IMLA – Institute for Machine Learning and Analytics

An autonomous system is a system that can make decisions on its own. To do this, it must perceive its environment, develop a model of that environment based on its perceptions, and act autonomously using that model. It makes decisions, which are then translated into planned actions. These actions, in turn, affect the environment. An autonomous system often interacts with other autonomous systems.

Examples include autonomous vehicles and robots that operate autonomously—systems that are not rigidly programmed but must react to changes in their environment. Projects involving autonomous systems at Hochschule Offenburg include the Audi Autonomous Driving Cup and the Shell Eco-Marathon Autonomous Challenge (autonomous driving), as well as Sweaty (robot soccer).

Business analytics focuses on the application of machine learning methods and statistical analyses to business management issues. These methods are applied across all areas of a company: from target audience analysis and campaign optimization in marketing, to sales forecasts, to the analysis of manufacturing processes based on sensor data. Traditional topics in data analysis and storage, such as business intelligence and data warehousing—as well as their extension through machine learning methods—are also the focus of our research interests. In the future, methods for processing natural language and other unstructured data will become increasingly important in business processes as well.

Learning analytics refers to the measurement, collection, analysis, and evaluation of data about learners and learning processes in order to provide support for learning and teaching using the collected data. A wide variety of machine learning methods and approaches are used in this process. The underlying data is typically generated in electronic learning environments and digital exam management systems. It is based, for example, on exam results, learning materials, online discussions, tests, and exercises. In all activities related to learning analytics, learners’ privacy and personal data must be protected. 

With the help of machine learning, computers are able to learn on their own, based on data, how to solve complex tasks without having to be explicitly programmed to do so. Machine learning algorithms use provided training and test data to search for patterns and learn a model that can be used to make decisions automatically in the future. The goal is for the computer to learn automatically and adapt its actions without human assistance.

There are a number of different machine learning methods, which can be categorized into supervised and unsupervised learning, as well as reinforcement learning. We work with all types of machine learning, with a particular focus on deep learning.

Unlike traditional, structured databases, multimedia databases store unstructured data—primarily images, audio, and video, but also free-form text. A viewer can understand the content of an image or a video, but its meaning remains hidden from a computer. As a result, searching for multimedia content in a database is only truly feasible if semantic metadata is available to describe that content. Today, this metadata can be generated using deep learning; examples include objects in images and videos, or faces/people and their emotions. This would make it possible, for instance, to search for images, sounds, and videos featuring a tiger.

Parallel computing is a prerequisite for the practical implementation of many machine learning techniques. For example, parallelization reduces the training time for deep neural networks—which can sometimes take weeks—to just a few days or hours, making the practical application of deep learning possible in the first place. Many other machine learning methods are also highly computationally intensive and require parallel or distributed computing.

We focus in particular on massive parallelization using graphics processing units (GPUs) and algorithms optimized for them, e.g., in the field of data clustering. The IMLA has several servers equipped with NVIDIA Tesla GPUs for highly parallel data processing.

Machine learning methods can also be used in a mobile context, particularly on smartphones. This can be achieved using various system architectures: On the one hand, the algorithms can be executed entirely on servers, meaning that the data to be analyzed must be transmitted from the smartphone via appropriate interfaces. However, this requires a constant connection to the servers. Alternatively, when using neural networks, for example, only the computationally intensive training can take place on the servers; the trained network is then transferred to the smartphone and used locally (see TensorFlow Lite). Another approach is to run machine learning applications entirely on the smartphone, for example, using hardware-accelerated SDKs such as ARCore (Android) or ARKit (iOS).

We are working on issues related to the ethical and societal impacts of AI.

Project website: www.ki-bohrer.de

In addition to the many operational advantages, however, the construction of geothermal plants in metropolitan areas also poses a significant technical and social problem: with currently available methods, the deep drilling required at urban sites leads to considerable and prolonged noise pollution for local residents. In particular, since drilling must continue continuously—24 hours a day, 7 days a week, for months on end—for technical and economic reasons, complying with legal noise limits of typically 35 dB (at night in residential areas) poses an enormous technical and logistical challenge. Current approaches to noise control rely on a combination of I) continuous monitoring of noise emissions through acoustic measurements in the surrounding area; II) optimization of construction site logistics, e.g., scheduling noisy activities for specific times of day; III) passive sound insulation through protective walls and enclosures around the drilling rigs; IV) active noise reduction through (currently) manual adjustment of drilling control parameters.

While Approaches I–III are currently part of the technical standard, we see considerable room for innovation and a need for further development in active noise reduction. The manual control methods currently in use are insufficient to address the comprehensive complexity of the optimization problem at hand: Noise control and the technically safe, economical advancement of the borehole are not only conflicting objectives; monitoring a wide range of operating conditions and simultaneously adjusting hundreds of control parameters can only be implemented suboptimally by manual means. Therefore, the goal of this project is to automate the optimization of drilling operations using state-of-the-art AI methods. We expect this fundamentally new approach to significantly improve noise protection for local residents while simultaneously increasing drilling speed. This is expected to significantly increase not only the public acceptance of geothermal projects but also their economic viability. Both of these improvements over the current state of the art could make a substantial contribution to the further expansion of geothermal energy and, consequently, to the achievement of overarching climate goals.

Project Website: https://www.llm-praxis.de/

In the context of LLMs—as with the introduction of many other disruptive technologies—a wide range of scientific, technological, legal, and societal issues arise, covering a broad spectrum and great depth. These issues are undoubtedly all of great importance, but it is impossible to address them all within a single project. Therefore, the proposed project will focus on a few core topics—both technologically and in terms of the implications of the technology and the broader context—that are of high practical importance during the early phase of testing and developing LLM solutions for SMEs and that minimize risks related to other aspects.

The LLMpraxis project is explicitly not intended to research new GPT algorithms or to design or train models from scratch. Instead, it will draw on the now-extensive range of so-called foundation models (available under open licenses) and open-source frameworks, with the aim of adapting them to the requirements of real-world applications at the four partner companies and deploying them productively. In addition to many practical issues regarding hosting, APIs for existing software, and the scalability of LLM deployments for growing user numbers on the companies’ own local systems, the research focus—as part of a doctoral program accompanying the project—will be on questions regarding the restriction of LLMs to their own data and the correct referencing of that data (citation).

At the heart of the project is the application-centered approach: in collaboration with four partner companies, five specific applications were defined in the two basic categories of coding assistant (AP5) and chatbot (AP6), whose requirements drive research and development. The goal is to deploy initial solutions into live operation early in the project and thereby identify the practical challenges (AP7).

Link to the project website:  https://www.lab2device.de/ 

Current list of publications on Google Scholar

• Pfarrer Dr. Gernot Meier, Evangelische Landeskirche in Baden