Group Data Efficient Automated Learning

One challenge in the use of AI, both within manufacturing companies and in logistics, is the procurement of a lot of high-quality data. In particular, the effort to label the data, i.e. to provide it with status information, is time-consuming and cost-intensive. The higher the amount of information and its quality, the better the AI model trained later. To make the process as efficient as possible here, on the one hand we are researching the application of methods from the field of Data Centric AI to increase the quality of the data and the labels as well as to generate additional labels efficiently. On the other hand, we use methods of Semi-Supervised Learning in order to still be able to train good AI models with little label information. In order to be able to scale these AI models to other similar applications, we are researching different approaches to be able to reuse existing models with as little additional data as possible. One example of this is transfer learning methods.

In addition to data efficiency, efficient handling in training, maintenance and servicing of the AI models is also essential. In particular, we counteract the shortage of skilled workers by researching comprehensible and automated solutions. The high manual effort of recurring processes, for example the creation and optimization of AI models, can be reduced by the sensible application of Automatic Machine Learning (AutoML). But also in the area of operationalization, i.e., the operation of the ML solution, Machine Learning Operations (MLOps), automated processes, such as Automated Testing & Monitoring, can increase personnel efficiency.

In automation, computational effort also plays an important role in generating sustainable solutions. For this reason, we are researching, for example, resource-efficient optimization methods and the use of pre-trained models.

Currently, the Machine Learning Lifecycle is only managed by highly qualified personnel, i.e. commissioned, serviced and maintained. To counteract the shortage of skilled personnel, we are researching comprehensible methods in the field of MLOps to simplify the maintenance and servicing of the models and thus generate low-threshold access to the use of AI.