Prof. Michael Beck

The recent rise in popularity for machine learning algorithms, specifically the field of convolutional neural networks created a large demand for labelled data. For example, an image is labelled if there exists accompanying information of what can be seen on the image and where these objects are located within the image. But how many data-samples does a machine learning model need to reach a good performance? What are the costs of collecting this data and is every sample "worth" the same? How can we identify and get rid off samples that hurt our model or even better avoid collecting them in the first place? How to utilize data the best in machine learning is a question that is often overlooked in research, but it has very real implications for all the people out there who want

Agriculture currently undergoes a digital revolution with many new technologies making their way into all aspects of the industry. In the Digital technologies in agriculture and rural areas briefing paper of the United Nations Food and Agriculture Organization we can read: "In the agriculture and food sector, the spread of mobile technologies, remote-sensing services and distributed computing are already improving smallholders’ access to information, inputs, market, finance and training. [...] Digitalization will change every part of the agrifood chain." The agriculture of tomorrow could look vastly different to what we are used to today. Autonomous agents equipped with intelligent cameras and sensors could evaluate crops and weeds on a plant-by-plant basis and take action not unsimilar to how we take care of our plants in our gardens. These are challenging tasks for today's robotic systems and machine learning models but solving them promises a reduction in resources used while improving yields and preserving our environment. This is exactly one of the things we do at the TeraByte project

Microplastics are plastic particles in the size range from 5 mm down to a few micrometers. They pose a global environmental challenge as their occurrences appear to be omnipresent in all water bodies investigated, including arctic ice and the deep sea. These particles enter the food chain and can impact the health of animals and humans, as they often release toxic additives. The pathways and sources of microplastics need to be better understood to to tackle this type of pollution. For this the research community is in dire need of automation when it comes to microplastic analysis. New technologies in the fields of deep neural networks and embedded systems are a promising candidate to fulfill this need.

Queueing networks in their most simple form consists of servers, their buffers, and jobs that need to be processed by service elements. The world is full of such systems: the check-outs at the supermarket are a simple one, the world wide web with millions of service elements in the form of switches and routers is another queueing system. In many situations we are interested in the worst-case scenario, rather than the average performance of a queueing network. For example, consider the exit times of visitors to a stadium in case of an evacuation: Here, the average time is a poor metric, and we are rather interested in the time needed to get even the last person out of danger. Another example are SLAs (Service Level Agreements) between a service provider and a client or the maximum transmission delay in a car that breaks-by-wire. Stochastic Network Calculus is a mathematical theory that provides bounds on these worst-case scenarios, excluding events with negligible likelihoods. This methodology is used in designing and modelling queueing networks. It has the advantages that complex queueing systems can be simplified to easier scenarios and thus provides us with an end-to-end analysis.