Prof. Dr. Felix Naumann

-          Welcome to new members, farewell to leaving members

-          Web pages

-          Research retreat(s)

-          Cooperation with research school branches

-          Encouraged research activities

-          Research school rules

-          Miscellaneous

Tim Repke

Text collections, such as corpora of books, research articles, news, or business
documents are an important resource for knowledge discovery. Exploring large
document collections by hand is a cumbersome but necessary task to gain new
insights and find relevant information. Traditional information retrieval systems
require users to have some prior knowledge to formulate relevant queries and
interpret how the results fit into the broader context. An overview of the entirety
of a dataset in the form of a two-dimensional map-like visualisation may help to
circumvent this issue. Such overviews enable novel interaction paradigms for
users, which are similar to the exploration of digital geographical maps. In particular,
they can provide valuable context by indicating how a piece of information fits
into the bigger picture. This presentation will provide an overview of approaches
for generating such maps from different kinds of data.

Alexander Rakowski

Genome-wise association studies need to be performed on quantifiable traits. Medical images (or other highly-dimensional data) potentially contain a high number of information. Most approaches focus on extracting/computing quantities of desire e.g., volumetric measurements. However this can limit the range of the study, and requires tailored methods - e.g., specialized software or expert annotations. Alternatively, TransferGWAS aims to learn meaningful features using the transfer learning paradigm, leveraging the representation learning abilities of neural networks.

This presentation will be in form of an interactive tutorial, walking through a complete example of the TransferGWAS method, showing how to train a neural network model on medical images, use it to extract meaningful feature representations, and use the features to perform a genome-wise association study.

Alexander Kastius

Reinforcement learning has been used for various applications to solve decision processes. Deep RL can be used to tackle challenges that incorporate high-dimensional inputs and large, continuous action spaces. Many of those examples have been achieved with data generated from simulated environments. In those, it usually becomes a question of resources to generate the tremendous amounts of observations required to allow convergence towards well-performing policies. (1) In many real-world applications, it is impractical or very costly to generate many observations. In many enterprise problems, exploration has to occur without simulation. (2) In other domains, we require robustness towards slightly varying environments, without having to learn from scratch for each of them. In both cases, (1) and (2), it becomes an interesting alternative to learn policies based on multiple related problems. This talk will introduce several methods to learn from multiple tasks, for example, policy transfer, representation transfer, and task mapping. For representation transfer, we have the goal of reusing elements of one model for a different task. Task mapping aims at learning a mapping of one process's state and action space to another process's elements. In policy transfer, independently trained policies influence each other to converge towards a unified strategy.

Wang Yue

In the last decade, the rapidly increasing number of IoT devices constitutes huge decentralized sensor networks, which provide a big challenge to today’s stream processing. By 2021, the installation of IoT devices is beyond 31 billion. However, the current SPEs centralize all data and regardless of the different characteristics of queries, which is a limit factor to stream processing in the sensor network and leads to unnecessary overhead, such as Apache Flink, Apache Spark Streaming, and Apache Storm. In addition, sharing computation results among multiple continuous queries can dramatically improve performance, especially for crossing intermediate results among arbitrary windows which have different types and aggregation functions. Therefore in our work, we propose a new decentralized -stream-oriented stream processing system, which is specific for the distributed data stream and able to process arbitrary windows as well as aggregations flexibly but does not require centralizing all queries and data in a single node. We design an optimizer that broke through the limitation of sharing intermedia results between the same window types and functions. On the contrary, it is able to cross the intermedia result among the queries which have arbitrary window types and aggregation functions.

Gerardo Vitagliano

Preparing data. Exploring data. Visualizing data. Machine learning data. Data-ing data. No matter what your task is, there is a fundamental requirement: the ability to load data.

Thankfully, the CSV format sets a nice standard to distribute and consume raw data, right? Unfortunately, this is not always the case. 

The flexible nature of the CSV format often leads to files that deviate from standards and best practices, e.g., containing preambles, multiple tables, strange dialects, etc., etc.

We define these files as "polluted".
This talk will address some of the research questions that arise with polluted files:

• What are typical, real-world CSV "pollutions" that prove challenging for data loading?
• How much "intelligence" can be expected in state-of-the-art systems to automatically load such messy files?
• How can we measure the correctness of data loading?

To answer these questions, we present Pollock, our data loading benchmark. 

Prof. Dr. Felix Naumann

During this moderated discussion, we will explore the definitions of availability, repeatability, and reproducibility, establish the need for repeatable experiments in computer science, review different community efforts to support repeatability, and discuss ways that we can support these efforts and thus establish better scientific practices. A good starting point are definitions and policies laid out by the ACM: https://www.acm.org/publications/policies/artifact-review-and-badging-current

Manoj Prabhakar

In NLP, it is very common to use crowdsourcing for data collection and annotation. This has created a lot of new opportunities for data-driven machine learning. The various online crowdsourcing platforms make it very convenient to get access to crowd workers for various kinds of tasks, including word similarity, textual entailment, question answering, etc. It has helped several researchers to collect and annotate data on their own rather than reusing existing data with some limitations. In this presentation, we will take a look at an overview of crowdsourcing with insight into one online platform that we used for our research.

