunities in their research topic, this talk will provide a low-level introduction to the underlying concepts, fundamental problems that arise in RL applications, and their solutions. A small outlook on current research topics will be given as well.

Dr. Michael Perscheid

After receiving his PhD from the HPI, Michael joined SAP and was (among other things) responsible for the HPI-SAP cooperation and research projects from the industry-side since multiple years. In 2020, he joined the HPI again as chair representative of Prof. Plattner and now sees the other side from an academic point of view as well. Please join, Michael in this talk and interactive discussion on experiences of industry-academic-joint projects, learn how to convince the other side, and finally what needs to be done to make a product out of it.

Sarah Dsane-Nsor

The Trust Determination Model: A Grounded Theory Approach

Expectant and lactating parents are inundated with diverse pregnancy, birth, nutrition, and childcare related information online via social media and websites. The parents' ability to determine trustworthy information amidst such an extensive resource of unsupervised content is critical, particularly for users seeking online health information. We applied the grounded theory approach to derive a model for first 1000 days (FTD) parents' information-seeking behaviour. The FTD is the period between conception and the child's second birthday. We present the lived experiences of how expectant and lactating parents seek and trust online health information.

Tezira Wanyana

Agent-based Knowledge Discovery and Evolution: A KDE ontology

Knowledge Discovery and Evolution (KDE) is of interest to a broad array of researchers from both Philosophy of Science (PoS) and Artificial Intelligence (AI), in particular, Knowledge Representation and Reasoning (KR), Machine Learning and Data Mining (ML-DM) and the Agent-based Systems  (ABS)  communities.  In  PoS, Haig recently proposed a so-called broad theory of scientific method that uses abduction for generating theories to explain phenomena. He refers to this method of scientific inquiry as the Abductive Theory of Method  (ATOM).  In this work, we analyse ATOM, align it with KR and ML-DM perspectives and propose an algorithm and an ontology for supporting agent-based knowledge discovery and evolution based on ATOM. We illustrate the use of the algorithm and the ontology on a use case application for electricity consumption behaviour in residential households.

Shahryar Khorasani

Tumors located inside or in the proximity of the motor cortex or corticospinal tract increase the risk of motor function damage after surgery. Neurosurgeons need to measure this risk in their surgical planning. They use navigated transcranial magnetic stimulation (nTMS) to locate motor function in the motor cortex and define tractography of the corticospinal tract using diffusion-weighted imaging. The tumor can induce changes in the neuronal excitability of the corticospinal tract, which can be measured by nTMS. The neuronal excitability, the tumor-tract distance, and the involvement of the motor cortex are the three biomarkers that are currently used for measuring the risk of post-operative motor deficit. In this project, we try to automate the risk stratification workflow and develop deep learning models for predicting the risk of motor function deficit.

Prof. Dr. Gerard de Melo

While regular software developers often have prespecified milestones and requirements that need to be met, researchers tend to have much more flexible, open-ended goals. Moreover, while traditional scholarship has tended to emphasize thorough reading, reflection, and learning nowadays many researchers find themselves in a fast-paced and demanding environment, so we often only focus on the next deadline. Hence, it is essential to reserve time to think about longer-term goals and evaluate which research ideas are worth pursuing, how many projects you can take on simultaneously, and how to cope with challenges and failure. This talk will attempt to inspire thoughts and discussions about possible strategies to deal with these sorts of questions.

Ziena Elijazyfer

The presentation introduces a new structural characterization for general graphs, which we call balanced crown decomposition, that captures the structure imposed by their connected induced subgraphs of a given size. Such subgraphs are a popular modeling tool in various application areas, where the non-local nature of the connectivity condition usually results in very challenging algorithmic tasks. The provided structural perspective given by the balanced crown decomposition on graphs is applicable for problem types like editing, packing, and partitioning. This will be illustrated by deriving improved kernelization and approximation algorithms for a variety of such problems

Prof. Dr. Felix Naumann

We will discuss the following questions around publication venues and how to choose them:

• The role of conferences vs. journals (vs. book chapters and others)
• Conference and journal rankings
• Predatory conferences and journals
• Reviewing modes
• Chairs and editors
• Venue topics

Stephan Mandt, Assistant Professor at the University of California, Irvine

Neural image compression methods have recently outperformed their classical counterparts in rate-distortion performance and show great potential to also revolutionize video coding. In this talk, I will show how methods from approximate Bayesian inference and generative modeling can lead to dramatic performance improvements in compression. In particular, I will explain how sequential variational autoencoders can be converted into video codecs, how deep latent variable models can be compressed in post-processing with variable bitrates, and how iterative amortized inference can be used to achieve the world record in image compression performance.

Please note that this talk starts at 18:00 MEZ instead of the usual slot.

Dr. Michael Perscheid

In his talk, Michael will cover the potentials as well as challenges that come with working remote.
Especially when working remote is no longer an option but a necessity, appropriate strategies and mechanisms to focus, effectively collaborate, but also to mentally recover are vital.
Michael will share learnings from industry as well as academia, give some first (tool-) recommendations, and looks forward to a vivid discussion in the end.

Dr. Timo Kötzing

A typical introduction to scientific work at a university consists of learning-by-doing. We learn how to conduct experiments, write papers, and then choose the next question to work on by running along with seasoned researchers in the field of choice. In this talk, I want to focus on why we do all these things. What is the nature of science and what is the aim of all these activities we do, other than earning a living and continuing traditions of scientists before us?

Prof. Dr. Bert Arnrich, Bjarne Pfitzner

A literature review is the first step to enter a research field. It provides a systematic overview of state of the art and identifies open research questions. It will be the basis of your related work section in your thesis, and if published, it will be very useful for other researchers in the field. In this talk we introduce methods for systematic searching and screening literature, for extracting relevant data from literature and for writing a literature review.

Entity Linking (EL) is a compelling task in natural language processing for understanding the semantics and extracting information from text. The task is to link the entity mentions in an unstructured text to their corresponding entities in a knowledge base/knowledge graph. Wikidata is one such knowledge graph which is mainly crowdsourced and has several challenges due to its noisy, non-standard entity titles. Currently, there are very few approaches that try to solve the EL problem over Wikidata, which has shortcomings. We propose CHOLAN, a transformer-based model to target end-to-end EL over Wikidata. It is a context-aware approach that feeds global context from an external knowledge source. We empirically illustrate CHOLAN on datasets, where it significantly outperforms on various baselines and also the role of external knowledge in the transformer-based models for EL over Wikidata.

We will meet in person in F-E.06 or you can join via Zoom. We will use the first session for a quick round of introductions for new members, a quick glance at the schedule for this semester, and to reassign duties.