1. Academia meets Industry: SAP HANA Research Campus (Daniel Ritter) 

   With SAP HANA Cloud, enterprises are able to run business and analytical applications on top of the same underlying database system. However, building cloud databases for heterogeneous applications is hard and their design is interconnected with emerging technology.

   To address the challenge of providing a commercial, enterprise-ready (cloud) database and - at the same time - leverage technological and scientific advances and find innovative solutions for practically difficult database problems, the SAP HANA Research Campus was founded in 2009. In comparison to well-known industrial research centers from IBM and Microsoft, the "HANA Campus" exclusively focuses on database system research based on practical challenges and the transfer of research results into the product. That requires a close collaboration of leading, international, academic research groups and the HANA development teams. If you want to find out how this can be done, join our talk in the DSE seminar!

2. Information-Theoretic Causal Structure Learning from Mixed and Nonlinear Data (Johannes Huegle) 

   Causal structure learning (CSL) has received widespread attention in the scientific field as the knowledge of underlying causal structures is the basis for decision support within many real-world scenarios, e.g., genetics or discrete manufacturing. As part of research at the chair of Prof. Plattner, I aim to reduce the barriers for the transfer of CSL into practice through the provision of a supportive IT-system and the incorporation of information-theoretic statistical methods. In this context, statistical conditional independence (CI) testing is a fundamental task underlying methods for constraint-based CSL and particularly challenging in practice due to the omnipresence of mixed discrete-continuous and non-linear data. In this talk, I present current work on a information-theoretic CI test that combines a k-NN-based conditional mutual information estimator with a local permutation scheme and allows for CSL from mixed discrete-continuous and nonlinear data.

Justin Albert

Gait analysis is an important tool for the early detection of neurological diseases and assessing the risk of falling in older people. Various gait parameters, such as walking speed, step length, and clearance, are good health status indicators. Usually, gait is assessed in specialized biomechanical laboratories using expensive motion capture systems that require professional operators. With the recent advances in human pose estimation due to deep learning, stable 3D joint positions of users can be tracked in real-time, even on 2D images only. This technology offers great potential for health-related applications also accessible for amateurs. This talk will present work in progress from an ongoing evaluation study using a 3D human pose estimation algorithm for treadmill gait analysis. For this purpose, a dataset was recorded containing high-quality ground-truth motion capture data and synchronized RGB videos. We assess the motion-tracking accuracy and the agreement of calculated gait parameters from the DL-based algorithm and compare them to those from the gold-standard system.

Prof. Dr. Tilmann Rabl

In this talk, we will discuss major aspects of good scientific practice according to the DFG Codex. Furthermore, we will look at common pitfalls for young researchers and solution strategies.

Michael Perscheid

Have you ever pitched your ideas to people who are not experts, low on time but still decide on your proposal? In this talk, Michael will present a (brief) framework on how to structure your arguments, prepare your slides, and focus on the main message in order to convince your audience and get the decision you want. While this view is very industry-specific and the experience is based on board-level decisions, the presented insights will also be valuable for research-related argumentations or for your later career.

Sidratul Moontaha

Epilepsy is a neurological disease causing sudden electrical bursts in the brain resulting in seizures. Approximately 30 percent of these patients have drug-resistance epilepsy and are suffering from multiple types of seizures for several days or even years. Seizure control is only possible by several simultaneously administered drugs with very insufficient guidance about the combination and dosage. Therefore, it is essential to deploy a model that could reliably adjust the medication dosages within a minimum period to reduce seizure frequency. As a result, we propose building a non-linear state-space model and employing extensions of the Kalman filter to estimate the states of the model. Validating this algorithm with a simulated dataset, we propose deploying this model to the data set from epilepsy patients.

Prof. Dr. Robert Ranisch

Research and development of AI and Big Data technologies are increasingly required to meet ethical guidelines. Especially when data-intensive research involves humans (or their data), they may also be subject to requirements for ethical oversight. In this context, "Ethik-Kommissionen" or Ethics Review Committees (ERCs) can play a role in supporting responsible research. However, these committees, which have their origins in biomedical and clinical research, are only slowly adapting to the peculiarities of data-intensive research. Conversely, there is sometimes a lack of experience in the field of data science with respect to the requirements for applications to an ERC. The talk will give a brief overview of ethics advisory boards (in Germany) and an introduction to the function and operation of ERCs. Current developments of such bodies in the context of Big Data research will be discussed as well as some requirements for a successful application.